Efficient Land-Cover Image Classification via Mixed Bit-Precision Quantization

Land cover (LC) image classification is essential for monitoring environmental changes, urban planning, and disaster management. Deep neural networks (DNNs) have achieved remarkable success in LC classification; however, their deployment on edge devices is constrained by high computational and storage requirements. Quantizing neural networks to reduce model size has proven effective in achieving low bit-width representations of parameters while maintaining the original network's performance. To facilitate deployment on edge devices, we propose a novel adaptive quantization technique that optimally reduces model size while preserving accuracy. This method evaluates layer importance through statistical measures, enabling the adaptive selection of bit-width precision for each layer. Experimental results show that the proposed quantization strategy effectively balances compression and accuracy for different DNN architectures like VGG19, ResNet18, and ResNet50, providing a practical solution for LC classification on EuroSAT dataset in resource-constrained environments.