Decoupling-Based Cross-Layer Connection Removal for Compact Remote Sensing Image Classification

Cross-layer connections can couple different layers of neural networks in diverse paths, improving the classification accuracy of remote sensing images. However, some structures are unnecessary for inference and not conducive to be deployed on starboard/airborne platforms. Roughly removing them lengthens the backpropagation path and causes the gradient vanishing problem, degrading the performance of existing end-to-end training-based lightweighting methods. To address this issue, we propose a decoupling-based cross-layer connection removal (DCCR) framework, which decomposes the complete deep network into multiple shallower stages based on the density of connections. By optimizing each stage independently, the backpropagation path is shortened and the gradient vanishing problem is expected to be alleviated. Moreover, to approximate the original network in each stage, we perform concrete reconstruction error suppression on branches, layers, and channels. Combining these strategies, redundant structures are removed while maintaining performance. Extensive experiments on multiple public datasets and hardware platforms demonstrate that this compact architecture achieves substantially improved deployability and inference efficiency at the cost of only marginal accuracy degradation.