A Swin Transformer-based method for Classification of Land Use and Land Cover Images

Satellite image classification plays a crucial role in land use analysis, environmental monitoring, and urban planning. Recent developments in computer vision have led to the development of algorithms for image classification that are becoming increasingly successful. These techniques are known as vision transformer. On the other hand, it is often important to overcome problems related with limited receptive fields and the need for complete training data if one wants optimum performance. This work aims to provide a fresh approach for enhancing the design of the Swin transformer thus improving the classification of land use and land cover on the Eurosat dataset. Depth-wise Separable Convolutional Multi head Self-attention (DWSC-MSA) methods are suggested to be included into Swin transformer blocks. This entails changing the Shifted Window Multi-Head SelfAttention (SW-MSA) in the decoder and encoder blocks respectively. The DWSCMSA method enables the extraction and prioritizing of specific features, resulting in enhanced classification performance. We performed experiments on the Eurosat dataset using many additional commonly used transformers, including swin-tiny, swin-small, swin-base, crossvit, and convit. The experimental results showcase the efficacy of our suggested framework in capturing spatial relationships and improving feature representation, thus pushing the boundaries of land use and land cover classification.