Performance of pre-trained deep learning models for land use land cover classification using remote sensing imaging datasets

Abstract

This study presents a comparative analysis of ten pre-trained convolutional neural network (CNN) models, evaluated across three remote sensing datasets: EuroSat, NWPU, and Earth Hazards (Land Sliding). We investigate the interplay between model architecture and classifier selection by incorporating five different neural network (NN) classifiers, emphasizing their impact on predictive accuracy and computational efficiency. Due to its densely connected architecture, DenseNet201 achieved the highest accuracy—97% on EuroSat, 99.40% on NWPU, and 97.80% on Earth Hazards. In contrast, MobileNetV2, while slightly less accurate, demonstrated superior computational efficiency, recording the shortest prediction times of 39.943 s on EuroSat, 27.482 s on NWPU, and 2.8986 s on Earth Hazards. Additionally, classifier choice significantly influenced performance, with the Wide NN classifier excelling in diverse datasets and the Medium NN classifier optimizing speed. Our findings underscore the importance of balancing accuracy and efficiency in selecting CNN models for remote sensing applications, suggesting future research should explore ensembling techniques and lightweight models to enhance performance.