Image-To-Image Translation Networks for Estimating Evapotranspiration Variations: SAR2ET

Abstract

Evapotranspiration (ET) plays a significant role in understanding the water necessities of crops during their growing season, and hence, aids to make a decision in agriculture (planting time, applying fertilizer, irrigation, yield prediction and etc.). In this context, over the past few years, a wide range of research studies have been implemented for learning field-level ET from low-resolution ET products by downscaling and/or data fusion strategies. Unlike these previous studies, this research aims to leverage deep learning based models to learn ET from temporally and spatially dense imaging data; Sentinel-1 and climate data; ERA-5, both provided by Copernicus Climate Change Service. The model is formed by weak supervision from high spatial resolution Sentinel-1 coupled with climate data and analysis ready ET product as target. We evaluated the framework across two geographically distributed regions, namely; The Balkans and The Aegean in order to understand how well weak supervision estimates ET over croplands in different ecosystems.The code for the SAR2ET model is publicly available at https://github.com/Agcurate/SAR2ET, where you can access all the details regarding the model.