Adversarial U-Net Adaptation with Targeted Augmentation Boosts Crop Classification in Data-Scarce Regions

Deep learning models for crop classification are crucial for food security (SDG 2) but often fail when deployed in new geographic regions due to domain shift. This is a major barrier in nations like Algeria, which lack large-scale labeled datasets. To address this, we propose an adapted Domain-Adversarial Neural Network (DANN) that effectively transfers knowledge from data-rich European regions to data-scarce Algerian environments for Sentinel-2 imagery. Our methodology is centered on a U-Net segmentation architecture trained with a DANN framework. Our primary contribution is the introduction of a feature-matching loss at the U-Net bottleneck, which forces the model to learn more robust, domain-invariant representations. To address the limited availability of local labeled data, we apply a targeted data augmentation pipeline (including random rotations and scaling) to the small set of labeled Algerian wheat and potato parcels. The proposed model demonstrates significant performance gains. A baseline U-Net trained only on European data achieved 62% accuracy on the Algerian test set. In contrast, our adapted DANN model, trained with only 50% of the available Algerian labels, increased the overall accuracy to 89%. This data-efficient approach yielded high class-specific F1-Scores of 0.93 for wheat and 0.89 for potatoes. This work provides a validated and scalable pathway for developing accurate crop classification systems in regions with limited data.