Cross-Year Crop Mapping with Thermal Calendar from Optical Satellite Image Time Series

Traditional approaches for crop type classification from optical satellite images typically evaluate algorithms using training and test datasets from the same year and based on calendar days. However, this experimental setup is not practical for real-world applications due to (i) year-to-year variations in crop growth caused by climate, which limit generalization, and (ii) the inability to apply a model to the current year if trained on current-year data. This work addresses these challenges by introducing a cross-year experimental setting and incorporating thermal calendars into our deep learning model. Specifically, we train an attention-based deep learning model on the 2021 Swiss crop dataset, validate it in 2022, and test it in 2023. Thermal calendars, derived from accumulated daily average temperatures, align crop growth with thermal time instead of calendar time, addressing temporal shifts caused by climatic variations. Our results demonstrate that integrating thermal calendars improves performance compared to baseline using standard calendar encodings, achieving better generalization across years and showcasing the potential for large-scale operational crop classification.