Comparative Analysis of U-Net Models Using ResNet34, InceptionV3, and VGG16 for the Processing of Satellite Images

Satellite image classification is a crucial component in remote sensing applications, facilitating the automated analysis of land cover and land use patterns. This research explores the effectiveness of the U-Net architecture and state-of-the-art deep learning techniques for satellite image classification. Utilizing diverse and well-annotated satellite image datasets, we demonstrate the capability of U-Net in capturing intricate spatial features within images, making it a powerful tool for discriminating land cover classes. Our experiments involved training the U-Net model with the ResNet34, InceptionV3, and VGG16 architectures. The U-Net-based ResNet34 model achieved the highest accuracy of 81.0% and a validation F1-score of 65.95% after 50 epochs. The findings underscore the potential of U-Net and deep learning techniques to advance the field of remote sensing, providing solutions for real-world challenges such as environmental monitoring, urban planning, disaster management, and precision agriculture.