Graph Neural Networks with Autoregressive Moving Average Graph Filter and Graph- Regularized Sparse Coding for Accurate Hyperspectral Image Classification on FPGA

Classification techniques based on hyperspectral images (HSIs) have gained significant importance in target identification, mineral mapping, and environmental management due to rapid advancements in hyperspectral remote sensing technology. Graph Neural Networks (GNNs) have become a key technique, attracting considerable attention in HSI analysis. However, GNN-based techniques frequently depend on a graph filter to extract the intricate spectral-spatial characteristics inherent in HSI data, which limits the ability to fully exploit the diverse and rich information available. This can lead to less comprehensive feature representation. To overcome these limitations, we propose a novel method called AMAGC-GNN, Autoregressive Moving Average Graph Filter and Graph-Regularized Sparse Coding for the GNN. The ARMA graph filters play a crucial role in spectral filtering, effectively capturing and preserving complex spectral relationships while mitigating the common oversmoothing problem in traditional GNNs, ensuring node features retain their uniqueness and discriminative power. Concurrently, graph-regularized sparse coding transforms the input HSI data into sparse coefficients, representing the most relevant features efficiently. This combined approach leverages the strengths of both techniques, providing adaptive and precise spectral filtering and compact, discriminative feature representation. AMAGC-GNN outperforms other techniques, achieving significant improvements in Cohen's Kappa coefficient (6.58% and 6.15%), per-class accuracy (6.24% and 5.57%), and overall accuracy (6.24% and 6.30%) across Kennedy Space Centre and Indian pines datasets. Furthermore, our implementation of AMAGF-GNN on Virtex-7 field-programmable gate arrays (FPGAs) demonstrates promising results for real-world applications in HSI classification, particularly in achieving highly accurate target localization.