Multi-source heterogeneous feature fusion and multi-scale edge enhancement method for open-pit mining land cover segmentation

Accurate land cover classification in mining areas is essential for environmental monitoring and resource management. However, this task is hindered by complex spectral-spatial characteristics and highly fragmented landscapes, where traditional single-source spectral analysis often fails to resolve “spectral ambiguity” among disparate land covers. To address this, we propose a Dual-Branch UNet (DBUNet), a dual-branch architecture that explicitly leverages multi-source heterogeneous features–integrating physicochemical spectral data, geometric topographic profiles (DTM), and biological vegetation indices (NDVI). The core innovations of DBUNet include: First, a dedicated branch for processing topographic and NDVI information, coupled with a Cross-branch and Cross-scale Interaction (ICMC-Trans) module and an Adaptive Fusion Module (AWS-Link), which synergistically captures discriminative elevation and vegetation priors. Second, an Edge Feature Enhancement (EFE) module employing learnable multi-scale Sobel kernels to adaptively refine complex boundaries. Moreover, we introduce a boundary-preserving loss function enhanced by the Segment Anything Model (SAM) to further supervise fine-grained segmentation. Experimental results on the Land Cover in Mining Areas (LCMA) dataset demonstrate that DBUNet achieves state-of-the-art performance with a 34.03% mIoU, outperforming recent Mamba and Transformer-based models. Furthermore, its robust generalization on the Gaofen Image Dataset (GID) underscores its transferability. The source code is available at: https://github.com/Venti666/DBUNet.