Deep learning for sustainable agriculture: automating rice and paddy ripeness classification for enhanced food security

Abstract

Accurate and timely classification of rice paddy ripeness is critical for optimizing harvest decisions, improving grain quality, and strengthening global food security. Traditional manual assessments remain subjective, labor-intensive, and poorly scalable, underscoring the need for automated solutions. This study presents a rigorous comparative evaluation of five fine-tuned deep learning architectures for real-time rice maturity assessment: YOLOv11 enhanced with an Attention-Guided Multi-Scale Feature Fusion (AGMS-FF) module, baseline YOLOv11, ResNet18, EfficientNet-B0, and MobileNetV3. Two publicly available datasets were utilized: one augmented to simulate diverse field conditions and another comprising raw, uncontrolled imagery to assess real-world generalizability. To ensure robustness and mitigate overfitting, we employed 5-fold cross-validation alongside a held-out test evaluation. Models were assessed across Accuracy, Precision, Recall, F1-score, ROC-AUC, and PR-AUC metrics. The AGMS-FF YOLOv11 achieved superior performance, with up to 99.6 % cross-validation accuracy (±0.21), ROC-AUC = 0.9877 and PR-AUC = 0.9526 on the augmented dataset, and 98.0 % test accuracy with perfect ROC-AUC and PR-AUC (1.000) on the raw dataset. Statistical validation confirmed the significance of these results through ANOVA (Dataset 1: F(4,20) = 158.4, p < 0.001; Dataset 2: F(4,20) = 92.7, p < 0.001) and McNemar’s paired tests (p < 0.05). These findings provide robust comparative benchmarks across lightweight and state-of-the-art models, reinforcing the viability of deep learning-based computer vision systems for sustainable rice farming and their potential for scalable field deployment.