Investigation of Land-Use and Land-Cover Classification with synergization of Resnet50, PCA, and Machine Learning Classifiers

Abstract

This research developed a novel strategy for land-use classification that combines feature extraction through deep learning with dimensionality reduction and machine learning (ML) classifiers. The UCMerced LandUse (UCMLU) dataset with 21 classes is used in this work, using pre-trained ResNet50 to extract high-level spatial and contextual features from high-resolution aerial imagery. Then, Principal Component Analysis (PCA) is used to reduce dimensionality, to combat overfitting, and to enhance computation efficiency. Four prominent classifiers, Logistic Regression(LR), Random Forest(RF), Gradient Boosting(GBst), and Support Vector Machine (SVM), are then employed to classify the reduced feature representations. Before training and testing, stratified train-test splits are used to ensure a balanced representation of the classes while training and testing the models. This study evaluated the performance of classifiers with various performance metrics and ROC curves.

The results indicate that the SVM outperforms other models, giving the highest accuracy (81.67%), AUC-ROC (0.9928), thus demonstrating its robustness towards high-dimensional data. LR and RF yield almost equivalent results with strong overall performance, while GBst shows moderate effectiveness. The application of PCA would greatly contribute to an efficiency and generalization by the model. This hybrid approach exemplifies the possibilities of effectively exploiting deep learning features for accurate and efficient LULC classification.