In a study published in Knowledge-Based Systems, researchers from the Changchun Institute of Optics, Fine Mechanics and Physics (CIOMP) of the Chinese Academy of Sciences developed a deep reinforcement learning-based pruning method that made convolutional neural networks more compact and faster while maintaining competitive recognition accuracy. The method, called Dynamic Clustering for Channel Pruning (DCCP), reduced the floating-point operations of ResNet-50 by 68.52% on a remote sensing dataset and achieved practical acceleration on both a GPU and a mobile embedded platform.

Deep neural networks have become increasingly important in remote sensing image interpretation, but their growing computational complexity has limited deployment on resource-constrained devices. This challenge is particularly significant for satellite-borne and embedded platforms, where computing power, memory capacity, and energy consumption are tightly constrained. Structured network pruning offers a practical way to reduce computational cost because it removes complete channels or filters while remaining compatible with general-purpose hardware.

However, many existing deep reinforcement learning-based pruning methods mainly determined how much each network layer should be compressed, while paying limited attention to the different functions of individual channels. To address this problem, the researchers used low- and mid-frequency response statistics of feature maps to describe channel-level network states. These features reflected both broad semantic information and texture-related characteristics, providing the pruning agent with more detailed information for decision-making.

The team further designed a Dynamic Clustering Agent based on hypergraph convolution. Unlike conventional graph models that mainly describe pairwise relationships, the hypergraph-based framework modeled higher-order similarities among multiple channels. During pruning, the agent dynamically updated channel clusters, estimated the pruning ratio for each layer, and ranked redundant channels according to their functional similarity. Deep reinforcement learning then guided the pruning strategy toward a balance between model accuracy and computational efficiency.

The researchers evaluated DCCP on the NWPU-RESISC45 and NaSC-TG2 remote sensing scene classification datasets using ResNet-50, MobileNet-V2, and VGG-16. On NaSC-TG2, DCCP reduced the FLOPs of ResNet-50 by 68.52% and the parameter count by 63.39%, while increasing Top-1 accuracy from 96.40% to 97.33%. On NWPU-RESISC45, it reduced ResNet-50 FLOPs by 64.03% with only a 0.27-percentage-point decrease in Top-1 accuracy. 

The method also showed clear deployment benefits. On an NVIDIA GTX 3090 GPU, DCCP reduced per-image inference latency from 0.348 ms to 0.246 ms, corresponding to about 1.41× acceleration, while GPU memory consumption decreased from 126.06 MB to 68.07 MB. On the Fudan Micro JFMQL100TAI embedded platform, pruning reduced total inference latency by 1.36 ms and increased the average frame rate by 34 fps.

These results demonstrated that DCCP could improve the efficiency of deep neural networks without introducing additional runtime modules after pruning. The approach may support more efficient remote sensing intelligence on satellites and edge devices. The researchers planned to further extend the method to remote sensing object detection for real-time satellite-borne applications under strict resource constraints.