Shapley Values Guide Token Pruning for Real-Time Remote Sensing AI

Vision Transformers (ViTs) have demonstrated exceptional capabilities in remote sensing scene classification; however, the inherent quadratic cost associated with the self-attention mechanism presents a significant barrier to achieving real-time inference when deployed on resource-constrained edge platforms, such as those found on-board satellites and Unmanned Aerial Vehicles (UAVs). While token pruning offers a pathway to accelerate ViT models, existing mainstream pruning methodologies typically rely on attention scores to rank tokens. The reliability of these attention scores is often contested, meaning aggressive pruning risks discarding patches that are critically important for accurate classification. To address this fundamental limitation, researchers introduced a novel approach utilizing the axiomatically grounded Shapley contribution as the definitive pruning criterion.

The core innovation lies in decoupling the computationally expensive Shapley value calculation from the real-time inference path. To prevent the exponential cost of Shapley computation from hindering deployment, the Shapley values are estimated offline and used as supervision. Subsequently, a lightweight head, comprising approximately 0.10 million parameters, is trained through a process called ranking distillation. This head is specifically trained to reproduce the ranking provided by the offline Shapley values. Consequently, the deployed model executes no Shapley computation during inference, ensuring that the pruning decision remains highly informed without incurring prohibitive runtime overhead.

The pruning process itself is implemented hierarchically across specific layers (layers 4, 7, and 10) and is complemented by a calibrated early-exit mechanism that is recalibrated after the pruning procedure is complete. When tested on the NWPU-RESISC45 dataset using a DeiT-S backbone, the criterion learned by the lightweight head demonstrated superior fidelity to the offline Shapley ranking compared to standard attention-based methods. Specifically, the learned criterion achieved a Spearman rho of 0.641, significantly outperforming the 0.401 observed with attention scores, alongside higher Insertion AUC and lower Deletion AUC metrics.

In terms of practical deployment, the model was tested on a Jetson Orin Nano equipped with TensorRT. The implementation achieved substantial efficiency gains, resulting in a reduction of 24.18% in latency and 32.27% in per-image energy consumption when compared to the unpruned baseline model. Regarding classification accuracy, the model achieved 93.04% accuracy at a 0.7 token retention rate, which represents a 0.53 percentage point improvement over pruning guided by attention scores. However, the study notes that this accuracy comparison is marginal, given that the comparison was based on only three paired seeds, leading to a non-significant result in a paired t-test (p = 0.067). Furthermore, an observed-effect diagnostic power calculation yielded only 0.51 at an alpha of 0.05.

Given these findings, the authors conclude that the primary contribution of this work is not framed as a general accuracy advantage, but rather as the introduction of a more faithful pruning criterion and the demonstration of a favorable accuracy-latency trade-off within a single-backbone, single-remote-sensing-benchmark setting.

Subject of Research: Developing a robust token pruning method using Shapley values to enable real-time scene classification on edge devices.

Article Title: Shapley-guided token pruning for real-time remote sensing scene classification on edge devices

Article References: Li, C., & Zhou, T. (2026). Shapley-guided token pruning for real-time remote sensing scene classification on edge devices. Scientific Reports. https://doi.org/10.1038/s41598-026-76578-7

Image Credits: AI Generated

DOI: 10.1038/s41598-026-76578-7

Keywords: Vision Transformers, Token Pruning, Remote Sensing, Edge Computing, Shapley Values, AI Efficiency, Deep Learning, Computer Vision, Shapley-guided, token, pruning, real-time