SAR2ET: End-to-end SAR-driven Multisource ET Imagery Estimation Over Croplands

Evapotranspiration (ET) is a crucial parameter in agriculture as it plays a vital role in managing water resources, monitoring droughts, and optimizing crop yields across different ecosystems. Given its significance in crop growth, it is essential to measure ET accurately and continuously to conduct precise analyses in agriculture. However, the continuous monitoring of ET changes is very challenging: while in-situ measurements are costly and not feasible for covering a wide geography, remote sensing-based ET products are typically dependent on optical satellites that can not operate and transmit data under certain weather conditions, especially in the presence of clouds. In this paper, we present the first comprehensive study on predicting ET from synthetic aperture radar (SAR) imagery, which we refer to as SAR2ET. Our work is motivated by the fact that SAR has the critical advantages of being all-weather available and sensitive to crop and soil changes. In handling the SAR2ET problem, we additionally incorporate non-optical meteorological and topographical input data from auxiliary data sources. We approach SAR2ET as a multi-modal image-to-image translation task, for which we train a UNet-shaped network. To evaluate the effectiveness of SAR-based ET predictions, we construct a benchmark data set over a large geographical region with image samples covering a whole agriculture season. Our experimental findings on this data set suggest that (i) the proposed approach leads to strong results, (ii) valuable information can be extracted from both SAR and auxiliary data sources, and (iii) SAR2ET is overall a promising research direction towards obtaining data-driven year-round ET estimates. The benchmark data set will be shared publicly upon publication to stimulate future work.