Efficient Semantic Segmentation of Remote Sensing Images Through Global-Local Feature Integration

The rapid acquisition of remote sensing information plays a significant role in the development of image semantic segmentation methods for remote sensing image interpretation applications. With the increasing variety and complexity of data recorded by satellite remote sensing images, accurately and effectively extracting information from remote sensing images has become crucial for interpreting these images using semantic segmentation methods. However, natural scene-based semantic segmentation methods often suffer from low segmentation accuracy when applied to high-resolution remote sensing images due to the complex background, diverse scales, and irregular shapes. To address these challenges, this paper proposes an efficient remote sensing image semantic segmentation model called Multi-GLISS, which integrates global and local features. The model captures global features through consecutive downsampling and Fourier transform while preserving spatial feature learning and boundary information using convolutional residual layers. Moreover, the local wavelet block utilizes wavelet transform to extract multi-scale local features, enhancing sensitivity to local textures and shapes. Subsequently, the model gradually increases the image resolution through a series of upsampling steps, reconstructing a high-resolution version of the image. Experiments on the Vaihingen dataset show that Multi-GLISS achieves a mean F1 score of 93.29, outperforming the baseline DeepLabV3+ by 1.17 points, while providing clear boundary and structural information for accurate and detailed segmentation.