Enhancing Land Use and Land Cover Classification in Satellite Imagery Using Vision Transformers: A Comparative Analysis with Convolutional Neural Networks

The accurate classification of Land Use and Land Cover (LULC) is imperative for spatial applications such as resource management, environmental monitoring, and satellite image analysis. Methods such as Convolutional Neural Networks (CNNs) can identify local patterns. Local patterns often do not identify or make full use of the global contextual information. This study uses Vision Transformers (ViTs), which interplay a self-attention mechanism, promoting better-perceived integration of local and global contexts. The task is to exploit both local and global features for efficient classification properties. Results show ViTs are bounded with many better values than CNN in classification accuracy, robustness, and the generalization aspect, several datasets indicate their power to appear as great tools in remote sensing/satellite image classification problems.