Multiscale feature fusion based CNN with feature ranking for satellite image classification

Satellite image classification plays an important role in various applications like monitoring, surveillance, land cover planning, etc. Deep learning models with many learnable parameters face overfitting as model depth increases and have struggled to extract both fine details and broader patterns simultaneously. To overcome the aforementioned issues, an efficient CNN model of <0.23 million parameters is proposed, utilising the extraction of multiscale features across different CNN branches. After concatenating the two branches, a global average pooling (GAP) layer is added to mitigate overfitting without introducing additional parameters. To overcome the limitations of standard CNNs in handling redundant and less discriminative features, chi-square test feature ranking is incorporated to emphasise the most informative representations, while the subsequent SVM classifier exhibits superior performance in separating complex feature patterns. Then, the model shows accuracies of 99.82% and 95.33% on RSI-CB256 and EuroSAT, respectively, making it a suitable candidate for satellite classification.