Improved Knowledge Distillation for Land-Use Image Classification

An improved Knowledge Distillation (KD) framework has been proposed for compressing deep convolutional neural networks, specifically using a VGG16 teacher model to train a lightweight MobileNetV2 student model for land-use image classification. The method integrates hard supervision with a soft supervision strategy that employs Kullback-Leibler divergence and Cosine Similarity losses, achieving an accuracy of 99.04% on three land-use datasets, surpassing baseline student training and single-loss distillation methods while maintaining significant model compression. This advancement is crucial for practitioners aiming to deploy efficient, high-accuracy models in resource-constrained environments.