Multisensor Machine Learning to Retrieve High Spatiotemporal Resolution Land Surface Temperature
In one sentence. 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.
Abstract
Climate change is making heat waves more frequent, long-lasting, and severe. While multiple satellite types provide data to monitor surface temperature, geostationary (GEO) sensors provide near-continuous, continental-scale observations which can better capture the diurnal variability of land surface temperature (LST) than intermittent observations from low-earth orbit (LEO) sensors. However, standard products from GEO satellites are available at coarsened spatial and temporal resolutions compared to the native sensor resolution. Using datasets from the NASA Earth Exchange, we leveraged co-located, co-temporal observations from LEO and GEO satellites to learn a data-driven mapping using a convolutional neural network. The resulting NASA Earth eXchange Artificial Intelligence LST (NEXAI-LST) achieved a mean absolute error of 1.73 K relative to the target LEO product and improves on both spatial and temporal resolution [2 km, 10 minute] compared to the GEO full disk standard product [10 km, hourly]. In validation against measurements from a ground-based sensor network, NEXAI-LST achieves similar or better fit than both LEO and GEO standard products, while depending none of the prior knowledge of land surface and atmospheric states required by physical-statistical models. Further, application of the model to unseen LEO and GEO satellites demonstrates robust generalization of the model across spatial region, time of day, and sensor. In support of NASA’s open-source science initiative, we make our NEXAI-LST product, model, and codes available to facilitate data exploration and further studies.
Labels
- Topic
- Observations & remote sensing
- Type
- New model or method
- Method
- CNN / U-Net
- Domain
- Regional
- Data
- Satellite
- Status
- Research
- Identifiers
- arXiv:2022.31986 DOI:10.1109/access.2022.3198673 OpenAlex:W4291653582
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.