Abstract
Crop type classification using optical satellite time series remains limitedin its ability to generalize across seasons, particularly when crop phenologyshifts due to inter-annual weather variability. This hampers real-worldapplicability in scenarios where current-year labels are unavailable. Inaddition, uncertainty quantification is often overlooked, which reduces thereliability of such approaches for operational crop monitoring. Inspired byecophysiological principles of plant growth, we propose a simple,model-agnostic Thermal-Time-based Temporal Sampling (T3S) method that replacescalendar time with thermal time. By subsampling time series in thisbiologically meaningful way, our method highlights key periods within thegrowing season while reducing temporal redundancy and noise. We evaluate theT3S on a multi-year Sentinel-2 dataset covering the entirety of Switzerland,which allows us to assess all applied methods on unseen years. Compared tostate-of-the-art baselines, our approach yields substantial improvements inclassification accuracy and, critically, provides well-calibrated uncertaintyestimates. Moreover, the T3S method excels in low-data regimes and enablessignificantly more accurate early-season classification. With just 10% of thetraining labels, it outperforms the current baseline in both accuracy anduncertainty calibration, and by the end of June, it achieves a performancesimilar to the full-season baseline model.