Deep Learning for Satellite Image Classification: A Comparative Analysis of CNN and ResNet-18

With the ease of classifying land through satellite imaging, remote sensing has captured the Earth observation domain. Traditional methods for analyzing satellite images relied on manual feature extraction and statistical models which was time-consuming and often failed to capture sophisticated patterns. Moreover, advancements in artificial intelligence (AI) and deep learning, particularly Convolutional Neural Networks (CNN), have significantly improved classification accuracy. This study investigates the effectiveness of CNN-based architectures on the task of categorizing satellite images, by evaluating the performance of a custom CNN model against a pre-trained ResNet-18 model. The comparative analysis showed that ResNet-18 achieved better classification accuracy, computational efficiency and generalization ability than the custom CNN by achieving a 89.3% test accuracy with the reduced training time. These results indicate the power of transfer learning for the classification of satellite images and provide an efficient and practical means for the analysis of remote sensing data. According to the study, these deep learning classification algorithms isolate an effective and automated framework for handling large swathes of satellite imagery making it significant for environmental monitoring, land use planning, and disaster management.