High resolution evapotranspiration map over Austria: leveraging AI and Sentinel-2

This paper presents a deep learning pipeline for high-resolution evapotranspiration (ET) estimation over Austria using Sentinel-2 multispectral imagery and a digital elevation model as inputs, and spatially-sharpened ECOSTRESS measurements as the training reference. The model, a ResNet18-based U-Net architecture pre-trained on Sentinel-2 data, produces daily ET maps at 10-metre resolution, bridging the gap between the 70-metre ECOSTRESS footprint and the spatial detail required for urban and agricultural applications. A validation study is presented at the AKH Vienna site, comparing model outputs against raw ECOSTRESS acquisitions and field lysimeter measurements. Results confirm that the model accurately reproduces its training target, achieving consistent agreement with ECOSTRESS across 2022–2024. Comparison with ground-truth lysimeter data reveals a systematic positive bias, identified as partially inherited from the ECOSTRESS training signal itself. The implications for training data selection, model improvement, and broader validation strategies are discussed.