A TRANSFORMER APPROACH FOR MULTI ORBIT PER PIXEL TIME SERIES FOREST CHARACTERIZATION WITH SENTINEL-1

Abstract

The Sentinel-1 (S-1) microwave measurements pose a unique opportunity for estimating forest parameters from satellite time series with increased data availability from overlapping orbits. Leveraging data from different orbits comes with varying viewing geometries and acquisition schedules, which can be considered in artificial neural networks. We adapt a transformer architecture for mapping the full S-1 data against median forest height values. By doing so, we propose per-orbit temporal encoders to handle different acquisition times by position, missing data by attention masking, and the addition of viewing geometry context. We show that our adjustments improve performance, with a greater impact of the additional data and a slight improvement in the masking. By optimizing the hyperparameters, our proposed method achieves a preliminary RMSE of 5.9 m and an rRMSE of 35 % in predicting the per-pixel vertical median forest height.