On-Chip Hyperspectral Image Classification Using Vlsi-Based Transformer Networks

Hyperspectral image (HSI) classification plays a vital role in remote sensing, agriculture monitoring, and environmental surveillance; however, its high spectral dimensionality and computational complexity pose significant challenges for real-time and on-chip implementation. This paper presents an on-chip hyperspectral image classification framework using a VLSI-based Transformer network, specifically optimized for energy-efficient embedded and edge devices. The proposed architecture integrates spectral–spatial tokenization, lightweight self-attention mechanisms, and reduced-precision arithmetic to efficiently model long-range spectral dependencies while minimizing hardware overhead. The Transformer model is co-designed with the underlying VLSI architecture using fixed-point quantization (8–12 bits), attention pruning, and on-chip memory reuse to reduce data movement and power consumption. Experimental evaluation on benchmark hyperspectral datasets demonstrates that the proposed method achieves a classification accuracy of 98.2% on Indian Pines, 99.1% on Pavia University, and 97.6% on Salinas, outperforming CNN- and RNN-based on-chip baselines by 2.8–4.5% in overall accuracy. Hardware synthesis results show that the proposed VLSI accelerator operates at 200 MHz, achieves a throughput of 1.9 GOPS, and consumes only 72 mW of power, resulting in an energy efficiency of 26.4 GOPS/W, which is a 38% improvement over existing HSI hardware accelerators. These results confirm that the proposed VLSI-based Transformer framework enables high-accuracy hyperspectral image classification with low latency and ultra-low power consumption, making it highly suitable for real-time on-chip remote sensing and edge-AI applications.