Enhancing Satellite Image Classification in Remote Sensing Applications Utilizing Deep Convolutional Neural Networks

Satellite image classification is a key task in remote sensing, where deep learning models are increasingly applied for their accuracy and automation capabilities. This study conducts a comparative analysis of multiple convolutional neural network (CNN) architectures, including a Simple CNN, ResNet50, and EfficientNet variants (B0 to B7), using two benchmark datasets: RSI-CB256 and NWPU_RESIS_small. While the Simple CNN achieved the highest classification accuracy on both datasets (88.30 % and 88.81 %, respectively), EfficientNet models, particularly B2 to B4, demonstrated superior generalization with higher precision, recall, and F1-scores. The evaluation also reveals that deeper EfficientNet variants (B5-B7) do not guarantee improved performance, highlighting the importance of matching model complexity with dataset characteristics. These findings emphasize that while simpler models may excel in raw accuracy, EfficientNet variants offer a more balanced trade-off between accuracy and robustness, making them well-suited for real-world remote sensing applications.