RemoteTrimmer: Adaptive Structural Pruning for Remote Sensing Image Classification

TL;DR

This work tackles the efficiency bottleneck in remote sensing image classification by introducing RemoteTrimmer, which combines Channel Attention Pruning (CAP) and Adaptive Mining Loss (AML). CAP amplifies inter-channel importance by mapping features into an attention space using SENet and BatchNorm scaling to guide structured pruning, while AML focuses fine-tuning on difficult samples to recover or exceed the original accuracy. Empirical results on EuroSAT and UC Merced with ResNet18 and VGG16 show state-of-the-art performance post-pruning, with substantial reductions in parameters and MACs and notable accuracy gains over prior methods. The approach offers a practical, principled pathway to deploy lightweight RSIC models without sacrificing performance, and the provided code enables reproducibility and further exploration in remote sensing applications.

Abstract

Since high resolution remote sensing image classification often requires a relatively high computation complexity, lightweight models tend to be practical and efficient. Model pruning is an effective method for model compression. However, existing methods rarely take into account the specificity of remote sensing images, resulting in significant accuracy loss after pruning. To this end, we propose an effective structural pruning approach for remote sensing image classification. Specifically, a pruning strategy that amplifies the differences in channel importance of the model is introduced. Then an adaptive mining loss function is designed for the fine-tuning process of the pruned model. Finally, we conducted experiments on two remote sensing classification datasets. The experimental results demonstrate that our method achieves minimal accuracy loss after compressing remote sensing classification models, achieving state-of-the-art (SoTA) performance.

Figures (3)

• Figure 1: Motivations of our proposed method. (a) Comparison of general (ImageNet) and remote sensing (UCM) images. Remote sensing images have top-down views from different heights, resulting in greater scale variations of the objects. This situation usually leads to the narrowing of differences between similar object features. (b) Remote sensing images often suffer from unclear images with atmospheric noise pollution, which makes lightweight models learning more difficult.
• Figure 2: Overview of our RemoteTrimmer. In the pruning phase, we leverage intermediate outputs from SENet and scaling factors from the BN layer to map channel importance into the attention space. During the post-pruning fine-tuning phase, we design a lateral inhibition loss function to emphasize difficult samples. Our method effectively addresses two key challenges in remote sensing model pruning: the lack of distinct channel importance and the prevalence of difficult samples.
• Figure 3: Ablation studies on different attention mechanisms (a) and pruning rates (b) on EuraSAT dataset. Our approach achieved the best accuracy across all pruning rates.