Few Shot Crop Mapping Using Transformers and Transfer Learning with Sentinel-2 Time Series: Case of Kairouan Tunisia

Two transformer-based deep learning models, PSE-TAE and PSE-LTAE, achieved nearly 93% accuracy for detailed land cover mapping in Kairouan, Tunisia, using Sentinel-2 satellite imagery. This breakthrough demonstrates the effectiveness of transfer learning from pre-trained models (e.g., French LPIS) to adapt to new, data-scarce regions with limited labeled examples. The study leverages multi-headed self-attention mechanisms, outperforming traditional deep learning methods like CNNs and RNNs in agricultural contexts. By combining pixel-set encoders with temporal self-attention, the models generalize well even with few in-field observations, offering a scalable solution for precision agriculture in regions lacking comprehensive land parcel data.

Keraani, M. K., Mansour, K., Khlaifia, B., & Chehata, N. (2022). FEW SHOT CROP MAPPING USING TRANSFORMERS AND TRANSFER LEARNING WITH SENTINEL-2 TIME SERIES: CASE OF KAIROUAN TUNISIA. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B3-2022, 899–906. https://doi.org/10.5194/isprs-archives-xliii-b3-2022-899-2022