Enhancing Remote Sensing Image Classification and Interpretability: A Multi-Stage Feature Extraction Approach and Grad-CAM

Abstract

Relevant and current land use data remain critical for effective spatial layout, environmental stewardship, and bestow resources. This study assesses the classification accuracy of machine learning classifiers developed for land use/ classification (LUC) and how effective are they when combined with feature engineering and augmentation strategies on a benchmark dataset. The processing pipeline uses a Convolutional Neural Network (CNN) model called ResNet-50 for image feature extraction, which captures spatial patterns and also performs well in classification tasks. Model performance is enhanced by using a combination of the pre-trained ResNet-50 as an input for self-compiled image classifiers and applying Principal Component Analysis (PCA) for dimensionality reduction. Advanced-Data augmentation(ADA) techniques further improve the generalization of the models. Of all tested classifiers (Logistic Regression, Random Forest, SVM, Gradient Boosting and XGBoost), SVM outperformed the rest with AUC-ROC at 0.993 and MCC at 0.843. In addition, Grad-CAM (Gradient-weighted Class Activation Mapping) visualizations applied to enable an understanding of the decision processes of the model, thus increasing its interpretability. This work demonstrates the combining of deep feature extraction with classical machine learning and makes land use classification scalable and robust.