Nano Transformer-HSI: A VLSI-Accelerated Attention Framework for Real-Time Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification has emerged as a crucial task in remote sensing, environmental monitoring, precision agriculture, and defense applications. Conventional convolutional neural network approaches suffer from limited capability in capturing long-range spectral-spatial dependencies and require substantial computational resources. Transformer-based architectures have demonstrated superior feature representation capability; however, their high complexity restricts deployment in edge and embedded systems. This paper proposes NanoTransformer-HSI, a VLSI-accelerated attention framework for efficient hyperspectral image classification. The proposed architecture employs spectral-spatial tokenization followed by lightweight multi-head self-attention and hierarchical feature fusion to capture discriminative information across spectral bands. To enable real-time operation, the model is implemented using a low-power VLSI hardware accelerator with parallel processing elements and optimized memory access mechanisms. Experiments conducted on benchmark datasets demonstrate classification accuracies of 99.12%, 98.76%, and 99.05% on the Indian Pines, Pavia University, and Salinas datasets, respectively. Furthermore, the proposed architecture achieves approximately 31% lower power consumption and 42% reduction in latency compared with conventional transformer implementations. The results indicate that the proposed framework provides an efficient and scalable solution for intelligent edge-based hyperspectral image analysis.