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

Abstract Vision Transformers (ViTs) excel at remote sensing scene classification, but the quadratic cost of self-attention hinders real-time inference on edge platforms such as on-board satellite and UAV payloads. Token pruning accelerates ViTs, yet mainstream methods rank tokens by attention scores, whose reliability is contested, so aggressive pruning may discard predictively important patches. We instead adopt the axiomatically grounded Shapley contribution as the pruning criterion. To keep its exponential cost out of inference, Shapley values are estimated offline as supervision, and a lightweight head (about 0.10 M parameters) is trained through ranking distillation to reproduce their ranking, so the deployed path performs no Shapley computation; pruning is hierarchical (layers 4/7/10) with a calibrated early-exit recalibrated after pruning. On NWPU-RESISC45 (DeiT-S backbone), the criterion learned by this head follows the offline Shapley ranking more faithfully than attention does (Spearman rho = 0.641 vs. 0.401), with a higher Insertion AUC and a lower Deletion AUC. On a Jetson Orin Nano with TensorRT, latency and per-image energy fall by 24.18% and 32.27% versus the unpruned model. Accuracy at 0.7 retention is 93.04%, 0.53% points above attention pruning; however, with only three paired seeds this comparison is marginal (paired t-test, p = 0.067), and an observed-effect diagnostic power calculation gives only 0.51 at alpha = 0.05. Accordingly, no statistically decisive accuracy improvement is claimed. The contribution is therefore framed as a more faithful pruning criterion and a favourable accuracy-latency trade-off under a single-backbone, single-remote-sensing-benchmark setting, rather than as a general accuracy advantage.