Spectral Representation-Enhanced Multitask SAM for Cropland Parcel Delineation

Accurate cropland parcel delineation (CLPD) from remote sensing imagery is crucial for fine-scale agricultural monitoring. The recent segment anything model (SAM), as a foundational vision model, demonstrates remarkable generalization capabilities, presenting significant potential for CLPD. However, SAM's design for RGB natural imagery limits its ability to model multispectral information effectively. To address these limitations, we propose a spectrally enhanced multitask segmentation framework built upon SAM for CLPD. Our approach enhances SAM in two key aspects: first, to better leverage multispectral data, we introduce a decoupled learning strategy designed to improve SAM's spectral representation. It is implemented through two components: learnable channel tokens (CTs) encoding band-specific spectral identities and a lightweight channel-attention module modeling interband relationships. Second, to enhance decoding robustness under high intraparcel spectral variability, we introduce a feature-structure regularization loss that enforces intraclass compactness and interclass separation, along with an auxiliary instance-guided decoding branch that captures fine-grained subparcel patterns. These patterns are integrated into the main segmentation stream via a token-based task interaction module. Extensive experiments on two public datasets demonstrate consistent improvements of the proposed method. On the Sen4AgriNet dataset, it achieves a pixel-level F1-score of 89.36% and reduces the object-level global total classification error (GTC) by 0.97% compared with existing methods. Ablation studies further validate the contribution of each component and confirm the effectiveness of our decoupling strategy. The source code will be made available at: https://github.com/halfhalfer/MSSAM.