Model Compression Meets Resolution Scaling for Efficient Remote Sensing Classification

Efficient land cover (LC) image classification is essen-tial for resource-constrained applications like environmen-tal monitoring and urban planning. While deep neural networks (DNNs) excel in LC classification, their deployment on edge devices remains challenging due to computational and memory limitations. We propose an adaptive pruning framework that compresses pre-trained networks while maintaining accuracy. Our approach introduces a Layer Pruning Coefficient to rank layers for pruning based on parameter ratio and standard deviation ratio. An adaptive bi-nary search algorithm dynamically determines the optimal pruning threshold, minimizing accuracy loss. Additionally, we integrate 8-bit quantization and Huffman encoding for further compression. Experimental results on the EuroSAT dataset demonstrate that our method balances compression and accuracy across DNNs such as VGG16, ResNet18, ResNet50, and GoogleNet. For the VGG16 model at 64 × 64 resolution, we achieve compression rates of 41× (pruning), 169 × (with quantization), and 687× (with Huffman encoding), with minimal performance loss. Our approach scales well with resolution and outperforms traditional pruning strategies, offering an effective solution for real-world de-ployment on edge devices.