Publication

Abstract : Hyperspectral imaging (HSI) is a powerful non-destructive technique for biochemical analysis of plants, as it can capture detailed spectral information across hundreds of continuous wavelength bands. This study mainly focuses on classifying plant parts of M. oleifera and analyzing their biochemical properties by predicting Vegetative Indices (VIs) directly from HSI. While VIs can be directly computed using combinations of a few bands, these traditional formulations often fail to exploit multi-band structural dependencies across the full spectrum. Hence, a novel framework is developed by integrating HSI with an Improved Hippopotamus Optimization Algorithm (IHOA) for band selection and Graph Neural Networks (GNNs) for classification and regression. We employed 3 GNNs: standard Graph Convolutional Network (GCN), standard Graph Attention Network (GAT), and a novel hybrid GCN-GAT architecture designed specifically for this task. The core objective of this work is to demonstrate that a drastically reduced subset of spectral bands that is selected by the IHOA, retains the comprehensive spectral information required to accurately reconstruct the VIs. Hyperspectral data of Moringa oleifera plant were collected in real time using a FigSpec-23 camera from flowers, leaves, and pods at tender, harvest, and over-matured physiological stages. First, the selected IHOA bands were used for plant-part classification using the GNN models. Subsequently, the optimized band subset that achieves highest classification accuracy was used to predict VIs related to the plant's biochemical properties. Experimental results show that the hybrid GCN-GAT model achieved strong classification using the full dataset of 1200 spectral bands, while the GAT model provided the best trade-off between computational cost and performance, achieving up to 96.82% accuracy utilizing just 25 optimized bands. This resulted in a 97.9% reduction in spectral dimensionality without affecting the model performance. Overall, the key contribution of this work is that the integrated HSI-IHOA-GNN framework successfully predicted multiple VIs with high accuracy, using only the optimized spectral bands, thereby reducing the computational cost.