T^3S: Thermal time-based crop mapping beats calendar methods
Replacing calendar dates with growing degree days boosts cross-year crop classification accuracy.
Crop type classification from satellite imagery typically uses calendar dates, but inter-annual weather shifts cause phenological mismatches, limiting generalization. Inspired by ecophysiology, researchers from ETH Zurich and partners introduce T^3S (Thermal Time-based Temporal Sampling), which re-indexes satellite observations by cumulative growing degree days. This simple, model-agnostic preprocessing aligns biologically equivalent growth stages across seasons, reducing temporal redundancy while focusing on informative periods.
Tested on three distinct model backbones using the new SwissCrop dataset (country-scale, multi-year Sentinel-2 with paired temperature) and the cross-region TimeMatch benchmark (Denmark and France), T^3S consistently outperforms state-of-the-art baselines including thermal positional encoding. Key gains include robust cross-year and cross-region generalization, better uncertainty calibration, strong performance under label scarcity, and earlier season predictions—all without modifying model architecture.
- T^3S re-indexes satellite time series by cumulative growing degree days instead of calendar time
- Evaluated on 3 backbones using new SwissCrop dataset and TimeMatch benchmark across 3 countries
- Improves cross-year classification, uncertainty calibration, and early-season prediction over baselines
Why It Matters
Enables operational crop mapping without current-year labels, critical for food security amid climate variability.