Hyperspectral Imaging and Advanced Vision Transformers for Identifying Pure and Pesticide-Coated Apples

The classification of hyperspectral images has become an essential task in agricultural analysis, as it assesses the quality, chemical composition, and overall health of produce. This study concentrates on the analysis of the hyperspectral images of apples to distinguish and classify pure apples and those subjected to fertilizers in different concentration levels. This classification is fundamental to ensuring food quality and optimizing fertilizer input and precision agriculture practices. The dataset did not initially satisfy requirements owing to the small sample size, which rendered it difficult to train any robust generalizable model. A conditional GAN variation (CGAN) has been proposed for data augmentation. This approach effectively handled class imbalance by generating high-quality, category-specific synthetic images that enriched the dataset with realistic and diverse samples. For classification, the ConvNeXt architecture was employed, integrated with the Simple Attention Module (SimAM) to enhance feature refinement and extraction. The SimAM module allowed the model to focus on the most relevant areas of hyperspectral images while suppressing unimportant information, leading to a highly refined representation of features. The proposed classification model achieved an impressive accuracy of 99.75%, showcasing its exceptional ability to distinguish between different categories of apples, including untreated ones and those with varying fertilizer concentrations. This high level of accuracy underscores the model’s robustness and reliability in agricultural image analysis.