Enhancing Crop Type Mapping with Out-of-year Data

The systematic acquisition of satellite data has enabled the production of up-to-date crop maps, essential for agricultural management, environmental monitoring, and informed decision-making. However, generating these maps typically requires high-quality labeled data, which is costly and time-consuming to collect. To alleviate the need for extensive labeled data, we propose a deep learning framework called REFeD (data Reuse with Effective Feature Disentanglement for crop mapping). REFeD leverages out-of-year reference data to improve the production of current crop maps by combining remote sensing and reference data from different domains (e.g., historical and recent data) via a disentanglement strategy based on contrastive learning. By separating domain-invariant and domain-specific features, REFeD addresses distribution shifts between data sources, enhancing the crop mapping process. Experimental validation on the Centre-Val de Loire region in France confirms the effectiveness of the proposed approach.