Multi-level Edge Priors Guided Progressive Network for Fine-Grained Cultivated Land Delineation

Fine-grained delineation of cultivated land parcels from optical remote sensing imagery is essential for agricultural monitoring and land resource management. However, existing approaches are often constrained by limited internal structural consistency and inadequate spatial separability, resulting in fragmented predictions and blurred boundaries. To address these limitations, a Multi-level Edge pRiors Guided progrEssive Network (MergeNet) is developed to enhance plot-specific feature representation and improve both intra-parcel coherence and inter-parcel discrimination. MergeNet adopts Res2Net-50 as the backbone to extract hierarchical features and integrates a progressive edge-guided decoding strategy consisting of three functional modules. Firstly, an edge enhancement module (EEM) is designed to strengthen boundary sensitivity and introduce complementary edge cues. Secondly, a structure-aware semantic enhancement module (SSEM) promotes semantic completeness through adaptive multi-scale feature aggregation. Thirdly, an edge-guided semantic refinement module (ERSM) exploits edge priors derived from the input image, EEM, and SSEM to progressively refine semantic features and improve boundary delineation. Deep supervision is applied at each decoding stage to facilitate network optimization and reinforce hierarchical feature learning. Extensive experiments conducted on three datasets, including FGFD (0.3 m), iFLYTEK (0.75-1.1 m), and a newly constructed Netherlands dataset (10 m) that covers diverse cultivated landscapes, demonstrate that MergeNet consistently outperforms five state-of-the-art (SOTA) methods in terms of accuracy and structural quality. The source code will be publicly available at https://github.com/lixinghua5540/MergeNet.