A hierarchical foundation model framework for generalizable agricultural and crop parcel delineation

• A novel PT-PEFT SAM framework adapts foundation models for agricultural mapping. • A hierarchical strategy uses fine-tuned and zero-shot SAM for multi-scale parcels. • Enables a scalable ‘train global, fine-tune local’ mapping paradigm. • Achieves state-of-the-art accuracy on large-scale European and Chinese datasets. Precise, multi-scale mapping of agricultural landscapes—ranging from management-oriented Agricultural Parcels (APs) to fine-grained Crop Parcels (CPs)—is critical for land administration and precision agriculture. However, automated delineation remains challenging due to the accuracy limitations and poor generalizability of existing methods. While the Segment Anything Model (SAM) offers powerful zero-shot segmentation capabilities, its direct application is hindered by a lack of semantic awareness, a domain gap with RS imagery, and the prohibitive costs of full fine-tuning. To address these limitations, this study introduces PT-PEFT SAM, a novel framework that efficiently adapts SAM for agricultural mapping. A sequential “guide-then-specialize” adaptation strategy is employed, utilizing Prefix-Tuning (PT) to align SAM’s representations with the remote sensing domain, followed by a proposed Orthogonalized Multi-Low-Rank Adaptation (Ortho-Multi-LoRA) to inject task-specific semantics efficiently. The framework is further augmented by a learnable spectral adaptation module for multi-band compatibility and a domain-invariant feature learning module to enhance intrinsic generalization. Furthermore, a hierarchical architecture is introduced: the fine-tuned model performs semantic AP delineation, while a fully automated pipeline leveraging the pre-trained SAM 2 is deployed for intra-parcel CP segmentation, exploiting its superior zero-shot capabilities. Evaluations across European and Chinese datasets demonstrate that our method achieves state-of-the-art accuracy (Boundary F1 > 0.92 for APs; Geometric Error [GTC] < 0.065 for CPs) and exceptional transfer learning performance. This research establishes a principled, scalable paradigm for adapting foundation models to specialized geospatial tasks, enabling the rapid and accurate generation of critical agricultural data.