Abstract:
Satellite image classification is crucial for various applications, driving advancements in Convolution Neural Networks (CNNs). While CNNs have proven effective, deep models often encounter overfitting issues as the network's depth increases since the model has to learn many parameters. Besides this, traditional CNNs have the inherent difficulty in extracting fine-grained details and broader patterns simultaneously. To overcome these challenges, this research presents a novel approach using an optimized parallel CNN (OPCNet) architecture with an SVM classifier to classify satellite images. Each branch within the parallel network is designed for specific resolution characteristics, spanning from low (emphasizing broader patterns) to high (capturing fine-grained details), enabling the simultaneous extraction of a comprehensive set of features without increasing network depth. The OPCNet incorporates a dilation factor to expand the network's receptive field without increasing parameters, and a dropout layer is introduced to mitigate overfitting. Evaluation of two public datasets (EuroSAT dataset and RSI-CB256 dataset) demonstrates remarkable accuracy rates of 97.91% and 99.8%, surpassing previous state-of-the-art models. Finally, OPCNet, with less than 1 million parameters, outperforms high-parameter models by effectively addressing overfitting issues, showcasing exceptional performance in satellite image classification.