Vision Transformer-Based Soil NPK Classification Using Infrared Heatmap Analysis and Optimization Techniques

Abstract

Accurate soil nutrient analysis is essential for optimizing fertilizer application and improving agricultural productivity. Traditional machine learning (ML) and deep learning (CNN-based) approaches have been widely used for soil classification; however, they face limitations in capturing long-range dependencies and complex feature representations. This study proposes a Vision Transformer (ViT)-based model for NPK classification from infrared heatmap images. The ViT architecture leverages self-attention mechanisms to enhance spatial feature extraction, improving classification accuracy. The experimental evaluation demonstrates that the proposed ViT model achieves a 94.2% classification accuracy, outperforming standard CNN architectures such as VGG19, ResNet-50, Inception-V3, MobileNet-V3, and EfficientNet-B2. The confusion matrix analysis highlights the model's robustness in distinguishing varying soil nutrient compositions, even under different moisture levels and fertilizer concentrations. The results validate the effectiveness of attention-driven feature extraction and optimization techniques in soil nutrient classification. This research establishes a strong foundation for precision agriculture, enabling real-time NPK monitoring and adaptive fertilizer management.