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. 

Called Dynamic Clustering for Channel Pruning (DCCP), the method 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 are increasingly used to interpret remote sensing images; however, their computational demands can limit their deployment on devices with limited resources. This is a particularly significant challenge for satellite-borne and embedded platforms, where computing power, memory capacity, and energy consumption are tightly constrained. Structured network pruning is a practical way to reduce computational cost because it removes entire channels or filters while remaining compatible with general-purpose hardware.

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

The researchers also designed a Dynamic Clustering Agent based on hypergraph convolution. Unlike conventional graph models, which mainly describe pairwise relationships, the hypergraph-based framework models higher-order similarities among multiple channels. During pruning, the agent updated channel clusters dynamically, 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 balancing model accuracy and computational efficiency.

The researchers evaluated the DCCP method using ResNet-50, MobileNet-V2, and VGG-16 on the NWPU-RESISC45 and NaSC-TG2 remote sensing scene classification datasets. On NaSC-TG2, DCCP reduced ResNet-50's FLOPs by 68.52% and its parameter count by 63.39% while increasing its Top-1 accuracy from 96.40% to 97.33%. On the NWPU-RESISC45 dataset, DCCP reduced ResNet-50's FLOPs by 64.03%, while only decreasing Top-1 accuracy by 0.27 percentage points.

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 decreasing GPU memory consumption 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 demonstrate that DCCP can improve the efficiency of deep neural networks without introducing additional runtime modules after pruning. This approach could support more efficient remote sensing intelligence on satellites and edge devices. The researchers plan to extend the method further to include remote sensing object detection for real-time, satellite-borne applications under strict resource constraints.