Answers are grounded in this paper and cite their sources.
Accurate island segmentation in high-resolution remote sensing images is essential for ecological conservation. Although deep learning-based methods have achieved significant breakthroughs, the low class separability and the imbalanced class distributions on islands still pose challenges for accurate segmentation. In this article, we propose a wavelet-enhanced U-shaped Mamba network (WTUMamba) for island segmentation. This network comprises two core modules. First, a multilevel frequency-enhanced wavelet convolution module is proposed to enhance the discrimination of features among different ground objects via the multiscale decomposition and reconstruction of wavelet transforms. Second, a spectral-spatial enhanced state space module is designed to enhance feature modeling performance under class-imbalanced conditions while preserving linear computational complexity. Moreover, we construct a new island dataset, namely the Sun Yat-sen University Island Semantic Segmentation dataset, to evaluate the effectiveness of WTUMamba. This dataset comprises 17 islands in the South China Sea region, covering a total area of 12.57 km<tex-math notation="LaTeX">²</tex-math> and providing manually annotated pixel-wise labels. Additional experiments on the Gaofen Image Dataset are also performed to further evaluate the segmentation performance of WTUMamba. Experimental results demonstrate that WTUMamba achieves superior performance over state-of-the-art methods across core evaluation metrics.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: