Self-supervised Sentinel-1 representations for crop-type mapping in Sindh, Pakistan

Details

Supervisors

Defourny, Pierre ; Bontemps, Sophie

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

Large-scale crop mapping has become essential in recent years, and remote sensing provides an effective method to achieve it. Optical sensors have been widely used for this purpose, yet they remain limited by atmospheric conditions such as cloud cover. Sentinel-1 Synthetic Aperture Radar (SAR), which overcomes these limitations, remains underutilised and underanalysed, particularly in Sindh, Pakistan, an agriculturally significant region with scarce ground-truth data. This thesis compares phase-stratified handcrafted features derived from Sentinel-1 and Sentinel-2 time series against self-supervised representations learned from unlabelled Sentinel-1, evaluating both for crop-type classification over the 2023-2024 Rabi season in Sindh, Pakistan. Seven experiments are conducted on 1,088 ground reference samples across five classes (Wheat, Banana, Mango, Other Crops, Non-Crop) under 5,000 m spatial block cross-validation. A self-supervised 3D convolutional encoder, the Spatio-Temporal Encoder (STEncoder), is trained on unlabelled Sentinel-1A time series using a Variance-Invariance-Covariance Regularization (VICReg) objective augmented with a temporal ordering loss to encode phenological progression, producing pixel-wise representations evaluated with a linear probe. Combining Sentinel-1 backscatter, interferometric coherence, and Sentinel-2 multispectral handcrafted features achieves overall accuracy (OA) = 0.884 and macro F1 = 0.879, against OA = 0.855 for Sentinel-2 alone. Late-season optical acquisitions are most critical: withholding them reduces OA to 0.863 and narrows the margin over the SAR-only self-supervised baseline to 0.013. Under SAR-only conditions, VV interferometric coherence accounts for around 10% of handcrafted feature importance. The self-supervised learned representations from STEncoder outperform phase-stratified handcrafted Sentinel-1 features (OA = 0.850 vs. 0.791), with the advantage widening to approximately 15 macro F1 percentage points below 20% label availability. The encoder also partially recovers coherence-like structure from backscatter alone, reducing the operational case for coherence under self-supervised representations. Self-supervised Sentinel-1 representations close within 0.005 OA of an optical baseline without using any optical data, and retain a label-efficiency advantage under limited annotation budgets.