An Effective DNN Based ResNet Approach for Satellite Image Classification

Abstract

This research study presents a comparative examination of convolutional neural networks (CNNs) and residual networks (ResNets) applied to the classification of satellite images. The study employs a dataset containing diverse land cover types and trains both CNNs and ResNets on this dataset. The performance evaluation of the models is carried out on an independent test set of satellite images. The experimental outcomes reveal that ResNets outperform CNNs in terms of accuracy on the test set. This superiority can be attributed to the ResNets' ability to effectively capture long-range relationships within images. While CNNs excel in capturing local relationships, their capability to learn long-range dependencies is comparatively limited. Furthermore, the study identifies ResNets as more resilient to overfitting compared to CNNs. This resilience can be attributed to the larger number of parameters in ResNets, which provides additional learning opportunities and helps prevent overfitting. To summarize, the results suggest that ResNets are a more favorable choice for satellite image classification. ResNets demonstrate higher accuracy and greater resistance to overfitting, making them well-suited for this particular task.