Exploiting SAR Time Series for EEFLUX-based Evapotranspiration Predictions via Attention LSTM Over Cotton Fields

One of the great strengths of black-box model predictions in biophysical parameter estimation is their increasing accuracy, which improves day by day with the availability of publicly labeled Earth Observation (EO) data. For all its faults, the black-box models demonstrate remarkable accuracy. However, these models can only be powerful tools for policymakers if they can be explainable and reproducible. Therefore, it is vital to understand the underlying relationships behind these prediction models, especially since EO data are inherently self-explanatory in terms of physics. In this paper, we discuss the performance of the prediction of evapotranspiration (ET) using Evapotranspiration Flux (EEFlux) acquired from the Earth Engine as a target variable, using Sentinel-1/-2 data coupled with meteorological and topographic data. To understand how these data and ET interact during the phenology of crops, we use dynamic attention mechanisms and Long Short Term Memory (LSTM) approach to build a prediction model, as these relationships are influenced by physical processes and temporal dynamics inherent in crop growth. For this reason, we try to interpret whether temporal attention weights are meaningful in terms of crop science or whether the properties (temporal, spatial resolution, etc.) of the input images dominate to leverage physical relationships among the inputs.Experimental results with cotton fields (~ 14K) in the South-eastern Anatolia region of TUrkiye reveal that the proposed method provides high prediction performance (coefficient of determination is 0.84, root mean squared error is 0.005 mm) with physically sound temporal attention weights.