WTUNet: a U-shaped semantic segmentation network combined with wavelet convolution

Landslide extraction plays a critical role in facilitating effective rescue operations and emergency response. However, existing methods face difficulties in accurately extracting landslide areas due to complex terrain conditions, where variable textures, shapes, and spectral features can obscure critical details. These challenges are particularly evident in texture features, which spatial-domain methods alone struggle to capture reliably under varying lighting and background conditions. To address these issues, this study introduces WTUNet, a novel U-shaped semantic segmentation model that integrates frequency-domain feature extraction to enhance texture representation. WTUNet’s key innovation is the integration of a wavelet transform layer following each convolutional layer, enabling the model to capture frequency-domain features through wavelet decomposition while retaining spatial features from the convolutional layers. This dual-domain feature extraction approach is further optimized by dynamic feature fusion technology, which merges these diverse feature sets to enrich the semantic representation within feature maps, thereby substantially enhancing extraction performance. To evaluate WTUNet’s effectiveness, we conducted training and testing on the Bijie and Jiuzhaigou datasets, benchmarking it against three leading semantic segmentation models. Results demonstrate that wavelet transform integration meaningfully strengthens semantic feature representation, yielding at least a 2% Mean Intersection over Union(mIoU) improvement on the Bijie dataset and a 3% improvement on the Jiuzhaigou dataset, underscoring WTUNet's superior landslide extraction capabilities.