Combining Deep Learning with Physics-based Modeling for Improved Remote Sensing-based Evapotranspiration and Land Surface Temperature Estimates

Abstract

Evapotranspiration (ET) and land surface temperature (LST) are critical variables for understanding land-atmosphere interactions and are widely used for climate monitoring, agriculture, disaster response, and urban planning. Given the limited availability of ground-based observations, satellite-based remote sensing has emerged as an efficient approach for monitoring ET and LST due to broad spatial coverage and diverse temporal resolutions. The accuracy and continuity of the existing remote sensing-based ET and LST products are, however, limited by coarse spatial resolution, inconsistent temporal coverage, and gaps caused by cloud cover. This dissertation employs a stepwise approach to address the resolution, accuracy, and completeness issues of ET and LST estimates by leveraging multi-source remote sensing data, physics-based modeling, and advanced deep learning techniques. Chapter 1 provides a general introduction to the research. In Chapter 2, we developed and evaluated 30-m daily ET estimates using the Priestley-Taylor Jet Propulsion Laboratory (PT-JPL) model, driven by ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS), Harmonized Landsat Sentinel-2 (HLS), and ERA5-Land reanalysis data. The ET estimates are evaluated against eddy covariance (EC) tower observations at 145 sites across the contiguous United States (CONUS). Results show that higher spatial resolution vegetation information combined with continent-wide EC measurements and deep learning postprocessing can significantly reduce bias in the ET estimates. Chapter 3 introduces an Attention based Super Resolution deep Residual Network (ASRRN) to generate 100-m LST estimates from 1-km Moderate Resolution Imaging Spectroradiometer (MODIS) LST using ECOSTRESS LST, auxiliary remote sensing information such as HLS surface reflectance, Sentinel-1 synthetic aperture radar, and ASTER digital elevation data as reference. The ASRRN model outperforms traditional data fusion methods such as Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) and demonstrates strong agreement with target 100-m ECOSTRESS LST and in-situ measurements from the U.S. Climate Reference Network (USCRN). Chapter 4 presents a multi-stage deep learning framework (namely MAE-ASRRN) that integrates both supervised and self-supervised methods to produce gap-free, high-resolution daytime and nighttime LST estimates at 100-m resolution. In the first stage, a self-supervised Masked Autoencoder (MAE) is trained on complete LST scenes to reconstruct missing or cloudy pixels in partially incomplete MODIS and ECOSTRESS thermal observations. In the second stage, an ASRRN model is trained using coarse-resolution North American Land Data Assimilation System (NLDAS) LST and MAE-reconstructed MODIS LST to produce continuous, gap-free 1-km MODIS-like LST. Then, a second ASRRN maps the reconstructed MODIS-like LST to ECOSTRESS-like outputs, generating 100-m, spatiotemporally continuous daytime and nighttime LST estimates. The MAE component of the MAE-ASRRN framework effectively addressed the issue of missing pixels in MODIS and ECOSTRESS data, producing spatially complete images with low reconstruction errors, while the ASRRN component reduced temporal discontinuities and generated consistent, gap-free LST estimates at 100-m spatial resolution. The resulting MAE-ASRRN LST estimates showed strong agreement with in-situ measurements from USCRN stations. Time series evaluations across different sites and overpass times demonstrated that the 100 m LST estimates generated by the MAE-ASRRN framework captured seasonal LST variations well. Finally, Chapter 5 synthesizes key findings from the three studies and provides recommendations for future work. Overall, the work in this dissertation demonstrates the potential of integrating physics-based models, multi-source satellite observations, and deep learning to produce high-resolution ET and LST estimates that are both accurate and continuous in space and time, and contributes to the next generation of remote sensing-driven environmental monitoring systems that can be useful for climate resilience, agriculture, and sustainable land and water management.