Exploring Modern Deep Learning Architectures for Satellite Image Analysis: A EuroSAT Perspective

Classifying satellite data to infer usage and land cover properties of the geographical features is very important.In this work, we have analyzed the comparison of 3 deep learning models named VGG16, ResNet50, and EfficientNetB0 using the EuroSAT dataset. It consists of 27000 satellite images with 10 different land classes. We assess the models based on accuracy and computational complexity. The study shows that ResNet50 has a lower mistake rate than VGG16, with a classification accuracy of 92.8% compared to 89.8% for VGG16 and 91.1% for EfficientNetB0. ResNet50’s residual connections enable it to capture a complex feature map, resulting in enhanced performance results. Furthermore, we recommend ResNet50 for use in scenarios with limited resource utilization capabilities because it occupies less computational time and memory. The results of this study show that ResNet50 works better than before, which means it could be useful for classifying images from space and using progressive CNN for some functions that cross disciplines.