Forecasts from minutes to about 6 hours, mostly precipitation and convection, learned from radar, satellite or station data.
Search and filter this topic Plain list on GitHub
NowcastNet unifies advection physics with learned conditional generation for extreme precipitation nowcasts up to 3 hours, ranked first by meteorologists in 71% of cases.
China Meteorological Administration (CMA)Tsinghua University (NowcastNet)University of California, Berkeley
MetNet-3 learns from sparse and dense sensors to forecast precipitation, wind, temperature and dew point up to 24 hours ahead, beating HRRR and ENS over CONUS and running in Google Search.
DGMR, a deep generative model of radar, produces skillful 5-90 minute precipitation nowcasts that Met Office experts ranked first in 88% of cases.
MetNet uses axial self-attention on radar and satellite inputs to forecast precipitation up to 8 hours at 1 km, beating NWP for 7-8 hour lead times over the US.
A U-Net treats 1-hour, 1 km precipitation nowcasting as image-to-image translation and outperforms optical flow, persistence and HRRR.
DiffCast decomposes precipitation into deterministic motion and stochastic residuals via residual diffusion, reducing blurring and position errors on four radar datasets.
Compares RainNet, ConvLSTM U-Net, recursive U-Net and GAN nowcasting on South Korean S-band radar, with RainNet reaching CSI 0.826 at 0.1 mm/h.
Transformer nowcasts radar composites up to two hours ahead from geostationary satellite data, with lightning and infrared channels proving key features in severe weather.
Combines multi-temporal optical flow with a conditional GAN for radar rainfall nowcasting, improving accuracy over MAPLE for 0–3 h lead times.
Object-based probabilistic deep learning on GOES-16 infrared imagery predicts convection initiation up to 1 hour ahead, beating logistic regression, with explanations of moisture and cloud drivers.
Tests 24 schemes for deep-learning radar nowcasting in South Korea, finding U-Net beats ConvLSTM and balanced MSE loss helps heavy rain.
A physics-aware ConvLSTM followed by a U-Net predicts km-scale rainfall occurrence from satellite radiances up to 4 hours ahead for the weather4cast challenge.
Reviews precipitation forecasting, covering rainfall physics and the growing role of machine learning methods, with discussion of computational challenges.
Preformer, a streamlined transformer with encoder-translator-decoder design, achieves state-of-the-art precipitation nowcasting with the fewest parameters among compared models.
Light3DUnet, a 3D U-Net on radar reflectivity, nowcasts cloud-to-ground and intra-cloud lightning within one hour over central China.
STIN uses spatiotemporally-aware convolutions to fuse temperature, humidity and wind with precipitation for nowcasting, improving on shared-kernel baselines.
U-Net nowcasts convection initiation and thunderstorms over the continental US within 60 min from GOES-16 imagery, with MRMS radar targets.
A multi-layer attention module improves ConvRNN radar echo extrapolation, preserving high-intensity echoes and long-term skill on HKO-7 and HMB radar datasets.
ML-assisted physical models extract cloud features from sky-imager and irradiance data for intra-hour solar nowcasting, reducing hardware supervision needs for solar energy.
3D CNN forecasts precipitation 4 h ahead from FY-4A multichannel satellite imagery, beating NWP on critical success index (0.25 vs 0.21).
A deep generative nowcasting model with a balanced loss and temperature input extends skillful heavy-rain nowcasts over the Netherlands to 60-minute lead times.
A temporal graph convolutional network with learned adjacency predicts radar-based precipitation over the Trento area, reducing smoothing and improving error measures over CNN baselines.
U-Net variants nowcast radar rainfall one hour ahead and feed HEC-HMS flood models; TransAtt-Unet keeps simulated peak errors under 20%.
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.
Proposes an asymmetric receptive field autoencoder for spatiotemporal prediction and releases RainBench, a radar echo dataset for precipitation nowcasting, with state-of-the-art results.
LSTM uses raindrop size distribution with radar and gauge data to predict Guizhou precipitation 0–180 min, with 60-min correlations near 0.93.
LightGBM with aerosol features and GLM satellite data nowcasts hourly CONUS lightning at 0.25°, reaching 94.3% accuracy and 75% probability of detection.
Generative ensemble diffusion produces precipitation scenarios over Central Europe from ERA5 data, outperforming U-Net baselines after a learned post-processing step.
TAFFNet adaptively fuses radar echoes with NWP using time-aware weighting, improving 0–12 h precipitation forecasts and storm initiation and dissipation.
A U-Net ConvLSTM nowcasts near-global IMERG precipitation every 30 minutes up to 4 hours; a classification-loss variant better predicts rates above 8 mm/hr.
NowcastNet unifies advection physics with learned conditional generation for extreme precipitation nowcasts up to 3 hours, ranked first by meteorologists in 71% of cases.
China Meteorological Administration (CMA)Tsinghua University (NowcastNet)University of California, Berkeley
Multi-scale deep learning framework with attention and a spatiotemporal loss improves 1-hour storm nowcast skill scores by 8.5–42.6% on SEVIR.
Random forest model for short-term rainfall nowcasting; abstract unavailable, so performance could not be assessed.
Physics-constrained deep learning (CGsNet) nowcasts convective wind gusts from radar at 0.01° and 6-minute resolution for 0-2 h, beating NWP-based nowcasting.
MACIN combines a CNN cloud detector with data assimilation to nowcast direct irradiance up to 20 minutes ahead, improving on persistence beyond one minute.
Attentional GAN with a spatial variability loss reduces the blurry effect in radar nowcasts and outperforms competing models on the test set and storm cases.
Tsinghua University (NowcastNet)Colorado State University (incl. CIRA)Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
LSTMAtU-Net combines ConvLSTM with U-Net and an attention module to improve 0-2 hour radar precipitation nowcasts, especially for medium and high intensities.
Nanjing University of Information Science & Technology (NUIST)
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.
Dynamic weight loss function added to ConvLSTM, ConvGRU, PredRNN and PredRNN++ reduces radar echo extrapolation MAE by up to 14.8% over 1 hour.
Nanjing University of Information Science & Technology (NUIST)
MetNet-3 learns from sparse and dense sensors to forecast precipitation, wind, temperature and dew point up to 24 hours ahead, beating HRRR and ENS over CONUS and running in Google Search.
A 3D CNN that fuses environmental grid fields with radar echoes raises CSI by 3.42% at 40 dBZ for convective extrapolation.
U-Net and ConvLSTM models forecast fog and low stratus over Morocco 5 h ahead from Meteosat SEVIRI, beating optical flow on SSIM.
An anticipative vision transformer forecasts solar irradiance 15 minutes ahead from sky image sequences, reaching 21.45% forecast skill over smart persistence.
CLGAN, a GAN with U-Net and LSTM generator, improves heavy-precipitation nowcasts up to 2 h over PredRNN-v2 and optical flow baselines.
Nanjing University of Information Science & Technology (NUIST)Forschungszentrum Jülich / Jülich Supercomputing Centre
FlashBench combines a LightNet-style deep model with observed field evolution for regional lightning nowcasts over West India, delivered via a cloud forecast system.
TrajGRU and ConvLSTM nowcast Himawari-8 cloud top height up to 2 h, with TrajGRU beating optical flow and persistence in most seasons.
Deep learning nowcasts 6-minute precipitation up to 2 h over Guizhou from station data; PredRNN beats optical flow at all tested thresholds.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
A latent diffusion model for radar precipitation nowcasting gives more diverse, well-calibrated ensembles than DGMR and PySTEPS, with more accurate predictions.
Adversarially trained 3D LSTM nowcasts sudden localized heavy rain from MP-PAWR radar with 10-min lead, outperforming an advection model and JMA's operational system at onset.
SimVP-based radar nowcasting over South Korea improves 120-min CSI at 1 mm/h to 0.46 and handles typhoons and narrow convection better than prior models.
Dual-polarization radar nowcasting of severe convective weather combines physical constraints with a deep neural network; abstract unavailable, so details are limited.
EfficientRainNet, a lightweight EfficientNetV2-based CNN, matches convolutional GRU rainfall nowcasts on Iowa radar data with far fewer parameters.
Compares U-Net and Transformer for precipitation prediction from cloud-layer information; U-Net reaches 96.89% accuracy versus 94.13% for the Transformer.
CNN-RNN models predict northeast Algeria precipitation from synoptic station data and NWP outputs, showing gains over NWP-based prediction.
Mutual-information reweighting and time-superimposing training improve radar precipitation nowcasts for convective scenarios across several state-of-the-art models.
Random forests using wind profiler divergence and vorticity predict rainfall onset 30 minutes ahead, outperforming inputs derived from ERA5 reanalysis.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
SAR-UNet adds residual connections to SmaAt-UNet for precipitation and cloud-cover nowcasting, outperforming baselines from 30 to 180 minutes ahead, with Grad-CAM explanations.
A 3D-UNet-LSTM extractor-forecaster for 0-1 h radar echo extrapolation on MeteoNet outperforms baselines in quantitative and case-study evaluation.
Res is a lightweight transformer for quantitative precipitation nowcasting on remote sensing data; only source code is released while the manuscript is under review.
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.
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.
Nanjing 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.
Nanjing University of Information Science & Technology (NUIST)
Tuned fusion deep learning model forecasts tropical cyclone passive microwave rainfall images from satellite data, with no quantitative results reported in the abstract.
Multi-scale sequence-to-sequence deep network predicts future satellite cloud images; the abstract is unavailable, so quantitative results are unknown.
Combines sky images and satellite observations in one ML framework for deterministic and probabilistic intra-hour solar irradiance nowcasts up to 60 minutes ahead.
WF-UNet fuses precipitation radar with wind fields to nowcast western European precipitation 1 to 3 hours ahead, giving 6 to 22% lower MSE than radar-only UNets.
Adds a hand-engineered advection-diffusion PhyCell to PhyDNet for radar nowcasting, but the physics prior does not improve prediction skill.
GCMT-ConvRNN adds motion-field and sub-regression auxiliary tasks with a gradient-correlation loss to improve radar-based precipitation nowcasting.
Multi-task learning models nowcast convective storm vertically integrated liquid and lightning using satellite and weather radar observations; abstract not provided.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Enhanced multi-scale ConvLSTM network predicts typhoon cloud images three hours ahead from satellite sequences with sharper detail than prior methods.
Conditional diffusion model for short-term rainfall nowcasting; the abstract is unavailable so the classification relies on the title.
EfficientNetV2-based memory-resilient network for rainfall nowcasting; the abstract is unavailable so the classification relies on the title.
Reviews neural network approaches to precipitation nowcasting and tests existing methods, providing a comparative evaluation for radar-based forecasting.
Lightweight physics-informed transformer nowcasts quantitative precipitation from VIL radar products with convection-diffusion constraints and beats state-of-the-art models.
TS-RainGAN splits heavy-rainfall nowcasting into coverage and intensity GAN modules, reducing blurriness and giving skillful district-level forecasts up to 2 hours ahead.
STUNNER uses a two-stream spatiotemporal fusion network with temporal differencing for radar echo extrapolation, achieving best results on two public datasets.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Pixel-CRN nowcasts convective storms 30 and 60 min ahead from 3-D radar and reanalysis, and works with only a small training dataset.
MM-RNN adds meteorological element guidance to RNN nowcasters, raising CSI at R≥10 by 23.4% over MS-MotionRNN on MeteoNet.
HST-AFP fuses radar echoes with NWP data via attention-based multi-encoder and fusion decoder, outperforming RNN and U-Net nowcasting baselines in South China.
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).
Lagrangian CNN models rainfall growth and decay from advected radar differences, improving heavy-rain nowcast POD over RainNet and LINDA on Finnish radar.
A DyConvGRU and U-Net forecast-refinement network with WGAN training improves high-reflectivity radar echo extrapolation over Shanghai, raising POD and CSI.
Ensembles of CNN and LSTM networks with automatic hyperparameter tuning improve multi-step rainfall nowcasting by 10-25% over univariate models in Tehran.
Two-stage deep learning predicts typhoon satellite images and converts them to rainfall maps, reaching correlation 0.81 at 3 hours ahead.
PreDiff is a latent diffusion precipitation nowcasting model with a physics-based knowledge alignment step that enforces domain constraints on SEVIR radar forecasts.
TempEE uses a parallel spatio-temporal transformer with one-step forecasting instead of autoregression to avoid error accumulation in radar echo extrapolation.
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.
MSSTNet uses multi-scale 3D convolutions to predict precipitation nowcasts from radar echoes, achieving state-of-the-art on MovingMNIST and improved real radar accuracy.
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.
Ranks dual-polarimetric radar variables for predicting heavy rainfall from mesoscale convective systems over Korea and uses them in ML early-warning models.
Probabilistic machine learning forecasts short-term rainfall from cumulative precipitation fields built from station observations; the abstract is missing, so this rests on the title.
Compares ConvLSTM and adversarial SAVP models for hourly 12-hour 2 m temperature forecasts over Europe; SAVP reaches MSE about 2.3 K² versus 3.6 K² for ConvLSTM.
Neural network and random forest classify sky-camera images to nowcast clear-sky index drops 1–5 minutes ahead for PV cloud-induced variations.
Evaluates deep-learning very-short-term rainfall forecasts over South Korea; abstract not provided, so specific skill numbers are unknown.
RainUNet with decoupled large-kernel 3D convolutions forecasts high-resolution radar rainfall movies from low-resolution satellite data in the Weather4cast 2022 challenge.
Hyperparameter, augmentation and test-time augmentation recipe makes satellite-to-rainfall models robust to spatial-temporal shifts, ranking first on the W4C22 Transfer dataset.
Simple CNN baseline (SIANet) using only satellite images reaches state-of-the-art on four of five W4C22 rainfall benchmark tasks without ensembling or fine-tuning.
Transformer ensembles predict eight hours of rainfall from one hour of satellite imagery, ranking third in the NeurIPS 2022 Weather4cast challenge.
An LSTM encoder-decoder nowcasts extreme rain and wind over Malpensa airport for one hour, reaching over 90% wind detection probability with station and GNSS inputs.
U-Net fusing satellite frame prediction and rain extraction won the NeurIPS 2022 Weather4Cast challenge, forecasting severe rain up to eight hours ahead.
CycleGAN-style model with forward and backward temporal cycles generates radar precipitation 10 minutes ahead and rolls out 2-hour nowcasts with a heavy-rain loss.
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.
Two-stage TrajGRU and attention-enhanced U-Net GAN refines radar echo extrapolations, outperforming ROVER and RNN baselines on detail and precipitation accuracy metrics.
ISA-PredRNN adds self-attention and a weighted loss to PredRNN-V2 for radar echo nowcasting, reaching CSI 0.70 at the 10 dBZ threshold.
Compares local, global and transfer-learning strategies for deep learning solar forecasting from sky images across three heterogeneous datasets; transfer needs 80% less data.
Deep learning fuses radar, lightning, satellite, NWP and terrain data to probabilistically nowcast lightning, hail and heavy rain at 1 km and 5 min up to 60 min.
A convolutional autoencoder segments clouds in sky images above 90% pixel accuracy, and particle tracking predicts cloud motion for short-term PV power.
Adds a stochastic noise generation step to a deep network for ensemble urban radar rainfall nowcasts, matching optical flow skill over a full year at 5-minute resolution.
ASOC attention fuses sparse ground-station observations with radar images, raising CSI for 1-6 h heavy and light rainfall nowcasts by 5.7% over image-only models.
Fuzzy deep neural network predicts rainfall by classifying cloud images and detecting rainy clouds; the abstract gives no quantitative results.
T-UNet combines UNet with TrajGRU for radar echo extrapolation, improving CSI and HSS by up to 10.57% and 7.80% over 60-minute nowcasts.
ConvGRU with a structure-preserving loss forecasts European cloudiness 3 hours ahead from SEVIRI, outperforming optical flow by 12.43% in R2 across time steps.
Occlusion-perturbed CNN forecasts 10-minute-ahead irradiance from sky images, reaching 5.62% nRMSE and a 2.77% nCRPS for probabilistic forecasts.
Multi-scale extrapolation with a conditional GAN predicts FY-4A meteorological satellite images up to 2 hours ahead, capturing weather-system evolution in detail.
UTrans-Net adds a transformer to a U-Net to predict short-term precipitation from prior observations, reaching 59%, 78% and 82% accuracy at 0.1, 3 and 100 mm thresholds.
Nanjing University of Information Science & Technology (NUIST)
3D-ConvLSTM extrapolates multi-altitude radar volumes for convective storm nowcasting, outperforming optical flow and 2D deep baselines at longer leads for 35 and 45 dBZ echoes.
Gated attention recurrent network with multi-order Markov memory predicts radar echo sequences for precipitation nowcasting, improving on first-order spatiotemporal models.
TSGAN uses spatio-temporal GAN with attention for radar nowcasting of strong convection, giving sharper forecasts than optical flow and deep baselines; it is reported as used in an operational system.
NeXtNow, a ResNeXt-based encoder-decoder, predicts future radar frames from Romanian and Norwegian data, improving CSI by 1–17% over XNow.
Video Swin-Transformer with cross-attention forecasts eight hours of hourly satellite weather frames on Weather4Cast 2021, reaching MSE 0.442 in transfer.
Rad-cGAN, a conditional GAN for radar precipitation nowcasting, outperforms ConvLSTM and U-Net, extending useful lead time to 80 minutes and transferring across Korean dam regions.
Systematic PRISMA review of 110 studies on statistical, NWP and ML short-term rainfall forecasting, highlighting hybrid models for nowcasting.
Adaptive deep learning ensemble for precipitation nowcasting; abstract unavailable, so details are inferred from the title and journal.
Self-attention spatiotemporal LSTM extrapolates edge-radar echoes for nowcasting in an Internet of Radars setting; the abstract is missing, so the assessment relies on the title.
Nanjing University of Information Science & Technology (NUIST)
Back-propagation neural network combines GNSS tropospheric delay with meteorological data for short-term rainfall warning in Ningbo, with 100% true detection rate.
CEMA-LSTM adds contextual correlation and attention blocks to recurrent radar extrapolation, improving POD, FAR and CSI over PhyDNet and PredRNN++.
Nanjing University of Information Science & Technology (NUIST)
ConvLSTM precipitation nowcasting on German radar data shows that finetuning across locations and thresholds extends skill to 100 minutes.
M5Images, a convolutional LSTM on radar images, nowcasts strong convective rainfall near Campinas, Brazil, outperforming persistence at all lead times.
Combines all-sky camera images with satellite observations in one ML framework, improving intra-hour solar irradiance forecasts up to 60 minutes ahead.
A Conv-LSTM encoder-decoder with rain-weighted MSE extrapolates radar echoes for 2 h, reaching 0.42 CSI at 30 dBZ for 1 h, above CTREC.
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.
A graph-guided ConvLSTM nowcasts precipitation from tropical cyclones, combining spatio-temporal graph structure with convolutional recurrent modelling.
A parallel multi-frame radar nowcasting model with noise dropout and semantic restore modules reduces error accumulation and achieves competitive scores against state-of-the-art methods.
Adds FlowNet optical-flow loss to a GAN-based McNet for radar echo extrapolation, improving precipitation nowcasting on Shenzhen radar data.
Random Forest nowcasts cloud-to-ground lightning over Italy and seas, finding that finest-resolution gridding of inputs gives the best skill.
Training a deep learning radar nowcaster on all of Japan and then transfer learning to regions improves heavy-precipitation skill over locally trained models.
Conditional GAN turns webcam images into photorealistic future weather views conditioned on COSMO-1 forecasts, matching forecast conditions in at least 89% of cases.
Compares U-Net, ConvLSTM and SVG-LP for precipitation nowcasting on precipitation maps, adding a patch extraction scheme and a loss that reduces blurring.
Spatiotemporal LSTM with self-attention forecasts hourly rainfall from radar echo images; no abstract is available, so results are unverified.
A recurrent-convolutional network nowcasts lightning occurrence over Switzerland for 60 minutes, reaching CSI 0.75 at 5 minutes and 0.32 at 60 minutes.
CNN on infrared all-sky images forecasts clear-sky index ranges 5 to 15 min ahead, reaching about 65% accuracy at 5 min for solar irradiance.
Tests seven optimizers and eight learning rates for a CNN forecasting hourly typhoon rainfall 1–6 h ahead; moderate rates (0.1–0.001) work best.
Nanjing University of Information Science & Technology (NUIST)
Combines a UNet with PredRNN_v2 to predict radar reflectivity fields in Vietnam, cutting training time to 30% while reaching SSIM 0.95 and CSI 0.94.
Gradient-boosted nowcasts of reflectivity, lightning and hail-related echo tops show radar is the most useful input and satellite helps across all hazards.
A GAN with convolutional recurrent units extrapolates Shenzhen radar echoes, raising POD and CSI by 5.5% and 10.4% over rgcPredNet.
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.
AA-TransUNet adds attention and depthwise convolutions to TransUNet for precipitation nowcasting, outperforming baselines on Dutch precipitation and French cloud cover data.
CLSTM-LFN video-prediction model fuses lightning history with NWP variables to forecast 0-3 h cloud-to-ground lightning areas, outperforming parameterization schemes.
MCCS-LSTM adds criss-cross attention and multi-scale spatiotemporal blocks to radar echo extrapolation, improving accuracy at longer lead times on CIKM and Guangdong radar data.
Nanjing University of Information Science & Technology (NUIST)
A ConvLSTM treats GOES satellite-based solar irradiance nowcasting as next-frame prediction, reaching 120 W/m2 RMSE for North America up to 3 hours ahead without a GPU.
Proposes M-ConvGRU, a convolutional GRU with an encoder-forecaster design for radar echo prediction, cutting MAE 6.29% and MSE 10.25% versus ConvLSTM.
Proposes a two-stream convolutional LSTM for radar-based precipitation nowcasting, combining two input streams to improve echo map prediction.
Nanjing University of Information Science & Technology (NUIST)
Random Forest and XGBoost ensembles trained on automatic weather station data predict 2-4 hour rainfall occurrence with about 80% accuracy.
Applies machine learning to nowcast extreme weather using different input datasets; no abstract available, so details are inferred from the title.
Convolutional LSTM nowcasts of meso-gamma convective storms from MP-PAWR radar over Tokyo; the ConvLSTM beats advection for total rain at 10 min lead time.
Earthformer is a space-time transformer with cuboid attention that reaches state-of-the-art precipitation nowcasting and ENSO forecasting on real benchmarks.
CNN and RNN nowcasters applied to IMERG satellite precipitation extend forecasts to 4.5 h, mitigating the latency of satellite rainfall products over CONUS.
A motion-guided global-local Transformer with adversarial regularization improves radar echo extrapolation, giving sharper heavy-rain predictions and better skill scores than CNN and ConvRNN baselines.
A variance-weighted loss improves neural lidar wind-field nowcasting during highly variable events, cutting errors by over 15.2% on wind shear onset and decay.
A self-attention GRU nowcasts radar echoes for landfalling hurricanes and extreme convective rainfall in the southeast US, aiming to improve skill beyond the first hour.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
NLED uses spatiotemporal self-attention in an encoder-decoder to extrapolate radar echoes over Shanghai and Hong Kong, improving long-term nowcast accuracy.
Focal Frame Loss weights prediction frames by difficulty, improving several radar nowcasting models without extra training cost on two large radar datasets.
DMSF-GAN extrapolates radar images for 0-6 h typhoon nowcasting and outperforms optical-flow and RNN baselines; the abstract gives no quantitative results.
A ConvLSTM encoder-forecaster nowcasts 100 m fine-scale wind fields from Hong Kong airport lidar up to 27 minutes, beating nine baselines including optical flow.
Systematically compares GAN and Wasserstein GAN adversarial regularization for radar echo nowcasting; multi-scale Wasserstein GAN performs best across cases.
Introduces InstGCN, an inductive spatiotemporal graph network for radar-based short-term rainfall forecasting, with a novel elliptic graph construction and subgraph decomposition for unseen nodes.
Systematic review of 65 articles (2011-2020) on sky-camera cloud nowcasting for solar power, finding a shift from classical statistics toward neural networks and CNNs.
PredRANN adds spatiotemporal attention to a convolutional recurrent network for radar-based precipitation nowcasting.
Self-supervised CNN-LSTM models forecast GOES-R satellite irradiance to nowcast solar PV output, reaching errors close to ground-truth-based models at 25 sites.
Spatial and spectrum GANs with masked style loss sharpen ConvLSTM and U-Net radar echo extrapolations on SEVIR, plus a new power spectral density score.
SmaAt-UNet, a lightweight U-Net autoencoder, forecasts eight hours of satellite-based weather fields from one hour of input in the Weather4cast 2021 challenge.
PTCT combines patch-based Transformer with 3D temporal convolution for autoregressive radar echo nowcasting, reporting state-of-the-art results on Guangzhou and HKO-7 datasets.
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.
Microsoft's operational ConvLSTM radar nowcasting model, conditioned on HRRR, beats optical flow and HRRR by 20–25% for lead times up to 6 hours.
A large-context neural network forecasts precipitation 12 hours ahead over CONUS, beating HRRR and HREF in skill.
Improves a recurrent-convolutional satellite nowcasting model for the Weather4cast 2021 competition via a shallower variant, AdaBelief and ensembling.
Variational U-Net combining VAE probabilistic modelling with U-Net detail recovery placed third in the Weather4cast satellite forecasting challenge.
Encoder-forecaster with residual ConvGRU and U-Net-style shortcuts for Weather4cast satellite image prediction; early forecasts stay sharp, later ones blur.
Two-stage spatiotemporal RNN with detail refinement improves radar echo nowcasting over PredRNN++ and ConvLSTM, reaching CSI 0.195 at 45 dBZ with 2 h lead.
Adding polarimetric radar variables KDP and ZDR to a deep-learning nowcasting model raises CSI by 13.2% at 30 min and 17.4% at 60 min.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
IrradianceNet is a satellite-based deep network forecasting surface solar irradiance up to 4 h over Europe, beating persistence, TV-L1 and ERA5 baselines via fewer cloud errors.
RAP-Net embeds a region attention block and recall attention mechanism into ConvRNN to improve radar-based precipitation nowcasting in heavy-rain regions.
Machine learning nowcasts solar irradiation by classifying cloud types from satellite imagery; the abstract is missing, so details are limited.
U-Nets forecast convection coverage from Himawari-8 brightness temperatures up to 120 minutes, beating persistence at lead times of 60 minutes and beyond.
Systematic study of deep learning storm nowcasting design choices (history length, loss, predicted frames) from radar data, yielding models that outperform existing implementations.
Deep ensemble learning on radar reflectivity videos issues severe thunderstorm warnings for Liguria, Italy, with a value-weighted skill score for probabilistic-to-binary alerts.
ATMConvGRU adds axial attention memory to ConvGRU for sequential weather prediction of temperature, humidity, wind and radar echo, outperforming state-of-the-art methods.
Deep learning precipitation forecasts from spatio-temporal inputs in a regional case study; no abstract available, so details are limited.
U-Net on five radar frames predicts convective reflectivity, matching TrajGRU and outperforming TREC with a simpler CNN.
Trains ConvGRU radar echo extrapolation with SSIM and MS-SSIM losses to reduce blurring; MS-SSIM gives the most realistic strong-echo forecasts out to 2 hours.
Nanjing University of Information Science & Technology (NUIST)Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
A 3D convolutional neural network applied to multi-parameter phased-array weather radar volumes improves short-term precipitation nowcasting.
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.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Residual convolutional encoder-decoder models for radar echo nowcasting match or beat operational systems and transfer across regions without performance decay.
MSDM combines optical flow, random forest and CNN to nowcast East China precipitation at 30 to 120 minutes from radar and satellite data.
Modular fully convolutional networks use radar echoes to forecast hourly typhoon rainfall 1 to 6 h ahead in southern Taiwan, outperforming an MLP baseline.
Radar nowcasting combines TVL1 optical flow with SINDy sparse dynamics to produce ensemble precipitation forecasts that beat benchmarks on typhoon and squall-line cases.
FDNet separates radar echo motion into optical flow and morphological deformation via two parallel encoders, achieving state-of-the-art precipitation nowcasting on two radar datasets.
Review of physical and data-driven sky-image methods for intra-hour solar irradiance nowcasting, identifying which image features matter most.
Neural network combines RadVOR radar nowcasts with postprocessed COSMO-DE-EPS ensembles, giving calibrated precipitation threshold probabilities that beat both inputs for 1 to 6 h.
DGMR, a deep generative model of radar, produces skillful 5-90 minute precipitation nowcasts that Met Office experts ranked first in 88% of cases.
Conv3D-GRU encodes radar echo sequences to predict rainfall intensity 1–2 hours ahead, improving short-term heavy rainfall accuracy over baselines.
CNN-LSTM model predicts weather radar echoes for agricultural applications; nowcasting from radar sequences is the core contribution.
Deep learning forecasts intra-day solar irradiance from satellite-derived estimates near a PV plant, a short-horizon nowcasting application.
SCENT pairs sparse correspondence tracking with a deep network for precipitation nowcasting; abstract unavailable, so the classification rests on the title.
Nanjing University of Information Science & Technology (NUIST)
LSTM post-processes MAPLE radar nowcasts for 180-minute rainfall at 13 urban stations in South Korea, outperforming MLR, MARS, MLP and basic RNN baselines.
Dual-encoder RNN fuses radar mosaics and station rainfall for 0–2 hour grid nowcasts, reaching 0.523 threat score at 0.5 h, about 3–4 times WRF.
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.
NASAColorado State University (incl. CIRA)University of Oklahoma (AI2ES)
Machine learning models trained on temporal properties of radar convective storms predict their future locations, showing usable nowcasts with slight error.
GRU models using radar reflectivity and station data forecast typhoon rainfall 1, 3 and 6 hours ahead in southeastern Taiwan, outperforming XGBoost and linear regression benchmarks.
Weighted broadcasting in an encoding-forecasting radar model emphasizes the latest frames, improving MAE by 1.74% and heavy-rain (30 mm/h+) skill by over 30%.
Transfers a CNN radar nowcaster trained near Beijing to Guangzhou via fine-tuning and MMD, improving skill with only small target-domain datasets.
Nowcasting model adds a discriminator for realistic rain maps and consecutive attention extending lead time to three hours, with a new radar benchmark.
Broad-UNet with asymmetric convolutions and ASPP predicts satellite-based precipitation and cloud cover maps more accurately than standard UNet with fewer parameters.
IDA-LSTM adds coupled convolutions and channel-position dual attention to ConvLSTM to reduce underestimation of high radar echoes in extrapolation on CIKM 2017 data.
A modified fuzzy inference model gives probabilistic typhoon rainfall forecasts 1–3 hours ahead, with median-based deterministic forecasts performing satisfactorily.
Self-attention ConvLSTM encoder-forecaster for radar-based precipitation nowcasting, evaluated on HKO-7 radar echo sequences with improved visual and quantitative results.
A 3-D CNN plus conditional GAN postprocessor sharpens radar echo extrapolations, improving intense echo and convective prediction over North China in 2015–2016.
NowDeepN ensemble of deep neural networks predicts future radar product values from past values, reaching under 4% normalised RMSE on Romanian radar data.
Fusing rain radar images with NWP wind forecasts in a deep network improves 30-minute rain nowcasts, scoring 8% higher F1 than optical flow and 7% higher than radar-only.
PFST-LSTM adds a spatial memory cell and position alignment to ConvLSTM for precipitation nowcasting, outperforming baselines on CIKM 2017 radar data.
Machine learning precipitation nowcasting from Hubei radar reflectivity handles missing radar data and outperforms optical flow, with a mechanism for single and cumulative altitudes.
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.
Applies the PredRNN++ video prediction model to São Paulo radar reflectivity sequences, forecasting up to 1 h ahead as a nowcasting aid.
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.
Random forest post-processing of radar nowcasts and NWP for hourly precipitation occurrence at 41 Norwegian stations, improving skill and reducing false alarms.
Fully convolutional learnable optical-flow network extrapolates six years of Guangdong radar mosaics, outperforming optical flow and other deep models for convective nowcasting.
Estimates short-term solar irradiance from sky images with low computational cost, forecasting up to 4 h ahead on 16 years of data.
Multimodal semisupervised graph convolutional framework fuses meteorological and non-meteorological data to nowcast precipitation, including areas lacking radar coverage.
CNNs on Meteosat imagery for cloud cover nowcasting; U-Net beats persistence and the AROME physical model on meteorological metrics.
An energy-based GAN combined with a convolutional RNN extrapolates radar echo maps with sharper, more stable results than ConvGRU and GA-ConvGRU baselines.
Graph-based multicell features with a random forest predict next-hour short-term intense precipitation in North China with 79.2% POD and 68.3% CSI.
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.
SmaAt-UNet, a lightweight attention U-Net with depthwise convolutions, matches baselines for precipitation nowcasting using a quarter of the parameters.
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%.
A physically structured deep model learns a low-dimensional space with minimal dynamics, outperforming model-free and hybrid nowcasting on precipitation and sea-surface temperature data.
POD combined with Kriging or Random Forest emulates Arome-NWC rainfall nowcasts over France at 1-minute resolution with much lower computational cost.
A generative neural network nowcasts radar precipitation up to one hour ahead and beats optical-flow methods on five years of Japanese radar data.
RainNet, a U-Net/SegNet-style CNN, nowcasts German radar precipitation to 60 min, beating optical flow on MAE and CSI but underpredicting heavy rain.
Optical flow on sky images tracks cloud motion, and a neural network links it to short-term whole-sky solar radiation changes for plant efficiency estimates.
Combines kinematic advection analysis with Lagrangian Koopman operator analysis to improve short-term precipitation forecasts, with stable long-period predictions.
Reviews radar-based precipitation nowcasting methods from environmental science and the machine learning techniques applicable to them, arguing for closer collaboration between the two communities.
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.
MetNet uses axial self-attention on radar and satellite inputs to forecast precipitation up to 8 hours at 1 km, beating NWP for 7-8 hour lead times over the US.
Investigates convolutional neural networks for rainfall nowcasting from multi-parameter phased array weather radar observations (abstract not available).
Convcast, a convolutional LSTM trained on NASA IMERG satellite precipitation, nowcasts up to 150 minutes with 0.805 mm/h RMSE at 30 minutes, outperforming optical flow.
Stacking TrajGRU ensemble members with orographic features via a CNN roughly doubles skill on extreme rain rates, matching Lagrangian persistence on TAASRAD19 radar data.
Dual-channel CNN with multihead attention and residual connections improves radar-based local precipitation nowcasting, reaching state-of-the-art accuracy on a radar echo dataset.
ConvLSTM with optical-flow features nowcasts convective radar echoes, better capturing rotation, merging and separation than traditional methods.
ConvLSTM with star-shape bridge and rain-oriented loss improves radar precipitation nowcasting over convolutional LSTM baselines and extrapolation.
Conference abstract applying deep learning to convective storm nowcasting; no abstract available, so the specific method and gain are unclear.
Pure deep learning approach to precipitation nowcasting from radar; abstract unavailable, so classification rests on the title alone.
Combines Koopman operator analysis with kinematic analysis for short-term precipitation forecasting; abstract is in Japanese and the ML contribution is modest.
Conditional GAN with a predictive learning module sharpens meteorological image forecasts, achieving 14% better sharpness than the baseline.
MCSIP Net uses adversarial spatio-temporal discriminators and meteorological analysis critics to predict multichannel satellite images for weather nowcasting.
A multi-scale deconstructed ConvLSTM encoder-decoder forecasts precipitation from multiple meteorological fields on EC and CM datasets, combining NWP data with deep learning.
A U-Net treats 1-hour, 1 km precipitation nowcasting as image-to-image translation and outperforms optical flow, persistence and HRRR.
A UNet++/ResNet-inspired CNN on satellite images predicts lightning and thunderstorms 15 minutes ahead with over 94% probability of detection.
An MLP neural network nowcasts hourly summer precipitation over the Eastern Alps, outperforming the INCA nowcasting system and a multiple linear regression baseline.
A CNN with cross-channel 3D convolution nowcasts convective storms from 3D Doppler radar in Beijing-Tianjin-Hebei, outperforming traditional ML baselines on 3.7 million samples.
A location-generalizing deep network forecasts short-term solar irradiance from satellite and NWP inputs without local ground data, matching local models (31.31% vs 32.01% rRMSE) across 25 Dutch sites.
Machine learning nowcasts lightning up to 30 minutes ahead from four surface variables at single stations, beating persistence, electric-field and CAPE baselines.
Nowcasts convective initiation 17-40 minutes ahead from FY-4A satellite data using optical flow and BP-AdaBoost, reaching detection probability 0.80 in South China.
CNN-based deep integro-difference model with an ensemble Kalman filter gives calibrated SST forecasts and 10-minute Sydney radar nowcasts, reusing the ocean-trained model.
Machine learning is applied to short-term rain forecasting; the abstract is missing, so the scope and results are unclear.
A GAN with conditional generator and two discriminators extrapolates radar echoes 0.5 to 1.5 h ahead, reducing blur and outperforming baseline models.
A convolutional recurrent network for 3D radar echo extrapolation uses a per-channel output block to fix cross-channel contamination and shows gains from augmentation.
A hybrid 3D CNN and ConvLSTM encoder-decoder fuses radar and 3D re-analysis fields to nowcast convective storms, outperforming extrapolation baselines.
MLC-LSTM with adversarial training extrapolates multi-level radar echoes, reaching average POD of about 0.65 for short-term echo forecasts.
Data-parallel CNN training across GPUs cuts precipitation nowcasting training time from 59 hours to about one hour.
Generative adversarial ConvGRU model produces sharper radar echo extrapolations than ConvGRU and optical flow for precipitation nowcasting.
Neural networks and quantile trees nowcast precipitation growth and decay in Swiss Alps radar archive, with persistence inputs cutting RMSE by 20-30% in mountains.
StarBriNet, an RNN with sub-networks per rainfall intensity and a star-shaped bridge, improves radar-based precipitation nowcasting over prior methods.
Two-stage GANs generate synthetic radar reflectivity sequences for extreme precipitation in Shenzhen to address class imbalance in radar QPE.
LSTM and RNN models predict hourly rainfall levels from weather radar sequences, compared with random forest and XGBoost; gains are not quantified in the abstract.
Yandex develops satellite-imagery-based precipitation nowcasting, extending coverage beyond radar regions and deployed in the Yandex.Weather map and alerts.
Proposes FORECAST-CLSTM, a hierarchical ConvLSTM with a new forecaster loss for predicting infrared satellite cloud images, evaluated on a 40K-image dataset.
Stochastic variational frame predictor with a learned prior for radar precipitation nowcasting up to 2.5 hours, giving less blur than ConvLSTM in case studies.
Sequence-to-sequence ConvRNN for radar echo extrapolation trained with SSIM losses, reducing blurry 10-step precipitation predictions.
NARX network fuses GNSS precipitable water and station data to forecast hourly intense rain in Lisbon, cutting false-positive rate to 21–36%.
Applies deep learning to precipitation guidance for short-term weather prediction in Japan; the abstract is unavailable, so details are limited.
Deep networks on sky-camera videos estimate and forecast solar irradiance up to 4 hours ahead, reducing error relative to satellite-based approaches.
Combines optical flow with machine learning for nowcasting and early warning of strong convective weather; abstract not available.
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.
Develops and interprets deep learning models for nowcasting convective hazards, with attention to model explainability.
Computer-vision pipeline with soft-mask inpainting fuses satellite and radar observations to extend precipitation nowcasting beyond radar coverage.
Applies machine learning to an early warning system for very short-term heavy rainfall; abstract unavailable, classification based on title only.
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.
An LSTM extrapolates cloud-top brightness temperature, then PERSIANN retrieves 0–6 h rainfall, beating RNN, optical flow and RAP over three US regions.
Bidirectional LSTM is compared with a 1D CNN for multivariate precipitation nowcasting, with sequence models favoured over feed-forward networks; no quantitative results are given.
Conference abstract proposing a convolutional neural network approach to weather nowcasting; no quantitative results are available.
Neural networks trained on CINRAD radar reflectivity and gauges nowcast Wuhan rain intensity, reaching over 50% accuracy at 36-minute lead for 5 dBZ.
Deep learning precipitation nowcasting paired with HPC-optimized data pipelines; abstract unavailable, so the contribution is inferred from the title only.
Recurrent dynamic convolutional network extrapolates radar echo images for short-term nowcasting on Chinese radar data with higher accuracy than traditional methods.
Machine learning applied to weather radar data for severe weather nowcasting; abstract unavailable, so details are inferred from the title only.
Residual convolutional networks are applied to spatiotemporal inputs to investigate deep-learning precipitation nowcasting; abstract is truncated, so results are unknown.
3D convolutional network for regional precipitation nowcasting; abstract missing, so details are limited to the title.
TrajGRU learns location-variant recurrent connections for radar nowcasting and is benchmarked on a Hong Kong Observatory dataset, outperforming ConvLSTM.
ConvLSTM recasts precipitation nowcasting as spatiotemporal sequence prediction, outperforming FC-LSTM and the operational ROVER algorithm on radar rainfall data.
A Conv1D-Transformer predicts 3-hour rainfall levels in northern Xinjiang from 24 hours of station data, a regional semi-arid case study.
Cascaded diffusion models forecast tropical cyclone tracks and precipitation up to 36 hours from satellite and atmospheric data, running in 30 minutes on a single GPU.
Live experiments with XGBoost, MLP, Transformer and LSTM nowcast temperature, wind and cloud cover; a variance-horizon method cuts compute by over 50% and concept-drift retraining cuts it by up to 90%.
Review of deep learning solar forecasting from sky cameras, satellites and weather stations that tracks cloud displacement, outlining gaps relative to NWP.
Applications & impactsUniversity of Cambridge (Aardvark)Stanford University
A CNN (PCINet) identifies day and night precipitation clouds from FY-4A imagery, outperforming five segmentation baselines and enabling near-real-time monitoring.
Observations & remote sensingNanjing University of Information Science & Technology (NUIST)
A fully convolutional network forecasts spatially distributed solar radiation over time, targeting renewable energy applications.
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.
U-Net reconstructs radar reflectivity from WRF simulations and Himawari-8 satellite data with 5.38 dBZ RMSE, revealing scale-separated atmospheric features.
Observations & remote sensingInstitute of Atmospheric Physics, Chinese Academy of Sciences
Random-forest blending of sky-imager, satellite and data-driven forecasts improves 90-minute solar irradiance nowcasts by up to 30% for GHI and 40% for DNI.
Uses an AlexNet-based CNN to identify tropical cyclones from aerial imagery for an alert system; a basic classification application.
Multimodal sky-image and meteorological end-to-end learning nowcasts renewable power and cuts system cost 30% versus unimodal baselines in a Dutch case study.
Probabilistic structural forecasts simulate geostationary IR radial profiles with PixelSNAIL and feed a CNN, giving intensity errors near NHC official forecasts.
Spatiotemporal graph neural network for rainfall forecasting; no abstract was provided, so the assessment rests on the title.
Sky-camera images feed ANN and SVM models that nowcast photovoltaic output 3 hours ahead, with errors close to the uncertainty of the solar resource itself.
Survey of feature engineering, convolutional, LSTM and attention methods for spatio-temporal patterns in meteorological image sequences, illustrated with solar forecasting and convection detection.
A ViT-style transformer handles irregular sensor data directly, cutting wind nowcasting RMSE from 9.24 to 7.98 and beating GNN baselines across several domains.
SVM, tree and KNN models using camera-derived visibility predict 30-minute fog onset at Poprad-Tatry airport with a 0.89 probability of detection.
Compares low-resource CNN models nowcasting UK solar PV yield from satellite imagery over 1 to 4 hours, weighing carbon costs against benefits.
Neural network blends two precipitation ensembles into calibrated hourly exceedance probabilities for +1 to +6 h, outperforming both inputs at all lead times.
Cluster-wise typhoon rainfall model combines SVM, RF, AdaBoost, CNN and LSTM with grid-search tuning, claiming 45–90% forecast efficiency gains for 1–6 h lead times.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)
Feature-constrained Zooming Slow-Mo network spatiotemporally downscales rain radar fields, outperforming U-Net variants and interpolation on a public radar dataset.
Downscaling & super-resolutionNanjing University of Information Science & Technology (NUIST)
A CNN-LSTM predicts cloud-cluster evolution from geostationary IR imagery to rolling-correct 1-6 h GRAPES_Meso precipitation forecasts over China, improving rain/no-rain scores.
Post-processing & calibrationChina Meteorological Administration (CMA)
Trains an LSTM on COAMPS ensemble output to predict Typhoon Mangkhut's intensity and identify drivers of rapid intensification from asymmetry and heat flux features.
Self-supervised deep networks interpolate radar precipitation to 5-minute resolution, cutting error by up to 20% versus linear interpolation over southern France.
Uses CNN models on IMERG satellite precipitation to forecast near-real-time rainfall from hurricanes, targeting flood-relevant extreme rainfall.
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)
Self-attention augmented convolutions with a highway layer forecast extreme precipitation up to 12 hours ahead from pressure and wind, outperforming CNNs by 12%.
Reviews a CNN and data-mining based real-time cloudburst prediction system motivated by the 2023 Sikkim flood; no quantitative results reported.
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.
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.
Observations & remote sensingNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Machine learning regression on ten years of airport station data nowcasts visibility and ceiling height up to 1 h, with MAE of about 324 m and 167 m.
Local & station forecastingEnvironment and Climate Change Canada (ECCC)
Dropout neural networks with neighboring satellite pixel inputs produce ensemble solar irradiance forecasts with post-processing for renewable energy applications.
A CNN classifies rain events and a GRU forecasts hourly rainfall in Thailand, with focal and autoencoder losses. It outperforms ARIMA and CNN-GRU on RMSE in most regions.
Deep learning forecasts satellite rain fade from satellite and radar imagery plus link power, outperforming prior methods in near- and long-term predictions.
LightNet+ fuses radar, station and NWP data in a deep network for lightning forecasting, beating three established schemes and improving as data sources are added.
Extremes & hazardsFudan University (FuXi)China Meteorological Administration (CMA)
RAIN-F+ fuses radar, station and satellite data over Korea and provides U-Net benchmarks, showing fusion beats radar-only precipitation nowcasts.
U-Net-style CNN reconstructs 35 dBZ radar reflectivity from Himawari-8 infrared channels over North China, a proxy product for regions without radar.
Observations & remote sensingInstitute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
QPENet deep learning on dual-polarization radar, trained on 11 typhoons, beats DSD-based parametric QPE for rainfall estimation in South China.
Observations & remote sensingNanjing University of Information Science & Technology (NUIST)Colorado State University (incl. CIRA)China Meteorological Administration (CMA)
GRU-CNN models forecast typhoon winds and wave heights 1 to 6 h ahead at Taiwan buoys using station data and radar reflectivity images.
ConvLSTM precipitation forecasts over India at 25 km reach correlation 0.67 at lead day 1 and 0.42 at lead day 2.
Artificial neural network predicts rainfall rate 30 seconds ahead at Durban from three prior measurements, for link-quality planning; the abstract is garbled and the study is small.
GeoDIP, an HPC data-integration platform, merges satellite and reanalysis inputs and raises 2-hour precipitation prediction F1 by 9% with a 97% speedup.
A multi-task CNN framework with consistency constraints and cross-branch attention improves Himawari-8 infrared precipitation estimation by 3–5% F1 across intensities.
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.
Observations & remote sensingChina Meteorological Administration (CMA)
Compares logistic regression, LDA, trees, SVM and random forests for precipitation-area classification from S-band radar, with spatial features helping.
Deep learning with wavelet transforms nowcasts wind power and wind power ramps in Brazil and Uruguay; no abstract was available, so this is judged from the title.
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.
Observations & remote sensingColorado State University (incl. CIRA)
Neural networks combine RAP forecasts, surface analyses and GOES-16 satellite data to correct temperature and moisture profiles for severe weather nowcasting.
Semi-supervised ConvLSTM encoder-decoder with dual label guessing forecasts aviation air turbulence from weather data, improving accuracy under scarce turbulence labels.
Characterizes compute, time and energy costs of training a UNet for precipitation nowcasting, showing mixed-precision and distributed training cut training time without losing accuracy.
Releases CloudCast, a 70,080-frame satellite dataset of 10 cloud types over Europe at 3 km and 15-min resolution, with ConvLSTM, GAN and optical-flow video prediction baselines.
Random forest and other supervised classifiers combine weather radar and NWP data to classify winter precipitation type, beating a temperature and thickness threshold baseline.
CNNs using radar images and proximity soundings predict next-hour tornado occurrence with AUC above 0.9, with a GridRad critical success index of 0.31.
Extremes & hazardsNSF NCAR / UCARUniversity of Oklahoma (AI2ES)
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.
Uses an RNN with dropout to predict tropical storm trajectories with Bayesian intervals, giving evacuation-relevant uncertainty ranges from about 500 Atlantic storms.
Stacked denoising autoencoder with logistic regression predicts thunderstorm occurrence up to 15 hours ahead over 400 km2 South Texas, beating shallow networks and forecasters.
Random forest on HRRR surface fields produces gridded CONUS probabilistic subfreezing road temperature forecasts, with a 2-h Brier skill score of 0.66.
Post-processing & calibrationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
U-Net with shortcut connections corrects NWP precipitation outputs for Kansai heavy rain, improving strong-rain representation over plain CNNs.
Coevolutionary evolutionary-programming postprocessor improves 72-h surface temperature forecasts and 60-min convection nowcasts over bias-corrected NWP and logistic regression.
Post-processing & calibrationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Deep optical-flow interpolation upsamples GOES-R ABI imagery from 15 to 1 minute, improving severe-weather temporal resolution over baselines.
Random forests, boosted trees and elastic nets predict storm longevity in NWS quasi-operational testing, with lower error than human forecasters.
Extremes & hazardsNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Applies machine learning to satellite data to nowcast photovoltaic energy output, linking cloud and irradiance nowcasting to solar power.
A CNN corrects cloud-resolving model precipitation output for very short-term rainfall forecasts, trained against JMA analysed rainfall over urban cells.
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.
Uses a GAN on satellite images plus reanalysis fields to predict typhoon centre tracks 6 hours ahead, with 67.2 km mean error versus 95.6 km for images alone.
Compares MLP, CNN and LSTM models using sky images to nowcast one-minute photovoltaic output; LSTM gives a 21% RMSE skill gain over persistence.
A GAN predicts one-step-ahead typhoon centres and cloud structure from 6-hourly satellite images, with 42% of centre errors under 80 km.
Compares generalized linear models and neural networks for rainfall retrieval from geostationary IR data, using OPERA radar as truth; neural networks reduce false alarms.
LSTM neural network model for typhoon track nowcasting; abstract not available, classified from title.
An early neural network predicts tornadoes from Doppler radar-derived storm attributes. Abstract not available; classification based on title.
Extremes & hazardsNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
