DENet: a dual-encoder network for land use classification

Land use classification serves as a cornerstone of natural resource management and urban planning. However, existing semantic segmentation methods often suffer from insufficient collaborative extraction of multiscale features, poor semantic feature fusion, and loss of spatial details when interpreting high-resolution remote sensing imagery. To address these issues, this paper proposes a Dual-Encoder Network (DENet) to improve the accuracy of land use classification from high-resolution remote sensing images. The network comprises three key modules: 1) A dual-encoder structure, which utilizes a pre-trained ResNet to extract local detailed features of objects and incorporates the Hiera architecture of SAM2 to capture global contextual information, enabling effective multiscale feature collaboration; 2) A feature fusion module that employs hierarchical normalization to unify feature distributions and adopts a channel-wise addition strategy to integrate outputs from the dual encoders, thereby eliminating semantic distribution discrepancies; 3) A decoder that introduces multilayer perceptrons (MLPs), an Efficient Channel Attention (ECA) mechanism, and cross-layer skip connections to enhance critical features while preserving spatial details. To validate the effectiveness of the proposed method, comparative and ablation experiments were conducted on the ISPRS Vaihingen dataset, achieving an mIoU of 83.52% and a Mean F1 score of 90.77%. Experimental results demonstrate that DENet effectively addresses the challenges of multiscale feature matching and spatial detail preservation in land use classification of high-resolution remote sensing imagery. Its classification accuracy and robustness outperform existing mainstream semantic segmentation models, offering reliable technical support for high-precision dynamic land use monitoring and sustainable natural resource management.