Lightweight Remote Sensing Image Segmentation Network with Dilated Convolution and Data-Dependent Attention Decoder

Accurate object segmentation in high-resolution remote sensing imagery is essential for urban planning, infrastructure monitoring, and environmental analysis. However, large-scale semantic segmentation models often entail prohibitive computational costs, limiting their deployment in resource-constrained environments. This paper presents a lightweight encoder–decoder architecture that balances segmentation accuracy and computational efficiency. The proposed framework integrates a Light Dilated Convolution (LDC) module, which uses cascaded depthwise separable dilated convolutions to expand the receptive field with minimal parameters, and a data-dependent attention decoder that fuses data-driven upsampling with channel attention for refined spatial recovery. Experiments on the WHU Building (including synthetic fog), Massachusetts Building, and WHU Road datasets demonstrate that our method achieves competitive accuracy with substantially reduced complexity. Specifically, it attains an IoU of 89.04% on the WHU Building Dataset using only 1.89 million parameters, which just 6.09% of D-LinkNet’s size and 7.34% of its computational cost, while maintaining comparable accuracy. These results confirm the effectiveness of the proposed approach for efficient, high-precision segmentation under constrained computational budgets.