BotWaveNet: semantic segmentation network for remote sensing images based on wavelet transform and attention mechanisms

Accurate semantic segmentation of remote sensing (RS) imagery faces significant challenges due to the presence of small targets, ambiguous features, and complex scenes. Conventional convolutional networks, although computationally efficient, do not adequately perceive the global context. In contrast, transformers excel in global context awareness but incur a higher computational cost. To bridge this gap, we introduce BotWaveNet, an innovative semantic segmentation framework designed for RS images. This framework is inspired by the global context proficiency exhibited by BoTNet. BotWaveNet reconfigures the fourth layer of ResNet by replacing conventional convolutional operations with a spatial attention mechanism (SAM), thus establishing an encoder–decoder structure that integrates attention mechanisms within deep network architectures. The SAM module is specifically engineered to enhance edge delineation by promoting pixel interconnectivity, thereby capturing global context more effectively. Furthermore, we have designed a feature fusion module (wavelet attention module) that fine-tunes the extraction of channel features, optimizing adaptability for both synthetic aperture radar and optical images. In an effort to reduce the computational demands associated with MaxPooling, we have incorporated a wavelet transform module. This module serves as an innovative replacement for traditional convolutional pooling methods, aiming to maintain efficiency while enhancing performance. The efficacy of BotWaveNet has been substantiated through its significant performance enhancements observed in the Gaofen Challenge and the Inria Aerial Image Labeling Dataset.