Convolutional Neural Network-Based Accelerator for Hyperspectral Image Classification

Hyperspectral image classification is of paramount importance in precision agriculture and environmental conservation.Current approaches predominantly process hyperspectral images offline, resulting in suboptimal efficiency.To mitigate this issue, this study proposes an FPGA accelerator for hyperspectral image classification to enhance processing speed while preserving classification accuracy.In this paper we develop a lightweight convolutional neural network (CNN) tailored for hyperspectral image classification, achieving an average accuracy of 95.3% on the Indian Pines dataset.Subsequently, we implement the model on an FPGA platform utilizing the Open Computing Language (OpenCL) architecture.Techniques such as layer fusion and fixed-point quantization are employed to reduce model parameters without compromising performance.As a result, the proposed accelerator achieves image classification within 0.06s with negligible accuracy degradation.