T³S: Think in Thermal Time for Generalizable Crop Mapping from Satellite Image Time Series

Official repository for T³S (Thermal Time-based Temporal Sampling), a simple, model-agnostic approach for improving the generalization of crop mapping models across years and regions.

🏆 Accepted at the 11th Workshop on Computer Vision in Plant Phenotyping and Agriculture (CVPPA), ECCV 2026, Malmö, Sweden

📄 Paper: arXiv:2506.12885
🌱 CVPPA 2026: Workshop Proceedings

Code release: implementation and reproducibility scripts will be added to this repository.

---

TL;DR

Crop development does not strictly follow calendar time. Depending on temperature, the same crop can reach the same phenological stage substantially earlier or later across different years.

T³S replaces calendar-time sampling with thermal-time sampling. Satellite observations are re-indexed using cumulative growing degree days (cGDD), so that observations corresponding to similar crop development stages are more consistently aligned across years.

This simple input-level strategy:

• improves cross-year generalization,
• substantially improves uncertainty calibration,
• remains effective in low-data and early-season settings,
• requires no architectural modification, and
• can be applied to different model families.

---

Method

T³S uses accumulated temperature as a proxy for crop development.

Overview of T³S. Instead of sampling observations at uniform calendar-time intervals, T³S re-indexes the season using cumulative growing degree days (cGDD) and samples observations at equal thermal-time intervals.

For day $i$, growing degree days are computed as

$$\mathrm{GDD}_i = \max\left( 0,\, \frac{T_{\max,i} + T_{\min,i}}{2} - T_{\mathrm{base}} \right).$$

The cumulative growing degree days up to day $d$ are then

$$\mathrm{cGDD}_d = \sum_{i=1}^{d} \mathrm{GDD}_i.$$

Instead of sampling satellite observations uniformly in calendar time, T³S samples them uniformly along the thermal-time axis.

The procedure consists of four steps:

1. Compute daily thermal units using growing degree days.
2. Accumulate GDD into cGDD, re-indexing the season from calendar time to thermal time.
3. Divide the cGDD range into equal thermal-time intervals.
4. Select the least-cloudy Sentinel-2 observation from each interval.

The resulting sequence can be passed directly to the downstream model.

No architecture changes are required.

Calendar time tells us when an observation was acquired; thermal time better reflects where the crop is in its development.

Because thermal time accumulates faster under warmer conditions and slower under colder conditions, T³S naturally adapts the temporal sampling density to crop development.

---

Results

On the SwissCrop cross-year benchmark, T³S consistently improves both predictive performance and uncertainty calibration.

Method	Accuracy ↑	mIoU ↑	IoU ↑	ECE ↓
U-TAE	71.2	17.3	55.3	5.1
MC-Dropout	71.9	17.3	56.1	3.4
Thermal Positional Encoding	71.3	17.1	55.5	7.7
Deformable Sampling	75.4	20.5	60.5	6.3
T³S + U-TAE	77.0	21.5	62.6	1.1

Compared with the standard calendar-time U-TAE baseline, T³S improves accuracy from 71.2% to 77.0% while reducing expected calibration error from 5.1% to 1.1%.

T³S also improves performance when applied to a pretrained Earth observation foundation model, showing that the approach is not tied to a particular architecture.

---

SwissCrop Dataset

The paper introduces SwissCrop, a country-scale, multi-year Sentinel-2 crop mapping dataset for Switzerland paired with daily temperature data.

SwissCrop is distributed as part of the extended SwissCrop25 benchmark.

Dataset

👉 SwissCrop25 on Hugging Face

SwissCrop contains the data used for the cross-year experiments in the T³S paper and is included within the broader SwissCrop25 benchmark.

The dataset combines:

• multi-year Sentinel-2 satellite image time series,
• field-level crop labels,
• daily temperature information, and
• country-scale coverage of Switzerland.

If you use the SwissCrop subset introduced with T³S, please cite both the T³S paper and the SwissCrop25 dataset paper.

---

Citation

If you use T³S, please cite:

@article{turkoglu2025t3s,
  title         = {{$T^{3}S$: Think in Thermal Time for Generalizable Crop Mapping from Satellite Image Time Series}},
  author        = {Turkoglu, Mehmet Ozgur and Ledain, Selene and Zweidler, Jeffrey and Lauber, Thomas and Aasen, Helge},
  journal       = {arXiv preprint arXiv:2506.12885},
  year          = {2025},
  eprint        = {2506.12885},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2506.12885}
}