Chlorophyll attention network using MobileViTv2 for on-device plant health monitoring

Abstract

Accurate assessment of leaf chlorophyll content is critical for precision nitrogen management in maize (Zea mays L.), yet conventional SPAD meters are labor-intensive and limited in scalability. To address this, a regression framework integrating MobileViT, Chlorophyll Attention (CA), and a multi-modal fusion mechanism for rapid, image-based SPAD value prediction was proposed. An ablation study quantified the contribution of each architectural component. The full model achieved a baseline coefficient of determination (R²) of 0.6832, confirming robust predictive capability. Removal of the MobileViT block caused the largest performance decline (R² = 0.6564), emphasizing its critical role in feature representation. Excluding the CA module (R² = 0.6721) or the multi-modal fusion mechanism (R² = 0.6368) further highlighted the necessity of synergistic component integration for optimal performance. Using 5-fold cross-validation, the base model attained a mean R² of 0.7099, which was further enhanced to 0.7345 through ensemble learning and four-way test-time augmentation. The improved model achieved a Root Mean Square Error (RMSE) of 4.264 SPAD units and Mean Absolute Error (MAE) of 3.649 SPAD units. Comparative evaluation demonstrated that the proposed approach outperformed state-of-the-art architectures, including EfficientNet-B3 (R² = 0.7078) and deeper ResNet variants, highlighting its robustness and generalization capability. The model’s rapid inference ensures suitability for real-time, edge-device deployment in field conditions. The study confirms that each component enhances predictive accuracy, and that ensemble learning with TTA further improves reliability, providing a scalable, high-throughput solution for efficient, data-driven chlorophyll assessment.