ML on satellite, radar and in-situ observations for weather and climate: retrievals, precipitation estimation, cloud detection, gap filling of observations.
Search and filter this topic Plain list on GitHub
Neural network predicts GNSS zenith tropospheric and wet delay from GNSS and meteorological data, reaching 5 mm (ZTD) and 3.6 mm (ZWD) error for water-vapour estimates.
Sim2Real pre-trains a ConvCNP on reanalysis and fine-tunes on German weather stations, outperforming reanalysis-only or station-only training for surface air temperature interpolation.
Microsoft Research (AI for Science)University of Cambridge (Aardvark)
Fuses multiple data sources with machine-learning spatial downscaling to reconstruct all-weather precipitable water vapour maps; no abstract available, so results are unverified.
Machine-learning-derived 1° global climatology of eight WMO cloud types from ESA Cloud_cci retrievals, almost gapless from 1982 to 2016.
Uses deep reinforcement learning with a transformer policy to learn improvement heuristics for climate sensor placement, comparing favourably with existing solvers.
U-Nets retrieve storm maximum vertical velocity from 3D radar reflectivity with probabilistic outputs, reaching R² above 0.65 on WoFS simulations.
TaNet, a transformer, estimates near-surface air temperature from FengYun-4A imagery, reaching correlation 0.990 with ERA5 and 0.959 with stations.
UPerNet with ConvNeXt segments aircraft contrails in GOES-16 infrared false-colour imagery, ranking in the top 5% of a competition.
A capsule neural network estimates tropical cyclone intensity from INSAT-3D satellite images; the abstract reports no quantitative accuracy.
A CNN trained on SEVIRI satellite imagery and DARDAR lidar retrievals provides a geostationary observational constraint on cirrus and mixed-phase ice cloud microphysics.
Autoencoder-based anomaly detection is proposed to flag faulty automatic weather station data for predictive maintenance, a low-novelty study with no quantitative results given.
Uses a neural network trained on reanalysis and cloud masks to inpaint cloud-obscured SEVIRI satellite dust plumes over North Africa, recovering up to 15% of summertime observations.
Over-parameterized neural networks predict GPM radar tropical Pacific rain-rate distributions from MERRA-2 profiles, capturing heavy tails better than other machine learning methods.
MLP retrieval of surface solar irradiance from MSG imagery can beat a physical method but depends strongly on training-set similarity, highlighting pitfalls.
Random Forest and ANN estimate hourly GNSS precipitable water vapour across Austria without meteorological inputs, with a single-hidden-layer ANN slightly outperforming other models.
An ANN estimates precipitable water vapour over Egypt from surface meteorology, reaching RMS 0.21 mm and outperforming eight empirical models including ERA5.
Classifies simulated radiometer bands into deep convective ice-storm cloud types using WRF digital twin data and ML classifiers, with over 80% accuracy on the tropical set.
DSRSS-Net jointly super-resolves and segments FY-4A imagery for snow-cover mapping on the Tibetan Plateau, improving classification accuracy by 4.45% over MOD10A1.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
U-Net models estimate aviation turbulence intensity at three altitude layers from GOES-16 ABI imagery; U-Net3+ matches pilot reports best.
Convolutional autoencoders with clustering classify 3D MRMS radar reflectivity profiles into 18 precipitation-structure clusters linked to HRRR environmental conditions.
Machine learning on SAR-derived features estimates atmospheric surface-layer Obukhov length, explaining about 80% of variance for convective scenes.
XGBoost trained on dual-polarized Sentinel-1 SAR collocated with SFMR and SMAP retrieves tropical cyclone surface winds with 2.53 m/s RMSE up to 70 m/s.
Neural network retrieves sea-surface wind speed from the MTVZA-GYa microwave radiometer, matching AMSR2 with 1.6 m/s RMS difference and mapping tropical cyclone winds.
A CNN (PCINet) identifies day and night precipitation clouds from FY-4A imagery, outperforming five segmentation baselines and enabling near-real-time monitoring.
Nanjing University of Information Science & Technology (NUIST)
Few-shot transfer learning with a novel Hough-space SR loss segments flight contrails in satellite imagery for climate-impact monitoring.
Deep learning framework fuses multi-source satellite precipitation products; abstract unavailable, so classification relies on the title.
CNN plus Bi-GRU classifies weather conditions from on-farm sky images to support farmers' decisions; no quantitative results given.
Machine learning model estimates urban Beijing planetary boundary layer height from ERA5 predictors.
Institute of Atmospheric Physics, Chinese Academy of Sciences
Self-organizing map clustering of geostationary cloud properties finds tropical cloud clusters with over 70% precipitation probability, e.g., CER ≥ 20 μm and CTT < 230 K.
Presents a RetinaNet detector, labeled dataset and annotation tool for mesoscale convective systems in Meteosat SEVIRI imagery, achieving mAP 0.75.
U-Net reconstructs radar reflectivity from WRF simulations and Himawari-8 satellite data with 5.38 dBZ RMSE, revealing scale-separated atmospheric features.
Institute of Atmospheric Physics, Chinese Academy of Sciences
Proposes eleven stacking and ensemble learners that blend PERSIANN and IMERG satellite precipitation with gauge data over CONUS, improving monthly accuracy.
A physics-constrained LightGBM model adds CLM surface energy balance inputs to remote sensing, giving more accurate and interpretable gapless land surface temperature.
End-to-end DNN, CNN and CNN-LSTM models detect clouds in FY-3D HIRAS infrared hyperspectral soundings, with CNN and CNN-LSTM reaching 0.96 accuracy.
A deep neural network turns FY-4B infrared imagery into simulated night-time visible images with R² of 0.99, aiding night cloud and fog monitoring.
Adaptive recurrent graph networks infer historic snow accumulation from deep ice-layer thickness in airborne radar data, outperforming prior models in polar ice records.
A dataset pairs GOES-16 ABI multiband images with CloudSat cloud-type labels over South America, enabling supervised cloud classification with a simple neural network.
A physics-constrained deep learning radiative transfer model over oceans recovers correct Jacobians, enabling radiance sensitivities for satellite data assimilation.
A U-Net reconstructs radar composite reflectivity over the ocean from satellite data, with interpretability showing cloud features matter most.
Institute of Atmospheric Physics, Chinese Academy of SciencesAlibaba (DAMO Academy)
Uses an artificial neural network with Earth observation data to produce high-resolution bioclimatic maps for Cyprus; abstract unavailable, so scope is inferred from the title.
ResNet50 classifier on an edge device classifies ARM Doppler lidar cloud and clear-air images with 97.6% and 94.7% accuracy.
Deep learning detects cold, warm, stationary and occluded fronts from ERA5 over the US, reaching CSI 0.60 for cold fronts and 0.71 for front/no-front.
Transformer detects concentric eyewalls in passive microwave imagery, producing a 30-year global dataset of 377 events with 92.9% precision and recall.
Neural networks trained on A-Train radar-lidar data add cloud height, layering and rain intensity to tropical cloud systems, enabling convective organization analysis for model evaluation.
An MLP-CNN network retrieves atmospheric temperature profiles from ground measurements and historical observations, with low bias and RMSE at all altitudes.
A neural network corrects sea-surface-temperature effects on HY-2A Ku-band scatterometer backscatter, making retrieved surface winds closer to WindSat and ECMWF.
Compares PCA, a convolutional autoencoder and a pretrained ResNet as latent representations of satellite images for classifying weather events; the autoencoder gives the highest threat scores.
An attention-based GAN reconstructs low-quality SAR wind speeds in tropical cyclones, halving RMSE versus the existing SAR product in affected regions relative to ECMWF.
XGBoost corrects precipitation gauge undercatch (51.4% at a 4147 m Tibetan site), beating traditional methods and transferring to Norway and the US.
Proposes compact U-Nets with Self-ONN layers for satellite active fire detection, reporting superior detection with reduced computational complexity.
Residual attention back-projection network super-resolves weather radar base echoes, outperforming compared CNN methods visually and quantitatively in echo detail and edges.
Reviews physics-informed and data-driven visual anemometry, inferring local wind from the motion of trees and flags for renewable energy and urban applications.
Massachusetts Institute of Technology (MIT)California Institute of Technology (Caltech)
Transfer-learned deep networks estimate western North Pacific tropical cyclone size from infrared imagery, producing a 1981–2017 DeepTCSize record with R of 0.76–0.85.
A sensor-adaptive leaf area index retrieval for FY-3D MERSI-II combines radiative transfer with a neural network backup and agrees well with MODIS (R2 0.85).
A physics-initialized inpainting network (PI-RFR) fills large gaps in Antarctic temperature and wind fields, improving on Kriging and IDW by 63 to 74 percent.
CNN estimates downward shortwave radiation over the ocean from all-sky RGB images, outperforming empirical parameterizations and MODIS-based methods.
Interpretable non-linear ML bias correction removes 3D cloud artifacts from OCO-2 XCO2, cutting unphysical variability by 31% over land and 55% over ocean.
CNN wind-speed estimator for Sentinel-1 SAR reduces rain-induced errors by 27% and 45% under rain above 1 and 3 mm/h, validated against buoy measurements.
Deep learning merges multiple precipitation sources over the Tibetan Plateau to produce improved precipitation estimates (abstract unavailable; assessed from title).
DRAIN, a U-Net on GPM passive microwave brightness temperatures, retrieves rain rates with performance similar to or slightly better than the official GPROF algorithm.
Introduces the 6,691-image WEAPD dataset of 11 weather phenomena and benchmarks five AI classifiers for image-based weather recognition.
Compares five ML anomaly detectors for cleaning noisy IoT sensor data to improve local weather forecast reliability; the abstract gives no headline numbers.
Neural network and Gaussian process models fill gaps and bias-correct ERA5-Land meteorological series against station data, reducing NRMSE to 2–11%.
Combines polynomial chaos expansion and Gaussian process regression to estimate and forecast wind fields from ADS-B aircraft-derived data, with confidence intervals.
CNN corrects MRMS radar precipitation estimates over complex western US terrain, beating radar QPE on several atmospheric river events in a year-long evaluation.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Applies LightGBM to probabilistically merge contiguous-US gauge data with PERSIANN and IMERG satellite precipitation, outperforming quantile regression forests at extreme quantiles.
An energy-function merging method fuses downscaled GPM IMERG with gauge data to produce 1 km monthly precipitation over China, outperforming IMERG in cross-validation.
A conditional GAN retrieves CloudSat 94 GHz radar reflectivity from MODIS cloud products over global oceans, performing best for deep convective clouds.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Cloud-MobiNet, a lightweight MobileNet-based CNN, classifies ground-based cloud images from the CCSN dataset with 97.45% accuracy for smartphone use.
TC-Resnet estimates tropical cyclone 34-kt wind radii from infrared and microwave satellite data with MAE of 11.3 nmi and r of 0.907.
Deep learning reconstructs monthly surface air temperature over the Tibetan Plateau for 1901–2020, filling gaps in the historical station record.
YOLOv3 detects tropical cyclones in IASI infrared sounder radiances over the North Atlantic, a proof of concept with AP 78.3% at IoU 0.1.
Localized statistical retrievals from IASI sounder data reduce dampening and inflating errors in extreme land surface temperature retrievals, from 1.9 K to 1.6 K.
Support vector regression and hybrid neural networks estimate 2 m air temperature from MODIS land surface temperature and vapour pressure, reaching about 2.25 °C accuracy.
Random forest retrieves Himawari-8 AHI water vapor 35–45% more accurately than 1DVAR at much lower cost, demonstrated for Typhoon Maria.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
A 3D deep neural network sharpens AIRS/AMSU temperature and humidity retrievals, better capturing planetary boundary layer top inversions and improving PBL height estimates.
Bayesian neural network retrieves 0.7–6 km tropospheric temperature profiles from Raman lidar with per-profile uncertainty, agreeing with the standard algorithm to within 0.05 K below 4.5 km.
A CNN classifies six cloud types in ground-based sky images from the SWIMCAT-EXT dataset for use in solar radiation forecasting.
U-Net with ResNet18 segments clouds in horizon-camera images for solar irradiance nowcasting, reaching mean IoU 0.6 with class imbalance limiting performance.
Nested genetic-algorithm and machine learning framework optimizes rain-gauge network design for spatial average rainfall estimation under missing data; the abstract is unavailable.
A physics-informed CloudMask loss improves deep networks that reconstruct 3D multilayer cloud masks from 2D satellite imagery over binary cross-entropy.
MF-ANN retrieves ocean surface wind speed from CYGNSS GNSS-R data with feature selection and outlier filtering, cutting RMSE by 29% over a single ANN.
Feedforward neural network detects optically thin cirrus in IASI-NG simulations, improving detection probability by 3% over IASI with false alarm ratio falling to 0.01.
Four machine learning methods recalibrate FY-3A MERSI near-infrared precipitable water vapor against GPS stations, cutting RMSE by 58–72% over China.
Deep CNNs with dense blocks estimate hourly rainfall from polarimetric WSR-88D radar, reaching MAE 1.58 mm and beating traditional Z-R relations.
A CNN-LSTM retrieves global ocean surface wind speed from CYGNSS GNSS-R data, reaching 1.34 m/s RMSE against ERA5 and beating CNN by 6.3%.
A CNN attention fusion network combines MODIS and ERA5 data to produce gapless all-sky land surface temperature maps, outperforming four state-of-the-art methods.
Adversarial data-to-data translation converts geostationary satellite albedo and brightness temperature into radar-like rainfall over Korea, outperforming GK2A, PERSIANN-CCS and IMERG products.
Bagged-tree ensembles retrieve ocean wind speed from CYGNSS GNSS-R, reducing RMSE by about 29% and 65% versus the Level 2 product, including typhoon cases.
DenseNet-based fusion of infrared and water vapor satellite images, with a tropical-cyclone-specific loss, improves TC center location accuracy.
Attention network with Kalman filter estimates tropical cyclone intensity from multispectral infrared images, cutting RMSE from 9.79 to 7.82 knots.
Random forest and deep neural network reclassify snow versus cloud in GK-2A geostationary imagery, raising F1-score from 73% to about 77-78% against ground truth.
Compares random forests, gradient boosting and XGBoost for merging PERSIANN/IMERG satellite rainfall with gauge data over the contiguous US; XGBoost performs best.
A single-layer perceptron on commercial microwave-link attenuation cuts rainfall-rate test MSE from 0.2 to 0.1 mm² h⁻¹, beating the ITU R-k method.
Random forest fusing MODIS, GPM and GLDAS factors reproduces the SPEI drought index for Shandong, China, from 2002-2020, outperforming XGBoost and SVM.
Comparison of eight machine learning algorithms correcting PERSIANN satellite precipitation against GHCN gauges over CONUS finds XGBoost and random forests most accurate.
XGBoost retrieves cloud-top height from Himawari-8 imagery with RMSE 1.74 km versus CALIPSO and corrects JMA's large biases for ice and multilayer clouds.
Near real-time framework fuses per-source statistical and ML models to quality-assess crowdsourced third-party weather station observations, tested on synthetic and real network data.
Dense neural networks retrain passive microwave rainfall and snowfall retrievals from GPM and CloudSat, detecting precipitation with over 0.95 probability and showing large global snowfall uncertainty.
Reviews deep learning weather forecasting and classifies weather conditions such as rainy, sunny and foggy from images; the abstract offers no quantitative results.
Neural network retrievals from AMSR2 microwave data estimate all-weather surface longwave radiation globally, with R² of 0.82–0.89 against ERA5 over land.
A deep convolutional network classifies satellite cloud images from MOSDAC over India with 94% accuracy, aimed at tropical cyclone prediction.
A U-Net emulates the WSR-88D operational velocity dealiasing algorithm, giving accurate, fast radar dealiasing that ports to other radar systems.
CO2-independent neural-network cloud mask for IASI radiances agrees about 87% pixel-by-pixel with the operational IASI L2 cloud product, suited to climate records.
Convolutional Gaussian neural processes outperform non-stationary GPs for Antarctic air-temperature sensor placement, giving more informative station siting in a simulated experiment.
Neural networks reconstruct daily and monthly 780 m Greenland Ice Sheet air temperature from MODIS land surface temperature, reaching R of 0.96 against automatic weather stations.
Deep neural network retrieves cloud fraction from CrIS infrared sounder observations; abstract unavailable, so scope is inferred from the title.
Deep neural networks retrieve global ice water path from FY-3B MWHS microwave sounder, with a cloud filter; accuracy is good above 100 g/m2 and tested on tropical cyclones.
CNN fuses MODIS and ERA5 to estimate precipitable water vapour; abstract unavailable, so scope is inferred from the title.
Rotation-invariant autoencoder clusters 22 years of MODIS imagery into 42 AI-generated cloud classes (AICCA), revealing cloud-type trends linked to stratocumulus decline.
ML merges satellite and ground soil moisture, then SM2RAIN derives a rainfall product outperforming GPM, CHIRPS, ERA5-Land and MSWEP over the Heihe basin.
Survey of feature engineering, convolutional, LSTM and attention methods for spatio-temporal patterns in meteorological image sequences, illustrated with solar forecasting and convection detection.
Nine ML and DL methods estimate daily land surface net radiation from MODIS; deep learning generally beats classical methods, with ResNet most robust across conditions.
A DNN estimating atmospheric turbulence (Cn2) from scintillation patterns can be reused at new propagation distances via analytic or wave-optics scaling, avoiding retraining.
Unsupervised deep learning on MODIS radiances learns cloud features and produces physically meaningful cloud classes without predefined categories.
A rotation-invariant autoencoder with clustering yields AICCA, a 22-year atlas of 42 unsupervised cloud classes from 198 million MODIS image patches.
Feedforward networks refine global VMF1 zenith tropospheric delay over Nigeria using local GNSS stations, lowering external RMSE to 0.03–0.04 m.
Evaluates the ANN-based PDIR-Now near-real-time satellite rainfall product during California atmospheric river events, validated against profilers and disdrometers.
Develops machine learning models of atmospheric parameters from a 53 MHz stratosphere-troposphere radar in India, aiming to predict winds outside the training interval.
Transfer learning merges rain gauge and TRMM satellite data to improve daily precipitation estimates in a data-scarce region.
Backpropagation neural network corrects underestimated high winds from HY-2B/C/D scatterometers in tropical cyclones, cutting RMSE by 20.4% to 3.40 m/s.
Revised neural network retrieves OCO-2 XCO2 without the weak CO2 band, capturing downwind emission plumes absent from training data, though latitude and date leakage was found.
Gaussian processes with non-stationary Gibbs kernels reconstruct Upper Indus Basin precipitation patterns with uncertainty, outperforming stationary and deep GP baselines.
A Random Forest precipitation estimator using GOES-16 ABI infrared channels and RAP predictors over CONUS beats SCaMPR, with correlation 0.46 and HSS 0.91 vs 0.19.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Probabilistic neural GPROF-NN retrieves GPM surface precipitation, with a 3D CNN variant, improving GMI mean absolute error from 0.079 to 0.059 mm/h.
PSO-tuned extended RBF network estimates NWM zenith tropospheric delay accuracy against GNSS references, with average error better than 4 mm.
A deep learning method that combines data assimilation ideas retrieves sea-surface wind speed time series from underwater acoustics, with up to 16% lower RMSE than prior data-driven methods.
Pretrained CNNs classify clear versus cloudy sky from ground-based photos, with VGG-16 reaching 91.4% accuracy as a tool for cloud-cover observation.
Backpropagation neural network fuses GNSS precipitable water vapour with ERA5, cutting ERA5 RMSE by 18.5% in China and raising correlation to 0.99.
Neural network reconstructs a monthly 0.05° Antarctic air temperature product from MODIS and 119 stations, with MAE 2.23 °C and better agreement with in situ data than ERA5.
Gaussian process regression outperforms neural networks in estimating GPS integrated water vapour from surface meteorology in Hyderabad, with correlation 0.93.
F2F-NN, a CNN retrieving sea surface wind fields from FY-3D microwave imager data, reaches 0.26 m/s RMSE and beats five existing products in a hurricane case.
Compares seven ML methods for CYGNSS GNSS-R sea surface wind retrieval; LightGBM gives RMSE 1.419 m/s at low winds.
Tutorial introducing persistent homology and persistence landscapes for classifying mesoscale cloud organization in satellite images, with interpretable results and example code.
Machine-learning quality control for ABI water vapor radiances blends seven bands to reject surface-contaminated pixels, improving Hurricane Harvey track forecasts.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
CNN segments Sentinel-1 SAR images into three rain regimes learned from NEXRAD collocations, outperforming Koch's filters for high-resolution rain mapping over ocean.
TICAENet, a LeNet/VGG-style CNN, classifies typhoon intensity from geostationary satellite images at 97.12% accuracy and estimates wind speed with 4.78 m/s MAE.
Workshop report on machine learning for Earth system observation and prediction, outlining opportunities and challenges across satellite and forecasting applications.
Attention-U-Net estimates near-real-time precipitation from FY-4A satellite imagery, beating operational satellite products and baseline deep learning models on POD, CSI and RMSE.
Deep learning on Sentinel-1 satellite imagery yields the first multi-year global ship-track climatology, showing 2020 fuel rules cut ship-track frequency to decades-low levels.
Random forest regression with hourly compounding predicts wind speed from underwater noise, reaching MAE 0.62 m/s and correlation 0.96.
First deep-learning semantic segmentation of scanned 1940-2000 US Navy Antarctic Peninsula photos, reaching 74% accuracy on eight test images despite analog and scanning artifacts.
Slide deck on machine learning classification of cloud types in satellite observations and climate models; no abstract or quantitative results provided.
Neural network retrieves ocean surface wind, SST and near-surface air temperature and humidity from AMSR2, with heat flux RMSE of 18-40 W/m2 against buoys.
Nanjing University of Information Science & Technology (NUIST)
U-Net based CLP-CNN classifies cloud types from FY-4A L1 data over China in 0.9 s per 800x800 scene, matching Himawari-8 products 76.8% of the time.
Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
Compares CNN, CNN+decision tree and CNN+SVM for classifying five weather image categories, reaching 94% accuracy with the CNN+SVM fusion.
Builds the first open 10 m building height map of Ireland from Sentinel-1/2 features; CNN is most accurate, but ridge regression is chosen for realism and low cost.
Deep-learning cloud-type classifier trained on MODIS and CloudSat labels is applied to coarse satellite data to enable process-oriented evaluation of clouds in climate models.
A VGG16-based CNN applied to wavelet covariance transform images retrieves planetary boundary layer height from lidar profiles without human intervention, validated against radiosondes and MPLNET.
Synthetic driving-simulator images increase CNN weather-condition classification training efficiency by up to 74%, addressing dataset bias in vision-based weather recognition.
A backpropagation neural network classifies cloud types from images with a 60.6% success rate on 92 training images.
Machine learning maps fine-scale urban air temperature from dense in-situ sensors, showing park cooling of up to 1.1°C during a 2019 heatwave.
LightGBM reconstructs all-sky 1 km MODIS land surface temperature using LDAPS analysis, improving cloudy-sky estimates over South Korea.
MARS regression fusing ERA5 and Meteosat cloud data estimates surface downward longwave radiation with 18.8 W/m² RMSE, beating the LSA-SAF operational formula.
A neural network retrieves total precipitable water from Meteor-M No. 2-2 MTVZA-GYa microwave data, with RMSE under 3 mm over water versus radiosondes and AERONET.
Deep learning detects polar lows in Sentinel-1 SAR images with F1 0.95, using a new public dataset labelled from ERA5; higher resolution improves detection.
A deep CNN identifies tropical cyclone fingerprints in satellite cloud images over the Northwest Pacific with 96% accuracy, with heat maps showing eye and rainband focus.
Neural network retrieves 13 years of IASI temperature profiles trained on ERA5, agreeing within 0.5 K of radiosondes between 750 and 7 hPa.
Neural network retrieves stratospheric temperature from single-footprint AIRS sounder data, cutting error versus operational products and resolving gravity waves.
Estimates daily Greenland Ice Sheet surface melt at 3.125 km from spaceborne radiometers using a neural network trained on surface energy balance data, matching regional climate models.
Compares linear regression and neural networks for filling gaps in micrometeorological station data in complex terrain, finding significant wind differences only in synoptic flow.
Self-supervised deep networks interpolate radar precipitation to 5-minute resolution, cutting error by up to 20% versus linear interpolation over southern France.
A deep neural network corrects bias in daily satellite precipitation estimates over tropical regions, using ground observations as the reference.
Stacking and EMOS-CSG merge four satellite precipitation products with gauges into a 1 km daily product with RMSE 4.33 mm in a basin study.
Vision Transformers with spatial self-attention detect rain and road surface wetness from freeway CCTV cameras, reaching 96.7% and 98.1% F1.
A Gaussian-uniform mixture density network interpolates crowd-sourced UK station temperatures while automatically rejecting spurious observations.
Stacked machine learning maps GNSS zenith delays to precipitable water vapor without meteorological inputs, achieving under 2.5 mm accuracy and 22-28% gains over GPT3.
Deep-learning pipeline with super-resolution and segmentation classifies cloud images; the abstract is missing, so scope is uncertain and significance is low.
A CNN with multi-task learning on geostationary IR images estimates western North Pacific tropical cyclone RMW and R34 with correlations of 0.95 and 0.93.
Mahalanobis transform K-near-means algorithm retrieves boundary layer height from micropulse LiDAR under cloudy conditions, matching radiosonde references better than three baselines.
Nanjing University of Information Science & Technology (NUIST)
TCR-GAN converts geostationary infrared images of tropical cyclones into passive microwave rainfall fields and releases a new benchmark dataset (TCIRRP).
An LSTM-based health monitoring framework for CINRAD-SA weather radar flags early faults from deviations between predicted and measured monitoring parameters, reducing false alarms.
Tsinghua University (NowcastNet)China Meteorological Administration (CMA)
SVM and general regression neural network models estimate western North Pacific tropical cyclone size parameters (RMW, R34, R50, R64) from infrared satellite imagery.
Tropical cyclone classifiers trained on simulated atmospheric data perform much worse on satellite observations, showing the simulation-observation gap must be bridged.
CNN retrieval from CYGNSS L1 delay-Doppler maps beats SVM and BP models for global sea surface wind direction, with filtered errors under 20°.
MANET fuses satellite images with date, season and location metadata to classify weather systems, clouds and land cover better than image-only models.
A CNN fuses LEO and geostationary satellite data into 2 km, 10-minute land surface temperature, with 1.73 K mean absolute error against LEO targets.
FY4-SR-Net uses transfer learning to super-resolve FY4 visible satellite images from 4 km to 1 km, cutting RMSE by 16.12% versus bicubic interpolation.
A machine-learning fusion of CERES, ERA5 and GLDAS yields a global hourly 0.25° longwave downward radiation product with R=0.97 against ground measurements over land.
A residual deep network fusing SAR backscatter, texture and morphology retrieves hurricane sea-surface winds up to 75 m/s with RMSE 1.72 m/s.
Transfer-learned Inception v3 identifies wind streaks in GF-3 SAR images to retrieve sea surface wind direction, reaching 92-95% accuracy with 9.12 RMSE against ECMWF.
An artificial neural network maps CubeSat TEMPEST-D microwave brightness temperatures to surface rainfall, reaching 0.78 structural similarity on independent storms against MRMS radar.
A ResNet extracts 500 m sea-surface wind fields from Sentinel-1 SAR over Venice, with 16° direction RMS against in-situ data and finer structure than ECMWF.
TCR-GAN generates passive microwave rainfall images from TC infrared imagery and releases the TCIRRP benchmark dataset for IR-to-rainfall prediction.
Bayesian neural networks map GPM microwave imager data to precipitation type with calibrated uncertainty and decompose satellite retrieval errors.
Neural networks map Landsat TOA reflectance to horizontal and inclined surface albedo over sloping terrain, achieving RMSE of 0.029 for HHSA against in-situ data.
A deep learning model learns the radar reflectivity-to-rainfall (Z-R) relation for Shenzhen, outperforming the traditional Z-R equation and matching a pure DL nowcaster.
Nonlocal residual CNN super-resolves weather radar echo images at x2 and x4, outperforming interpolation in quantitative metrics and edge detail.
UATNet, a Swin-transformer U-shaped network, classifies cloud types in Chinese meteorological satellite images on a new nine-class dataset (CRMSCD).
A CNN remaps FY-3D MWRI brightness temperatures to NWP grids, reducing observation minus background bias and spread in GRAPES 4D-Var.
Streaming tensor decomposition of ceilometer lidar aerosol profiles across a US network tracks mixed boundary layer height in near real time over long archives.
Random forest and neural network models on GOES-ABI imagery improve low-cloud detection in multilayer scenes, raising probability of detection from 0.685 to 0.815 versus the CCL product.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Uses a Bayesian neural network to fuse chemistry-climate model output with sparse satellite and sonde data, producing a continuous vertically resolved stratospheric ozone record with uncertainty.
Physics-informed CNN estimates pixel-wise fog visibility from single images via Koschmieder's law, trained on a 3,000-image synthetic AirSim dataset.
Neural network retrieval of temperature and humidity profiles from Meteor-M MTVZA-GY microwave data achieves about 2-3 K temperature RMSE against radiosondes.
Support vector regression corrects rain-affected Ku-band scatterometer winds over the tropics, reducing rain-induced bias against ASCAT as reference.
Hybrid quantum-classical CNNs generate synthetic weather radar from satellite and NWP inputs, proposed as a benchmark for quantum advantage. No quantitative results given in the abstract.
UWILD random forest classifies ice versus liquid particles in Southern Ocean clouds from 2D-S probe images with 95% accuracy, beating two existing methods.
A multitask CNN estimates global tropical cyclone precipitation from HURSAT-B1 imagery, reaching 0.68 detection probability and beating MLR and random forest baselines.
Nanjing University of Information Science & Technology (NUIST)
Spatial random forest downscales and calibrates IMERG satellite precipitation in Sichuan using gauges and covariates, outperforming GWR, ANN and kriging.
A deep neural network retrieves near-surface air temperature over South Korea from satellite land surface temperature and vegetation indices, reaching R of 0.98 and RMSE of 2.19 K.
Neural network retrieves 13 years of IASI temperature profiles, agreeing with ERA5 and radiosondes within 0.5 K between 750 and 7 hPa.
Machine learning estimates 24-hour cloud cover from ground-based sky images, with support vector regression giving RMSE 1.45 tenths against human observers.
InceptionV3, ResNet50 and VGG models classify 9081 optical satellite images into five weather events, with InceptionV3 reaching 92% average accuracy.
CHEESEHEAD19 land-atmosphere flux observations are presented as training data for machine learning to scale energy fluxes across heterogeneous surfaces to model grid cells.
TempNet uses residual CNNs to raise the temporal resolution of radar rainfall maps, outperforming optical-flow interpolation and a CNN baseline.
Global planetary boundary layer height retrieved from ICESat-2 and CATS lidar using a threshold method and a CNN, validated against MERRA-2 with good spatial agreement.
A two-step neural network retrieves aerosol parameters from OCO-2 O2 A-band spectra, improving XCO2 retrievals relative to the Level-2 standard product.
CNN bias correction of CMORPH satellite precipitation over US coastal mountains raises correlation from 0.55 to 0.88 against Stage IV gauge-radar data.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
U-Net-style CNN reconstructs 35 dBZ radar reflectivity from Himawari-8 infrared channels over North China, a proxy product for regions without radar.
Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
SVR, PCA-SVR and CNN models invert high sea-surface winds from CYGNSS GNSS-R data; the CNN reaches 2.71 m/s MAE for winds above 25 m/s.
A random forest hybrid estimates mesoscale convective system rainrate from FY-4A cloud-top temperature features, outperforming the operational FY-4A precipitation product.
Machine learning bias correction of billions of smartphone pressure observations cuts errors 43-57% and yields analyses 21-31% better than MADIS networks.
QPENet deep learning on dual-polarization radar, trained on 11 typhoons, beats DSD-based parametric QPE for rainfall estimation in South China.
Nanjing University of Information Science & Technology (NUIST)Colorado State University (incl. CIRA)China Meteorological Administration (CMA)
Neural-network retrieval VACOS uses MSG/SEVIRI thermal channels and ECMWF profiles to detect volcanic ash and retrieve cloud-top height and particle radius, day and night.
A feedforward neural network grid model of GNSS weighted mean temperature over China improves on GPT2w-Tm by about 8.6–10.9% in validation.
GAN with channel attention synthesizes night-time visible satellite images from infrared data and ECMWF NWP fields, using FY-4A observations.
Hybrid neural network for hyperspectral atmospheric correction that infers atmospheric state without auxiliary met data, achieving MAE under 0.02 for emissivity across 29 materials.
Deep neural networks retrieve temperature and humidity profiles from FY-3 MWHTS microwave sounder and bias-correct its observations, outperforming shallow networks in accuracy and stability.
A machine-learning classifier on GOES-16 ABI channels with NWP inputs identifies precipitation occurrence with 97% accuracy and convective types at 76%.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
A neural network on MODIS brightness temperatures detects Atacama Desert fog better than a cloud-height method (Heidke skill 0.56 vs 0.49), enabling a fog frequency climatology.
A multi-task CNN framework with consistency constraints and cross-branch attention improves Himawari-8 infrared precipitation estimation by 3–5% F1 across intensities.
XGBoost model estimates hourly 1-km air temperature across 13 northeastern US states, with RMSE about 1.6 K versus 2.5 K for NLDAS-2.
EfficientNet classifies road surface conditions from street and highway camera images into five classes at 90.6% validation accuracy, complementing station-based road weather data.
A deep learning regression retrieves radar composite reflectivity maps from Fengyun-4A geostationary satellite data, with R² of 0.88-0.96 and MAE of 0.3-0.6 dBZ.
Random forest fuses thermal and passive microwave data to produce 0.01° all-weather land surface temperature over China, RMSE 2-5 K against in situ data.
A neural network and gradient boosting combination retrieves cloud top height from passive microwave radiometers, with MAE 2.1 km and correlation 0.8 against radar truth.
Compares logistic regression, LDA, trees, SVM and random forests for precipitation-area classification from S-band radar, with spatial features helping.
A deep neural network fuses NUCAPS, ABI and RTMA data to cut low-level temperature and moisture profile RMSE by over 30% versus ERA5.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Institute of Atmospheric Physics, Chinese Academy of Sciences
Compares linear regression, KNN, trees and neural networks to reconstruct ERA5 wind fields from 1% of points, with k-means sensor placement improving accuracy.
Neural networks retrieve Arctic temperature and humidity profiles from FY-3D HIRAS under clear skies, closer to ERA5 than AIRS products.
Generalized regression neural network calibrates MODIS and ERA5 precipitable water vapor against GNSS over China, improving accuracy most at monthly timescales.
Hierarchical mixture-of-experts model estimates rainfall rate from automotive lidar point clouds, reaching RMSE 2.42 mm/h after filtering uncertain predictions.
Applies an artificial neural network to CYGNSS GNSS-reflectometry data to retrieve coastal surface wind speed.
Position paper proposing research agendas for AI in Earth science and remote sensing to move beyond disciplinary comfort zones.
A convolutional network on one-minute GOES-16 visible and infrared imagery detects convection, trained against MRMS radar precipitation types as a proof of concept.
MeteCNN classifies 11 weather phenomena in a new 6,877-image dataset, reaching about 92% test accuracy.
Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Random forest fills cloud-caused gaps in satellite land surface temperature during heatwaves using land cover, elevation and vegetation indices.
Supervised classifier on 20 years of MODIS imagery maps Swiss mountain lake ice, finding freeze duration falling about 0.8 days per year.
Regression and neural network models using Jason altimeter ocean parameters estimate tropical cyclone 10 m winds up to 35 m/s with 1 m/s accuracy.
Random forest combined with Gaussian mixture clustering characterizes bias and standard error of water vapour atmospheric motion vectors, addressing a known gap in uncertainty estimation.
Rotation-invariant autoencoder (RICC) clusters satellite cloud imagery without predefined classes, with evaluation showing physically meaningful, spatially coherent and orientation-invariant cloud clusters.
Builds least-squares SVM regional zenith tropospheric delay models over Europe from GNSS and ERA5, with the best scheme reaching 9 mm RMSE.
A UNet segments airborne dust at night from GOES-16 infrared bands, aiming to support dust storm warnings for forecasters.
Builds a global 2005–2016 high-temporal-resolution precipitable water vapour dataset using a neural network to estimate atmospheric mean temperature.
Volume-to-point machine learning maps gridded QPESUMS radar sequences to Taiwan station rainfall, performing comparably to the operational scheme and excelling in heavy-rain detection.
Compares CDF-matching and neural network retrievals of ASCAT soil moisture for NWP assimilation, finding inversion better suited than forward modelling.
A CNN-BiConvLSTM reconstructs missing weather radar frames, outperforming interpolation, optical flow and ConvLSTM baselines across missing patterns.
Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
Machine-learning bias correction of 2016 smartphone pressure data reduces bias, and dense analysis resolves convective structure in a Beijing hailstorm better than station data.
Neural network replaces the eigenvalue solver in a discrete-ordinate radiative transfer model, with only a 3x speedup and generality retained; a proof-of-concept.
Random forest retrieves tropical overshooting cloud-top height from Himawari-8 infrared channels against CloudSat, reaching r=0.92 and RMSE 222 m versus 1305 m for interpolation.
Neural networks trained on satellite lidar, radar and reanalysis data map 3D radiative heating of tropical upper clouds, finding about 0.3 K/day net heating that enhances latent heating by 25%.
Regression, ANN and GWR corrections of IMERG late-run rainstorm precipitation in southern China slightly improve accuracy, with the best method varying by region.
Attention U-Net segments tidewater glacier calving fronts from SAR imagery end-to-end, matching U-Net with a 1.5% better Dice score and attention maps for tuning.
Gap-filling of satellite observations from reanalysis using imbalanced learning; no abstract was provided, so scope and results are inferred from the title only.
LM-BP neural network estimates clear-sky surface shortwave radiation from ASTER imagery with RMSE 27.34 W/m2 against site measurements.
Compares random forests, XGBoost, SVM and EM for imputing up to 40% missing precipitation at over a hundred stations; XGBoost combinations reach nRMSE 0.01–0.07.
Random forests generate gap-free daily 100 m and 500 m NDVI and land surface temperature products to support urban heatwave and heat-island assessment.
Survey of 2000–2018 satellite cloud classification methods, from thresholds to machine learning, recommending hybrid approaches combining ML, thresholds and physical parameters.
MIIDAPS-AI deep-learning retrieval of temperature, moisture and cloud from microwave and infrared sounders runs about 100 times faster with comparable accuracy to operational methods.
A supervised model on 16 Himawari-8 spectral bands classifies typhoon, cloud, sea, land, snow and ice with 96.5% accuracy.
3DeepCT, a neural network tailored to stationary cloud fields, reconstructs 3D cloud volumes from multi-view images with orders-of-magnitude faster inference than physics-based methods.
Statistical machine learning cloud masking for Proba-V satellite scenes provides per-pixel cloud flags to reduce errors in land and sea biophysical retrievals.
Transfers CNN features learned for one IASI physical variable to retrieve other atmospheric parameters, cutting retraining cost while matching models trained from scratch.
A recurrent convolutional network with temporal memory detects clouds in MSG SEVIRI image time series day and night with pixel-level confidence maps.
Quantile regression forest fuses 691 crowd-sourced citizen weather stations with open data to map Zurich urban temperature at 10 m resolution during a 2019 heatwave.
Random Fourier-feature kernel approximation enables fast retrieval of atmospheric parameters from IASI sounders, PROSAIL emulation and cloud detection on millions of samples.
Triplet-based ordinal deep learning estimates fog visibility from images, outperforming multiclass baselines on synthetic and real dense-fog datasets.
CNN classifies outdoor photographs into cloudy, rainy, shine and sunrise weather classes, comparing four CNN architectures with different depths.
U-Net-based deep learning improves segmentation of atmospheric cloud images from aircraft and satellite sensors, with claimed real-time capability for weather monitoring.
A fully connected deep network emulates CRTM clear-sky VIIRS thermal radiances over global oceans, with mean biases near 0.01 K and standard deviations under 0.1 K.
A CNN called GREMLIN fuses GOES-R ABI and GLM data into synthetic radar reflectivity, proposed to feed convective-scale data assimilation and latent heating initialization.
CNNs infer airmass properties such as moisture and instability from GOES-R ABI imagery, offering an automated alternative to sparse radiosondes for NWP and forecasters.
Neural network imputation of missing and anomalous station temperature and humidity data outperforms interpolation and other supervised methods for building simulation inputs.
Deep CNN approach detects polar mesocyclones and polar lows in satellite mosaics, addressing class imbalance and scale filtering for satellite data.
CNN retrieves near-real-time precipitation from FengYun-2 satellite imagery over Xinjiang, China; no abstract was available, so this summary is based on the title.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
PERSIANN-CWB uses a deeper convolutional autoencoder CNN on Himawari-8 infrared and water vapor channels, improving rainfall retrieval over Taiwan relative to the prior CNN version.
Semantic segmentation of VIIRS, MODIS and webcam images maps Swiss lake ice cover and ice-on/off dates, reaching mIoU above 93% for satellites and releasing a webcam benchmark.
A transfer-learned EfficientNet-b6 CNN classifies airborne cloud-probe ice-crystal images with 98% average precision, supporting cloud microphysics research.
Uses an enhanced back-propagation network with Gaussian activation and Levenberg-Marquardt training to calibrate an atmospheric pressure sensor, cutting MSE from 5.93 to 0.77 on a microcontroller.
PDIR-Now provides hourly quasi-global infrared-based precipitation at 0.04° with 15-60 min latency, improving the critical success index to 0.53 from 0.47 over PERSIANN-CCS.
Compares six ML models for near-surface air temperature from FY-4A satellite data; XGBoost gives R² 0.902 and RMSE 1.946 °C.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Input-space clustering scheme estimates retrieval uncertainty for neural-network IASI temperature profile retrievals, useful for NWP assimilation.
EfficientNet encoder with UNet decoder classifies satellite shallow-cloud organization patterns, scoring a Dice of about 66% on a Kaggle benchmark.
A deep neural network dynamic radiometric bias correction for ATMS radiances in MiRS improves temperature and humidity profile retrievals over ocean.
A Faster R-CNN detects global cyclones and anticyclones from ASCAT scatterometer wind fields, reducing false alarms and tracking Hurricane Irma.
A DNN trained on principal components of CrIS hyperspectral spectra, with VIIRS cloud masks as truth, detects cloudy fields of view with 93% accuracy.
U-Net and fully connected networks classify four rain types from GPM GMI brightness temperatures, reaching 0.95 total accuracy against DPR labels.
A neural network quality-control filter for CYGNSS ocean wind retrievals reaches over 75% outlier detection with under 20% false alarms for 5–18 m/s winds.
XGBoost corrects MAIAC satellite column water vapor errors, cutting RMSE by 26.9% for Terra and 16.5% for Aqua against AERONET sites.
Convolutional autoencoder corrects pixel-wise bias of PERSIANN-CDR daily satellite precipitation against APHRODITE over the Mekong, outperforming a standard-deviation method.
A deep learning framework reconstructs continuous space-time environmental fields from irregular station measurements using temporal basis functions and gridded spatial coefficients.
Triple collocation merges CMORPH, PERSIANN and ERA5 precipitation over China, improving on each input product against rain-gauge data; the method is statistical rather than ML.
XGBoost retrieves rainfall from radar reflectivity at multiple elevation angles, beating Marshall-Palmer and regression baselines for typhoon rainfall in Taiwan.
Random forest and SVM classify polar stratospheric clouds from MIPAS infrared limb spectra, with random forest the most promising and less prone to overfitting.
Karlsruhe Institute of Technology (KIT)Forschungszentrum Jülich / Jülich Supercomputing Centre
Machine learning maps gridded radar reflectivity volumes to hourly station precipitation in Taiwan, outperforming the operational QPE scheme for heavy rain.
Random forest and other supervised classifiers combine weather radar and NWP data to classify winter precipitation type, beating a temperature and thickness threshold baseline.
Gaussian process regression on water vapor isotopes estimates shallow convective moistening over the SE Pacific better than conventional fields alone.
Artificial neural networks retrieve temperature and humidity profiles from FY-4 GIIRS hyperspectral radiances, outperforming a simulation-trained network and matching AIRS consistency.
Three-stage random forest and XGBoost ensemble estimates hourly 1 km air temperature over Israel from geostationary satellite skin temperature and station data.
A neural network removes rain-induced bias from HY-2A scatterometer backscatter, using ECMWF winds as reference, so retrieved sea-surface winds deviate much less in rain.
Deep learning first classifies marine atmospheric duct type from sparse propagation factor measurements, then estimates refractivity parameters for real-time radar-performance applications.
Random forest fuses AMSR2 microwave data with MODIS annual-cycle terms to produce all-weather 1 km land surface temperature over South Korea, including cloudy days, with daytime RMSE of 2.5 °C.
SegCloud, an encoder-decoder CNN, segments ground-based all-sky camera images into cloud masks and estimates cloud cover better than traditional methods.
A CNN converts GOES-R radiances and lightning into synthetic radar reflectivity, enabling convective radar assimilation with much better skill than pixel-wise methods.
Multi-task CNN classifies and segments cloud organization types in single colour satellite images, learning from coarse noisy masks and ranking in the top 3% of a competition.
A CNN retrieves atmospheric boundary layer height automatically from LiDAR signals on real, non-ideal data, enabling real-time sensing of aerosols.
Random forest and CNN framework corrects GPM DPR orographic precipitation retrievals using HRRR, cutting false alarms by 77% over the Southern Appalachians.
Random forest cloud mask from GOES-16 radiances matches the operational GOES-R clear-sky product and aids solar irradiance forecasting.
Two-step classifier and regressor impute missing 30-minute rainfall at 37 UK weather stations, outperforming a neighbour-based surface fitting baseline.
Two-stage random forest hybrid estimates gridded 1 km intra-daily near-surface air temperature over Israel from satellite skin temperature, with out-of-sample R2 of 0.96 for air temperature.
CNN applied to audio from a low-cost acoustic sensor classifies rain intensity into seven classes, reaching 75% accuracy and 93% when adjacent-class errors are ignored.
Proposes a radar-mosaic CNN for estimating hourly typhoon rainfall in Taiwan, outperforming Z-R relationships at three stations; it reduces relative RMSE to 0.713 at Hualien.
Proposes a task-based graph convolutional network for ground-based cloud classification and releases a large labeled sky-image database (GRSCD).
Trains machine-learned models to estimate lightning distance from single-receiver VLF/LF sferics, reaching 53 km RMSE on observed Kansas data.
DeepLab-based models segment lake ice in webcam images with about 70% IoU across unseen cameras and winters, and introduce a new benchmark dataset.
Fuzzy-logic ALPHA algorithm fuses satellite, NWP and radar data to detect high ice water content for aviation, with AUC 0.85 against in situ measurements.
A deep CNN segments lake ice from Sentinel-1 SAR imagery across three Swiss lakes, reaching mIoU above 90%, independent of cloud cover.
DeepCTC neural network classifies cloud types from GOES-16 ABI imagery using CloudSat labels and evaluates PERSIANN-CCS precipitation by cloud type against MRMS.
Applies machine learning to satellite fire radiative power to improve fire representation in smoke and weather forecasting models; abstract not available.
A deep belief network fuses satellite, station, simulation and socioeconomic data to map daily maximum air temperature across China at 0.01°, achieving RMSE 1.996°C.
Neural networks trained on micrometeorological features improve eddy covariance flux estimates at an alpine lake on the Tibetan Plateau, with stable 10-fold cross-validation bias.
Random forest automates quality control of tail Doppler radar data from Hurricane Michael's rapid intensification, classifying over 99% of radar gates correctly.
Two-step deep detection and segmentation locates tropical cyclone centers in satellite infrared images with about 0.237° average position error.
Retraction notice for a CNN-based ground-based cloud image recognition system that duplicated prior work; the underlying topic is cloud classification.
Random forest downscaling of Landsat LST fused with MODIS tracks Chengdu summer urban heat island intensity rising from 2.78 °C to 4.04 °C over 2009–2018.
Compares multivariable ML methods for downscaling coarse nighttime urban land surface temperature; daytime visible-NIR predictors perform as well as physical variables, MAE within 2 K.
Himawari-8 brightness temperatures and NWP predictors estimate summer air temperature over China; a neural network reaches R=0.97 and RMSE 1.72°C against stations.
Logistic regression over land and extra-trees over sea detect clouds in FY-4A GIIRS hyperspectral infrared data, reaching HSS 0.78 land and 0.95 sea.
An unsupervised neural network finds hierarchical trade-wind cloud organization regimes in 10,000 GOES-16 images, linking regimes to radiative properties for climate model evaluation.
Uses a Gaussian process to decide when to release sensors from a weather balloon, targeting better stratospheric analyses and forecasts, tested on real flights.
A deep neural network applied to lidar backscatter images identifies planetary boundary layer height and outperforms a double-derivative method in preliminary results.
Review arguing modern AI can augment or replace parts of satellite remote sensing and NWP processing chains, from observations to forecasts.
Colorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
PERSIANN-CNN applies CNNs to geostationary infrared and water vapor channels, beating PERSIANN-CCS by 54% in CSI and cutting RMSE by 37%.
Deep belief network retrieves horizontal visibility from MODIS aerosol optical depth and ECMWF data over Eastern China, with validation correlation near 0.70.
Dynamic clustered Bayesian model averaging merges four satellite precipitation products over Pakistan, outperforming each input product against gauges.
Backpropagation neural network interpolates GNSS zenith wet troposphere delay across Austria from meteorological station data at 40 reference sites.
PDIR, a neural-network infrared precipitation algorithm with climatology-based dynamic rain-rate mapping, improves real-time estimates over the western CONUS, including atmospheric river events.
Deep convolutional networks retrieve atmospheric profiles from satellite radiances as a statistical retrieval, though the abstract is unavailable.
Deep learning emulates a satellite remote-sensing surrogate model, reproducing its outputs with acceptable accuracy and often faster runtime than the surrogate.
A neural network infers gridded marine surface visibility over the Indian Ocean from satellite optical thickness and reanalysis, with mean error under 8%.
An MLP fuses geostationary and LEO satellite data with dual-polarization ground radar to improve precipitation estimates over Dallas-Fort Worth.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
K-nearest neighbour classifies convective and stratiform rain from Doppler radar using TRMM labels, achieving 78% POD for convective and 95% for stratiform.
GAN-based super-resolution of degraded weather radar echo images produces sharper reconstructions with higher PSNR and SSIM than classical methods.
A conditional GAN generates nighttime visible-band COMS/MI satellite images from infrared data, reaching correlation 0.88 with observed images in summer.
Neural network retrieves ocean surface wind speed from Gaofen-3 quad-polarization SAR, reaching 0.74 m/s RMSE versus 0.88 m/s for CMOD7.
Real-time deep learning detector locates typhoon eyes in images for meteorological information; the abstract is unavailable, so accuracy is unknown.
Nanjing University of Information Science & Technology (NUIST)
Deep optical-flow interpolation upsamples GOES-R ABI imagery from 15 to 1 minute, improving severe-weather temporal resolution over baselines.
Compares random forest, XGBoost and a deep neural network for retrieving total precipitable water from Himawari-8 infrared channels; the DNN performs best against radiosondes.
An LSTM estimates atmospheric wind from quadcopter motion without a wind sensor, with lower errors than the wind-triangle method in simulated turbulence.
CNN classifies weather conditions into six classes from photographs and derives temperature estimates from class scores; the approach is not benchmarked.
A conditional GAN reconstructs CloudSat-style cloud vertical structure from MODIS inputs, capturing multilayer clouds and providing uncertainty estimates.
An optical instrument uses a random forest to classify cloud particle phase, achieving under 5% misclassification above 3 µm in lab tests.
Neural-network emulators of radiative transfer models speed up imaging-spectroscopy retrievals for greenhouse gas and coral reef monitoring by a multifold factor.
Crowd-labelled about 50,000 shallow cloud clusters in satellite images to train deep learning that maps four organization patterns and builds global climatologies.
Max Planck Institute for MeteorologyTechnical University of Munich (TUM)
Coupled CNN-RNN estimates local wind speed (0.75-11 m/s) from video of flags and trees, matching cup anemometer measurements.
California Institute of Technology (Caltech)Stanford University
Applies Gaussian process regression to infer marine boundary layer electromagnetic duct height from radar propagation factors, including noisy-input handling.
Builds an ensemble of anomaly detectors with adaptive thresholds for meteorological station time series, integrated into a road weather information system.
Bayesian model using SAR spectral parameters retrieves ocean surface wind with about 1.7 m/s speed and 49° direction standard deviation against ECMWF.
A light-weight CNN, CloudSegNet, segments clouds in ground sky-camera images across day and night in one framework, reaching state-of-the-art results on public databases.
Neural network combined with 1D-variational retrieval of FY-3C microwave temperature and humidity profiles during Hurricane Matthew, improving WRF data-assimilation track forecasts.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Multilayer perceptron retrieves evaporation duct heights in the marine boundary layer from sparse electromagnetic propagation measurements, enabling fast near-real-time estimates; architecture choices are documented.
Neural networks with selected geophysical inputs convert EarthCARE clear-sky TOA radiance to flux with 2.7–4.0 W m-2 error, under the 10 W m-2 mission threshold.
Neural network classifies typhoon low-altitude turbulence from Hong Kong terminal Doppler radar, validated against commercial aircraft turbulence measurements.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Retrieves temperature and humidity profiles from NPP CrIS hyperspectral radiances using autoencoders and a self-normalizing network trained on ECMWF data.
Proposes P_SegNet and NP_SegNet, modified SegNet architectures for cloud segmentation in remote sensing imagery, with improved accuracy over FCN baselines.
Builds a sky-camera and weather-station system and uses a CNN to classify sky condition with 97.8% accuracy for local short-term prediction.
Compares neural networks, bagged trees and random forest for filling missing station climate records in Minas Gerais, Brazil; random forest performs best.
Machine learning merges FLUXNET eddy-covariance towers with remote sensing and meteorological data into a 147-product global land-atmosphere energy flux database with uncertainties.
Random forest and SVM estimate summer Arctic air temperature from AMSR2 microwave data and buoys, reaching R² of 0.84–0.88 and RMSE of 1.31–1.53 °C.
Random forest estimates summertime precipitation from Himawari-8 infrared brightness temperatures plus GFS predictors, reaching 0.87 classification accuracy and 2.0 mm/h RMSE against GPM.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Adversarial color-space adaptation from synthetic to real cirrus cloud images, improving cloud segmentation by 6.59% for climate-related image analysis.
An ANN fuses HY-2A radiometer and scatterometer data to retrieve typhoon wind fields, outperforming operational HSCAT winds against GRAPES and H*Wind references.
An ANN combined with a particle filter retrieves ocean surface wind speed from BeiDou GEO GNSS reflections, reaching about 1.9 m/s RMSE against buoys.
A deep neural network retrieves high-latitude rain and snow from GMI and MODIS data, largely outperforming GPROF on snowfall and reducing rain underestimation.
NASACalifornia Institute of Technology (Caltech)Tsinghua University (NowcastNet)University of California, Los Angeles (UCLA)University of Oklahoma (AI2ES)
Quantile regression neural networks give case-specific posterior uncertainty for Bayesian retrievals, applied to MODIS cloud top pressure with consistent uncertainty; code released.
New open dataset of rain, snow and fog images with a super-pixel augmentation method, benchmarking ten CNN architectures for camera-based weather classification.
Compares generalized linear models and neural networks for rainfall retrieval from geostationary IR data, using OPERA radar as truth; neural networks reduce false alarms.
Estimates convective rainfall from GMS-5 infrared images and NWP products with an artificial neural network, validated against gauges (r=0.63 hourly) and gaining about 2 h lead time for flood forecasting.
Neural networks retrieve MODIS cloud-top height using CALIOP training labels, cutting mean absolute error by at least 33% against operational algorithms.
Bottleneck-channel neural compression of 8461 IASI channels matches NAPCA for temperature, water vapor and ozone retrievals while preserving physical meaning.
Back-propagation network retrieves sea surface pressure from ATMS microwave radiances with 2–3.5 hPa RMSE across clear, cloudy and rainy skies.
Machine-learning detection of rainfall from meteorological sensor data reaches an 87.4% true detection rate with a 32.2% false alarm rate.
Gaussian image entropy with piecewise stationary time series estimates expressway visibility from video, reaching 99.14% of relative errors under 10%.
Deep learning classifies cloud particle images from digital in-line holography to enable labeling of large datasets on precipitation initiation.
Shape and motion features with a trained classifier detect comma-shaped cyclone-associated cloud patterns in satellite imagery to support severe weather forecasting.
Neural network retrieves snow and liquid water path and water vapour from airborne millimetre radiometer data; submillimetre channels halve snow retrieval error.
A neural network method fuses spaceborne and ground radar reflectivity into three-dimensional fields. The abstract is missing, so the assessment relies on the title.
Nanjing University of Information Science & Technology (NUIST)
CNN and RNN models estimate ambient air temperature from outdoor images alone, with the RNN outperforming prior methods on image sequences.
Machine learning corrects smartphone barometer biases using ancillary sensor data, cutting average pressure bias by 82% and supporting use in NWP.
A DeepLab-based CNN trained on large-eddy simulation fields segments cloud base from total sky imager images to improve boundary-layer cloud cover measurement.
ContrailMod, a CNN applied to Himawari-8 imagery, outperforms a conventional contrail detection algorithm and is used to evaluate potential contrail coverage from ERA-Interim over South China.
Nanjing University of Information Science & Technology (NUIST)
Machine learning error correction makes low-cost Mini-AWS pressure readings meet WMO ±0.1 hPa accuracy, offering a cheap complement to automatic weather stations.
A stacked denoising autoencoder bias-corrects PERSIANN-CCS satellite rainfall, cutting precipitation-rate bias by 78–98% and false alarms by 33–43%.
Early neural network method that estimates precipitation rates from remotely sensed satellite cloud and infrared information.
A NeuralGCM-based hybrid model trained on satellite precipitation cuts climate-model precipitation biases and beats the ECMWF ensemble at mid-range forecasts.
Introduces WeatherReal, a benchmark from global in-situ surface observations showing reanalysis-trained AI models diverge from station data for temperature, wind and precipitation.
FuXi-DA is a deep learning data assimilation framework that assimilates FY-4B AGRI satellite radiances, reducing analysis errors and improving forecasts.
Data assimilationChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Transformer nowcasts radar composites up to two hours ahead from geostationary satellite data, with lightning and infrared channels proving key features in severe weather.
A residual U-Net postprocesses RAP forecasts with satellite data, improving dewpoint profiles and CAPE by up to 40% with well-calibrated uncertainty for severe-weather forecasting.
Post-processing & calibrationColorado State University (incl. CIRA)
Gradient-boosted hybrid models combining macro-meteorological baselines with local observations cut optical turbulence error by 29% with one day of data and 68% with 180 days.
A neural network trained on tri-band VHF radiometric sounding data forecasts short-term ground-layer temperature, humidity and precipitation from daily measurements.
A deep-learning segmentation model with transfer learning from synthetic fields detects and characterises trapped lee waves hazardous to aviation, faster than spectral methods.
An XGBoost calibration of the VMF3-FC zenith wet delay forecast cuts bias by 93% and RMSE by 6% versus the operational product, validated on 492 radiosondes.
A generalised regression neural network corrects GRAPES and ERA5 weighted mean temperature, nearly removing bias and cutting RMSE by 19–21% for GNSS water vapour.
3D CNN forecasts precipitation 4 h ahead from FY-4A multichannel satellite imagery, beating NWP on critical success index (0.25 vs 0.21).
An unsupervised neural network manifold maps the continuum of mesoscale shallow cloud organisation, showing organisation alters albedo by about ±1.4% beyond cloud-fraction changes.
A 3D-convolution U-Net downscales SEVIRI land surface temperature to hourly 1 km WRF air temperature, with MAE 1.36 °C versus 1.49 °C for ridge regression.
LSTM nowcasts 10-minute rainfall fields reconstructed from 26 commercial microwave links, reaching correlation 0.91 at one minute and falling to 0.20 at ten.
Machine learning corrects WRF urban air-temperature biases per grid cell using 402 crowdsourced personal weather stations over London.
Random forest downscales daily IMERG precipitation from 10 km to 1 km using MODIS cloud properties, cutting RMSE by 29% against gauges.
Vision Transformer and CNN regressors estimate tropical cyclone intensity from satellite cloud images, and a fused two-stage strategy gives the best accuracy.
Neural network combined with an improved GPT3 model forecasts Japanese precipitation from GNSS-derived atmospheric data; abstract unavailable, so details are limited.
Random Forest spatial interpolation with satellite covariates downscales reanalysis and satellite air temperature and precipitation products over Turkey, with AgERA5 and MSWEP the best products.
WEY-TCNet combines wavelet fusion and YOLOX with channel attention to detect tropical cyclones in Fengyun-2D infrared images, reaching 91.5% mAP versus 86.7% baseline.
CatBoost model estimates western North Pacific tropical cyclone intensity from FY-2 cloud-top brightness temperatures, achieving 3.74 m/s RMSE and beating prior CNN approaches.
A deep neural network combining GOES cloud-top satellite data with reanalysis parameters estimates severe hail likelihood, reaching CSI 0.511 against radar MESH.
Neural network sensitivity analysis of observations separates aerosol primary convective invigoration (72% of cloud-top cooling) from meteorological effects in deep convection.
U-Net with ResNet18 segments clouds in horizon-facing sky images with IoU 0.6, aimed at short-term cloud motion and solar power nowcasting.
DeepLCZChange links airborne LiDAR urban structure to Landsat 8 surface temperature in New York, showing a cooling effect of urban forests.
Downscales GPM satellite precipitation from 0.1° to 1 km using a BP neural network with terrain and vegetation predictors, improving accuracy in complex terrain with sparse gauges.
Downscaling & super-resolutionChina Meteorological Administration (CMA)
UNet CNN fuses GOES-16 cloud observations with COAMPS forecasts to correct high-cloud forecasts at 3-12 h lead times, improving accuracy and sensitivity.
DownScaleBench uses an iterative super-resolution CNN to produce 300 m urban precipitation from 10 km GsMAP data over Austin, improving on cubic interpolation.
Downscaling & super-resolutionNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Dynamic-balance CNN estimates tropical cyclone intensity from satellite imagery with 8.32 kt RMSE, 35% lower than the Dvorak technique.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)
Neural-network downscaling of ERA5 daily wind speed to a 1 km China dataset for 2015, with R2 of 0.95 against station observations.
Compares ML downscaling of satellite precipitation, finding SRF and XGBoost best and GDA calibration beneficial, while residual Kriging degrades performance.
ML models using VIIRS retrievals and HRRR variables estimate dead fuel moisture content over CONUS, outperforming climatology by 24% in hourly RMSE and beating MODIS-based models.
Extremes & hazardsNSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Review of deep learning methods for extracting tropical cyclone intensity and wind radius from satellite infrared imagery, with an outlook on future directions.
Trs-net applies a 2-D CNN to multispectral satellite images to analyse and classify tropical revolving storms. No abstract was provided.
Rotation-equivariant convolution combined with a Transformer estimates tropical cyclone intensity from multi-platform satellite and environmental data, with RMSE of 8.19 kt.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)
Reviews IoT sensor-based weather monitoring and deep learning forecasting, including self-attention and hybrid CNN-LSTM models, with challenges and future directions.
Random forests using wind profiler divergence and vorticity predict rainfall onset 30 minutes ahead, outperforming inputs derived from ERA5 reanalysis.
NowcastingNanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Uses machine learning to map city-scale air temperature from sparse meteorological station data for urban climate adaptation. No abstract available; scope inferred from title.
WR-Net combines TV-L1 optical flow warping with a refinement network to interpolate and predict GK2A geostationary imagery from 4- to 2-minute intervals.
Curates an hourly-interpolated satellite image and wind-speed dataset for tropical cyclones, plus a community competition whose winners improved on the benchmark by about 50%.
Pred-SF combines PredRNN-V2 with a stepwise refinement network to predict 4-hour precipitation for a target area from ERA5 multivariate fields and GPM data.
Proposes MS-ConvLSTM, adding echo top, VIL and radar-retrieved winds to reflectivity inputs to improve precipitation nowcasts over conventional deep learning baselines.
NowcastingNanjing University of Information Science & Technology (NUIST)
Review of deep learning for weather radar analysis covering beam blockage correction, echo extrapolation and precipitation nowcasting, with open challenges.
NowcastingNanjing University of Information Science & Technology (NUIST)
Neural network applied to MOSDAC satellite imagery for cyclone identification and monitoring, reported only qualitatively without a quantitative benchmark.
Tuned fusion deep learning model forecasts tropical cyclone passive microwave rainfall images from satellite data, with no quantitative results reported in the abstract.
Downscales GPM IMERG precipitation from 10 km to 1 km over Taiwan using decision trees, SVR and random forests with NDVI and terrain predictors, validated against gauges.
Bi-LSTM tuned by nomadic people optimisation detects and tracks tropical cyclones; the abstract is missing, so performance cannot be confirmed.
Pre-training on a suitable pretask improves U-Net generalization for segmenting unseen natural hazards from satellite imagery across regions without target labels.
Convolutional autoencoder bias correction of TRMM and PERSIANN-CDR satellite rainfall over the Mekong basin, using APHRODITE gauge data as benchmark.
A CNN detects clouds in camera photos to build a forward operator, enabling convective-scale assimilation of cloud cover into the ICON-D2 model; proof of concept over 6 days.
Multi-task learning models nowcast convective storm vertically integrated liquid and lightning using satellite and weather radar observations; abstract not provided.
NowcastingNanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
A neural network classifies satellite cloud morphology, showing that closed-cell cloud changes contribute about 0.06 W/m²/K to global shortwave cloud feedback under CO2 quadrupling.
Climate analysis & attributionNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)NSF NCAR / UCARUniversity of Washington
YOLOv7 detects smoke columns in camera imagery for early wildfire warning, reaching F1 of 0.74 at confidence 0.298.
Random forests linking direct upward lightning tower measurements to reanalysis meteorology identify drivers of lightning risk at wind turbines, with skill on independent data across Germany.
Multi-input CNN with wavelet-decomposed satellite imagery estimates Atlantic hurricane maximum wind speed with 15-knot RMSE.
Deep learning network detects tropical cyclones; no abstract is available.
Differentiable operator enforces exact conservation of low-resolution values in CNN super-resolution, tested on radar, satellite and climate model data.
NFC-Net fuses FY-4A satellite, DEM and ERA5 data with a UNet to correct ECMWF forecasts, cutting 2m temperature and 10m wind speed RMSE by about 50%.
Post-processing & calibrationNanjing University of Information Science & Technology (NUIST)
Wavelet-LSTM-ARIMA hybrid predicts GNSS precipitable water vapour, cutting RMSE by about 55% versus other models over 3 h horizons.
CWNNet, a ConvRNN fusing Fengyun-4A satellite and radar inputs, predicts 0-1 h reflectivity and convective initiation, outperforming single-source models (POD 0.87 at 1 h).
Neural style transfer makes simulated cloud images resemble satellite data, raising tropical cyclone precursor classification performance by 40–90% at 3–7 days lead.
Gradient boosting and attention-LSTM models predict cirrus ice water content and crystal number (R² = 0.49), with SHAP attributions linking dust aerosols to ice crystal changes.
SmokeyNet fuses wildland camera imagery with weather sensor data to detect smoke, with a few-minute time-to-detection for early wildfire alerts.
TSRC, a residual convolutional network using previous-frame cues, improves 3-hour radar-based precipitation nowcasts over China versus optical flow and UNet, including typhoon cases.
A random forest on a modified MOD14 scheme with Himawari-8 AHI data detects early-stage forest fires in Australia at about 90% precision and recall.
CSIP-Net, a convolutional network that predicts FY-4A infrared satellite images, reaching 4.77 K mean error at 2 h and 0.855 accuracy for precipitation-area detection.
Affinity Propagation clustering on a 2009-2021 Sicilian rainfall dataset identifies extreme rain events, including anomalous recent events near Catania.
Graph convolutional network estimates tropical cyclone intensity from satellite imagery; the abstract is missing, so details are inferred from the title.
Lightweight multitask CNN estimates tropical cyclone intensity and size from satellite data, with intensity RMSE of 8.40 kt, 33.5% below the ADT baseline.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)
Causal machine learning on satellite observations estimates aerosol-cloud interaction with uncertainty bounds, finding only one of three climate models reproduces the cooling trend.
LSTM forecasts Himawari-8 brightness temperatures and random forest converts them into 10-minute rainfall estimates over Indonesia, with RMSE of 1.54 mm per 3 h.
Reviews 2021 extreme events and barriers to deploying ML models on Earth observation imagery for disaster damage detection, highlighting InSAR as a tool.
Extremes & hazardsColumbia University (LEAP, Lamont-Doherty)
Doubly sparse spatiotemporal Gaussian process builds a probabilistic European climate-field reconstruction from the Last Glacial Maximum to mid-Holocene, combining simulations with pollen proxies.
Deployed camera-based system detects wildfire smoke with a high true-detection rate and alerts faster than human observers.
A convolutional autoencoder segments clouds in sky images above 90% pixel accuracy, and particle tracking predicts cloud motion for short-term PV power.
Machine learning replaces the radiative transfer observation operator for AMSU-A radiance assimilation in NICAM-LETKF, performing near RTTOV but slightly worse with simpler bias handling.
A neural network trained on ERA5 and MODIS skin temperature quality-checks IFS physiographic upgrades, improving land-surface temperature prediction by 0.37 K for updated lakes.
Deep learning estimates 34, 50 and 64-kt tropical cyclone wind radii in four quadrants from geostationary IR images, cutting errors 12-36% over prior methods.
Scattering-transform features with channel attention from multi-band satellite imagery estimate tropical cyclone intensity and predict lightning occurrence.
Neural network with a spatiotemporal encoding module estimates tropical cyclone intensity from infrared satellite imagery.
Classification algorithms detect and predict Alpine foehn events from GNSS troposphere products at Altdorf, reaching 70-85% detection and alarm measures.
Compares MLR and ANN downscaling of ERA5 and MERRA2 winds against South China tall-tower observations; ERA5 with ANN gives r=0.69 and RMSE 2.11 m/s.
Occlusion-perturbed CNN forecasts 10-minute-ahead irradiance from sky images, reaching 5.62% nRMSE and a 2.77% nCRPS for probabilistic forecasts.
Machine-learning regression maps AROME proxy variables to pseudo MTG lightning flash extent, giving a learned observation operator for storm-scale EnVar assimilation.
Deep learning on Himawari-8 and FY-4 imagery identifies typhoon vortices (88.6% accuracy above strong tropical storm grade) and estimates intensity with 3.8 m/s MAE.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Iterative SRCNN downscales 10 km JAXA GsMAP satellite precipitation to 300 m over Austin, Texas, beating cubic interpolation on PSNR and mutual information.
Downscaling & super-resolutionNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Detects tropical cyclones in South Pacific and Atlantic coastal imagery by combining optical flow estimation with a ResNet deep learning classifier. Abstract unavailable.
Combines all-sky camera images with satellite observations in one ML framework, improving intra-hour solar irradiance forecasts up to 60 minutes ahead.
A ConvLSTM encoder-decoder with spatial attention forecasts 24 h western North Pacific tropical cyclone tracks from satellite images, about 30 km better than ECMWF-HRES and GFS.
LightningCast uses a CNN on GOES-16 ABI imagery to predict next-hour lightning over CONUS, giving about 20 min or more of lead time and transferring to GOES-West.
Machine learning bias correction of simulated hourly precipitation, combined with quantile mapping, improves spatial frequency distribution but not temporal distribution of local precipitation.
Low-cost station sensors (temperature, humidity, rain and others) feed an ANN and decision tree that classify rain with scores of 0.92 and 0.96.
Random Forest nowcasts cloud-to-ground lightning over Italy and seas, finding that finest-resolution gridding of inputs gives the best skill.
Supervised classification of MODIS images builds the greenCuDb dataset, revealing that continental green cumulus form under similar large-scale conditions worldwide.
Feed-forward neural network driven by WRF rainfall output predicts rain attenuation fading on a geostationary satellite downlink during a dust-storm event.
Combines GOES-13 satellite image features with station meteorology in neural networks to estimate solar radiation; rRMSE stays above 20%, rated only fair.
Explainable ML with Shapley values on satellite and reanalysis data shows liquid water path rises nonlinearly with droplet number, with sensitivity dependent on precipitation and SST.
Climate analysis & attributionKarlsruhe Institute of Technology (KIT)
Faster R-CNN recognizes hail areas in Doppler radar imagery with over 86% accuracy and 97% precision on a test set.
U-Net nowcasting with a radar-prediction pre-training scheme and a new loss; heavy-rain CSI improves up to 95.7% and 43.6% at 5 h lead time.
Applies deep convolutional neural networks to improve tropical cyclone wind field estimates for extreme event analysis.
CNN with residual and attention blocks estimates tropical cyclone intensity from Himawari-8 cloud products, reaching RMSE 4.06 m/s on independent tests.
Extremes & hazardsColumbia University (LEAP, Lamont-Doherty)
Nonparametric neural-network test links 24-hour tropical cyclone infrared image sequences to rapid intensification events, with guaranteed type I error control.
Self-supervised network learns atmospheric dynamics by predicting temporal distance between ERA5 fields; the resulting distance loss improves downscaling and missing-data interpolation.
Introduces QpefBD, a minute-scale radar-based benchmark dataset for machine learning precipitation estimation and forecasting (abstract not provided; classified from title).
Benchmarks, datasets & evaluationChina Meteorological Administration (CMA)
An ANN trained on Kalpana IR and water-vapour satellite pixel values forecasts 24-hour precipitation at six Western Himalaya stations with a 0.78-0.95 probability of detection.
Meta-analysis of machine learning for dust storm detection and prediction, finding SVMs dominate satellite detection and SVM/random forests dominate meteorological-data prediction.
CNN and RNN nowcasters applied to IMERG satellite precipitation extend forecasts to 4.5 h, mitigating the latency of satellite rainfall products over CONUS.
Fully connected network emulates the CRTM microwave radiative transfer model for ATMS brightness temperatures over ocean, with mean biases of hundredths of a kelvin for data assimilation.
LSTM and CNN models using GNSS radio occultation profiles forecast regional wind fields over Beijing-Tianjin-Hebei, with about 1.4 m/s speed error and 1 s inference.
A hybrid method uses ML regression inside 3DVAR to retrieve fine-scale boundary-layer winds from lidar and in-situ data, improving 2-D and 3-D velocity estimates by 42.3% on average.
Physics-informed network downscales urban land surface temperature to 30 m using 3D point clouds, with errors under 0.2 K against satellite measurements.
3D-attention CNN estimates tropical cyclone intensity from multichannel IR, water vapour and microwave imagery, with RMSE 9.48 kt, 25% better than the Advanced Dvorak Technique.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)
Mask R-CNN detector, wind speed filter and CNN classifier detect tropical cyclones in satellite images with 97.10% precision on test images.
Competition-derived deep networks with oversampling, ensembles and cost-sensitive losses classify imbalanced cloud images for tropical cyclone precursors, improving performance by 65.4% over conventional methods.
MFSP-Net, a multi-input recurrent network, fuses precipitation grids, radar echoes and reanalysis for 0–4 h precipitation nowcasting, improving heavy-rain nowcasts over Southeast China.
A point-cloud network with attention estimates 3-D storm motion from radar volumes, trained on synthetic Perlin-noise data, with 5-15% lower endpoint error.
Machine learning downscales MODIS land surface temperature from 1 km to 30 m over Antarctic Dry Valleys using Landsat 8 as reference, achieving R² of 0.78.
A neural network using a microwave rain index and 24 h wind changes identifies tropical cyclone rapid intensification with near-perfect skill and reveals environmental drivers up to 1050 km away.
Extremes & hazardsNASACalifornia Institute of Technology (Caltech)
Simple machine learning models refine satellite-based advanced Dvorak technique features, cutting global tropical cyclone wind RMSE by 23–30% over 2017–2018 test years.
Deep belief network fuses satellite, model and station data to produce 1 km daily air temperature over the Yangtze basin, with RMSE 1.96 °C.
U-Nets forecast convection coverage from Himawari-8 brightness temperatures up to 120 minutes, beating persistence at lead times of 60 minutes and beyond.
Machine learning on polarimetric S-band radar storm cells predicts lightning flash rates up to 30 minutes ahead, with polarimetric gains mainly for short datasets.
NowcastingNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Releases a 2016-2019 statewide Iowa rainfall event dataset built from NEXRAD radar with quantitative precipitation estimation, intended for ML and physical modeling.
DR-Transformer estimates tropical cyclone intensity from satellite images with rotation and distance-consistency features, reaching 7.76-knot RMSE on the TCIR dataset.
Integrates multi-sensor Earth observation data and machine learning to monitor meteorological drought in central India; the abstract is missing.
Random forest downscales GPM satellite daily precipitation to 0.01° and cokriging merges it with gauges, retaining accuracy while improving spatial detail.
Deep learning model (DLztd) predicts daily GNSS zenith tropospheric delay at four West African stations from surface meteorology, filling data gaps.
CNN-based fusion of satellite and ERA5 predictors detects 97.1% of developing tropical cyclones in the western North Pacific at a 20.3% false alarm rate.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)
Interpretable machine learning methods inspired by the Dvorak technique estimate tropical cyclone intensity from satellite imagery; the abstract is unavailable, so details are limited.
CNNs estimate western North Pacific tropical cyclone intensity from Himawari-8 infrared images, reaching 88.9% accuracy and 4.62 m/s RMSE with focal loss and binary framing.
CNN classifiers (VGG, ResNet, Xception) on satellite imagery detect tropical cyclone rapid intensification, with Xception reaching ROC-AUC 0.911 and FAR 0.182.
A fast-frugal decision tree ranks seven long-term precipitation datasets against 813 China gauge grids, placing APHRODITE first; machine learning is a light evaluation tool.
Tests 13 machine learning algorithms with pre- and post-integration of satellite precipitation products for drought indices over India; Bayesian-regularized neural networks perform best.
Video-based CNN with ConvLSTM locates tropical cyclone centers from hourly geostationary infrared sequences, giving a 19.3 km median error and 42% lower error than baseline.
Random forests combining GOES-16 satellite, lightning and radar fields improve severe storm warning prediction over radar alone, identifying cloud-top divergence as most informative.
NowcastingNASAColorado State University (incl. CIRA)University of Oklahoma (AI2ES)
Gradient boosting fuses eight reanalyses with MODIS-based lapse rates to produce a 1-km daily air temperature dataset over the Tibetan Plateau for 1980-2014.
A cascaded CNN classifies and estimates northwest Pacific tropical cyclone intensity from infrared satellite images, reaching 8.60 kt RMSE against best-track data.
Deep learning is applied to downscale TRMM satellite precipitation products over northeast China; the abstract is missing, so methods and results are unverified.
Physics-augmented deep learning estimates tropical cyclone intensity and size from satellite imagery; classified from title only, as no abstract was available.
A nonlinear Gaussian-process Granger causality method attributes global vegetation greenness to precipitation and soil moisture over 30 years of gridded satellite data.
Attention-based U-Net downscales ERA5-Land, TRMM and IMERG monthly precipitation to 1 km over Central Texas, using PRISM as target over 2001–2018.
Gradient boosting, SVR and ANN downscale TRMM precipitation from 25 km to 1 km using vegetation indices, with the ANN performing best in the Mekong region.
Compares two regression and three ML methods (ANN, SVM, RF) for downscaling IMERG to 1 km over Gansu; ML beats regression against 80 rain gauges.
Simulated oblique Earth-space link network reconstructs rainfall fields (RMSE under 3.46 mm/h) and a deep learning model nowcasts rain cell motion and peak intensity.
CNN on polar-coordinate satellite images profiles full tropical cyclone structure parameters and wind profiles, with a new labeled benchmark dataset as baseline.
GAN-CNN framework trains on visible and passive microwave data and estimates tropical cyclone intensity from IR and water vapor images every 15 minutes, versus 3-hourly before.
A multi-task two-stream CNN fuses radar and satellite data for convective storm nowcasting, using pixel-wise sampling to handle class imbalance.
Multi-GRU-RCN model forecasts intra-hour cloud motion from FengYun-2G infrared imagery, outperforming optical flow, ConvLSTM and GRU baselines.
A deep CNN super-resolves NEXRAD radar reflectivity, outperforming interpolation at 4x and 8x resolution increases on objective and perceptual metrics after six months of training data.
Support vector regression on ATMS satellite warm-core temperature anomalies estimates Atlantic hurricane intensity, with Pmin error of 7.23 hPa when day of year is included.
TCLNet, a lightweight fully convolutional network with a new typhoon center dataset, locates typhoon centers with 14.4% higher accuracy and 92.7% fewer parameters.
A tensor-algebra CNN with multitask learning estimates tropical cyclone wind speed and intensity class from FY-4 multispectral imagery, handling invalid bands.
A quasi-supervised Mask R-CNN reidentifies Northern Hemisphere extratropical cyclones, finding 8.29% more cyclones than the traditional automatic method.
Climate analysis & attributionNanjing University of Information Science & Technology (NUIST)
A GCN-LSTM model classifies typhoon levels from Himawari-8 cloud images with 92.35% accuracy, and 95.12% for typhoons and super typhoons.
Tutorial-style guidance on evaluating, tuning and interpreting neural networks for meteorological satellite images, covering receptive fields, relevance propagation and image-to-image translation.
Explainability, uncertainty & physicsColorado State University (incl. CIRA)
Machine learning using satellite inner-core precipitation and outflow temperature beats the NHC operational rapid-intensification consensus, improving Peirce skill score by 12 to 138%.
Extremes & hazardsIBM ResearchNASACalifornia Institute of Technology (Caltech)
Interpretable satellite-derived convective structure features with logistic lasso classify tropical cyclone rapid intensity change, matching the accuracy of environmental predictors.
Combines GNSS integrated water vapour with tomography wet refractivity profiles in an ML model predicting 0–2 h storm occurrence over Poland, above 87% accuracy.
A cloudiness-classified CNN framework nowcasts PV output from sky images, reaching 2.20 kW RMSE and improving on an end-to-end CNN by 6%.
Unsupervised clustering of 850-hPa circulation over the NE Pacific links synoptic regimes to marine boundary-layer cloud droplet properties and cloud radiative forcing.
SVM and BPNN downscale IMERG daily precipitation over Zhejiang; SVM raises R² from 0.344 to 0.615 against gauges.
Four U-Net models detect tropical and extratropical cyclone regions from GFS precipitable water and GOES imagery, with Dice scores of 0.51 to 0.76 and 3x speedup over a heuristic.
An artificial neural network estimates IASI cloud-top pressure and fraction closer to optimal values, cutting retrieval time and improving 1D-Var cloudy-sky assimilation with near-neutral forecast impact.
Data assimilationMet OfficeKorea Meteorological Administration (KMA)
Two downscaling methods (GWR and BP-ANN, plus bias correction and PCA) refine TRMM rainfall from 25 km to 1 km over Hubei, validated against 27 gauges.
PCA/ICA with BP network and LSTM model GPS zenith tropospheric delay in West Antarctica; six-hour forecasts reach 7.2 mm RMS.
LightningNet, a 3D convolutional encoder-decoder fusing Himawari-8, radar and lightning data, gives 0-1 h cloud-to-ground lightning nowcasts over China with threat score 0.453.
Artificial neural networks classify 500 hPa circulation patterns, showing the 2019 Greenland melt season was about five standard deviations anomalous in atmospheric persistence.
Climate analysis & attributionNASAColumbia University (LEAP, Lamont-Doherty)
PreciPatch uses dictionary learning to downscale hourly GCM precipitation, outperforming bicubic, RainFARM and DeepSD on IMERG data.
Compares ray tracing on numerical weather data with a machine learning model for predicting atmospheric refraction-induced image displacement, consistent with field measurements in overall amplitude and phase.
Conditional GAN generates realistic cloud reflectance fields from sea surface temperature and humidity, a step toward stochastic cloud parameterization.
Uses machine learning on Darwin C-band radar data to derive transition functions for a stochastic convective-stratiform cloud model aimed at global climate model parameterization.
Hybrid physics-ML & parameterizationUniversity of ReadingBureau of Meteorology (Australia)
ConvLSTM with optical-flow features nowcasts convective radar echoes, better capturing rotation, merging and separation than traditional methods.
Machine learning on satellite imagery estimates high-resolution temperature during heat waves; abstract unavailable, so details are unverified.
Multi-task deep learning fuses COMS geostationary images with CFSv2 forecasts, cutting 6h and 12h tropical cyclone intensity forecast RMSE by 13% and 16%.
CNN estimates FY-4A satellite tropical cyclone intensity off southeast China with over 97% class accuracy and 1.75 m/s mean absolute error.
CNN classifies storm versus non-storm satellite images for cyclone detection after augmentation, reporting 90% accuracy, with linear regression for track prediction.
Hybrid CNN with classification and BP network estimates tropical cyclone intensity from satellite images with RMSE of 8.91 kt.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
2D and 3D CNNs estimate tropical cyclone intensity from multispectral geostationary imagery, reaching 8.32 kt RMSE, about 35% better than single-channel CNNs.
Statistical models predict tropical rainfall type and amount over the next 6 hours from MERRA-2 atmospheric profiles, trained on TRMM data, capturing time-mean spatial patterns.
Deep convolutional networks detect two anthropogenic aerosol-driven perturbations in satellite cloud imagery to assess their prevalence and climatic effects.
Cumulo is a global dataset pairing one year of 1km MODIS imagery with CloudSat cloud labels, with a flow-based baseline for cloud classification.
Benchmarks, datasets & evaluationUniversity of OxfordThe Alan Turing Institute
Cubist regression downscales PERSIANN-CDR to 1 km and calibrates it, yielding 1983–2015 Tibetan Plateau precipitation with bias reduced from 41% to about 5%.
Downscaling & super-resolutionChina Meteorological Administration (CMA)
A two-branch CNN estimates northwest Pacific tropical cyclone intensity from infrared and water vapour images, reaching RMSE 5.13 m/s with 60×60 pixel inputs.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)
Uses a CNN-LSTM on COMS satellite channels to estimate typhoon maximum wind speed, reaching 9.84 knot RMSE with the WV channel.
Neural networks optimized by PSO, ICA and GA downscale satellite precipitation over Austria, with ICA best and RMSE of 5.26 mm against 54 rain gauges.
Deep-PHURIE uses deep learning to estimate hurricane intensity from infrared satellite imagery; the abstract is unavailable, so accuracy is unknown.
Classifies tropical cyclone intensity from infrared satellite images using GLCM texture features and an SVM, reaching 88% accuracy against the Saffir-Simpson scale.
CNN with a novel probability-to-object step detects synoptic warm and cold fronts in reanalysis grids, achieving 0.73 probability of detection at 250 km.
Climate analysis & attributionNSF NCAR / UCARUniversity of Oklahoma (AI2ES)
Neural-network soil moisture retrievals from SMOS are assimilated into ECMWF's land model, improving Southern Hemisphere and some Northern Hemisphere forecasts through atmospheric feedback.
DeepMicroNet, a CNN on passive microwave satellite imagery, estimates tropical cyclone intensity with 9.6-14.3 kt RMSE, matching known best-track biases.
Conditional GAN generates realistic MODIS marine low cloud reflectance fields from large-scale variables, proposed as a stochastic cloud parameterization for sub-grid cloud distributions.
Random forest, decision tree and SVM detect western North Pacific tropical cyclone formation from WindSat with 94–96% hit rates, 26–30 h early.
Compares seven statistical post-processing methods for probabilistic HARMONIE-AROME solar radiation forecasts; quantile regression and generalized random forests perform best.
Deep learning applied to phased-array weather radar volumes detects tornadic vortices in Typhoon Nanmadol (2017); the abstract gives no quantitative results.
A CNN trained on satellite infrared and microwave data estimates tropical cyclone intensity with 8.39 kt RMS error, offering an objective alternative to the Dvorak technique.
Random forest with precipitation zoning downscales satellite rainfall to 1 km over the Lancang–Mekong basin, cutting RMSE by 1–17%.
Random forest on Himawari-8 brightness temperatures and NWP water vapour classifies weak, medium and severe pre-convective storms over East Asia with 0.79 accuracy.
Compares regression, ANN and spline downscaling of IMERG satellite rainfall in Austria; all improve accuracy, with spline best at correlation 0.56.
Dissertation proposing an LSTM model to forecast infrared cloud-top temperatures, combined with a satellite-to-rainfall mapping for 0-6 hour precipitation forecasts.
Thesis develops sky-camera and satellite-based short-term solar radiation nowcasting, including machine learning blending models.
Computer-vision pipeline with soft-mask inpainting fuses satellite and radar observations to extend precipitation nowcasting beyond radar coverage.
Deep learning model estimates tropical cyclone intensity from satellite data at higher temporal frequency, deployed in a production monitoring system with a portal.
Uses optical-flow error from satellite imagery as features for tree and neural classifiers to predict lightning up to five hours ahead, with 96% accuracy at 15 minutes.
PHURIE uses machine learning on infrared satellite imagery to estimate hurricane intensity; abstract unavailable, classification based on title only.
An LSTM extrapolates cloud-top brightness temperature, then PERSIANN retrieves 0–6 h rainfall, beating RNN, optical flow and RAP over three US regions.
Deep CNNs estimate tropical cyclone intensity directly from infrared satellite imagery as an objective alternative to subjective Dvorak analysis.
Marginalized Gaussian CRF predicts continental US precipitation from partially observed climate-variable graphs and identifies weather stations to cut data-collection costs.
Neural networks trained on CINRAD radar reflectivity and gauges nowcast Wuhan rain intensity, reaching over 50% accuracy at 36-minute lead for 5 dBZ.
