SViT-UAV: energy-efficient spiking vision transformer for remote sensing scene classification in UAV scenarios
Accurate scene classification is critical for unmanned aerial vehicle (UAV) perception, yet achieving high accuracy under strict energy constraints remains challenging for conventional artificial neural networks (ANNs). Although spiking neural networks (SNNs) provide an energy-efficient event-driven computation paradigm, existing SNN-based methods for remote sensing scene classification still face limitations in capturing both local spatial details and global contextual information. We present SViT-UAV, a hybrid spiking vision transformer that addresses this limitation by integrating convolutional feature encoding with transformer-based global reasoning under a unified spiking paradigm. A spatio-temporal gating block (STGB) is incorporated to enhance spike representations along spatial and temporal dimensions with negligible parameter overhead. Experiments on four widely used UAV-oriented remote sensing scene benchmarks—WHU-RS19, UC-Merced, AID, and NWPU-RESISC45—show that SViT-UAV achieves ANN-competitive classification accuracy with an energy consumption of only 9.58 mJ, substantially outperforming existing SNN methods. These results highlight the effectiveness of SViT-UAV for accurate and energy-efficient remote sensing scene classification in UAV scenarios.
Authors
- Lili Fan (ORCID: https://orcid.org/0000-0002-9506-3762)
- Keke Zha (ORCID: https://orcid.org/0000-0003-0427-102X)
- Yanan Ding
- Xu Liu
Institutions
- Jinling Institute of Technology (CN)
- Nanjing University of Aeronautics and Astronautics (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1038/s41598-026-80329-z
- Primary Topic
- Remote-Sensing Image Classification
- Type
- article
- Field-Weighted Citation Impact
- 0.00