Encoding Contextual Information by Interlacing Transformer and Convolution for Remote Sensing Imagery Semantic Segmentation

Contextual information plays a pivotal role in the semantic segmentation of remote sensing imagery (RSI) due to imbalanced distributions and ubiquitous intra-class variants. The emergence transformer intrigues revolution vision tasks with its impressive scalability establishing long-range dependencies. However, local patterns, such as inherent structures spatial details, are broken tokenization transformer. Therefore, ICTNet is devised confront deficiencies mentioned above. Principally, inherits encoder–decoder architecture. First all, Swin Transformer blocks (STBs) convolution (CBs) deployed interlaced, accompanied by encoded feature aggregation modules (EFAs) encoder stage. This design allows network learn patterns distant dependencies their interactions simultaneously. Moreover, multiple DUpsamplings (DUPs) followed decoded (DFAs) form decoder ICTNet. Specifically, transformation upsampling loss shrunken while recovering features. Together decoder, well-rounded context captured contributes inference most. Extensive experiments conducted on ISPRS Vaihingen, Potsdam DeepGlobe benchmarks. Quantitative qualitative evaluations exhibit competitive performance compared mainstream state-of-the-art methods. Additionally, ablation study DFA DUP implemented validate effects.