AI Based land use and land cover classification using RESNET-18 with CBAM

Accurate Land Use and Land Cover (LULC) classification using satellite remote sensing is vital for urban planning, environmental monitoring, and sustainable resource management. While standard Convolutional Neural Networks (CNNs) have automated Earth observation workflows, conventional backbones treat spatial pixels and spectral channels uniformly, causing feature redundancy and boundary confusion in spectrally similar categories. In this paper, we present an attention-guided deep learning framework combining a compact 18-layer Residual Network (ResNet-18) with the Convolutional Block Attention Module (CBAM). By sequentially applying channel and spatial attention mechanisms after each residual block group, the architecture adaptively highlights salient semantic regions while suppressing background noise. The proposed model is benchmarked on the 27,000-patch EuroSAT (RGB) dataset across 10 distinct LULC classes using a 60:20:20 stratified split. The framework achieves a peak validation accuracy of 97.37%, an overall test accuracy of 96.28%, a Macro F1-score of 0.9619, and a micro-average ROC-AUC of 0.9995 with only ~11.20 million parameters. The proposed lightweight model outperforms traditional machine learning models and substantially deeper vision backbones while maintaining computational efficiency.