Quantifying Visual Symptom Severity in Plants Using Deep Learning: A Severity Scale Derived from Classification Model Outputs

Plant pests and diseases pose a major global threat to food security and agricultural sustainability, making accurate assessment of plant symptoms important. This study proposes a simple and practical method for quantifying visual symptom severity using binary deep learning classifiers trained only on asymptomatic and severely affected images. Visual symptom severity was estimated from class probabilities, and classification accuracy is additionally used when image groups with similar symptom severity are available. While the framework enables symptom severity quantification at the group level, direct application to individual images is challenging due to overconfident predictions for mildly symptomatic cases. To address this issue, temperature scaling and label smoothing were evaluated, and label smoothing was found to improve reliability for individual image assessment. The method was validated using mosaic and wilting symptoms across multiple architectures. Results showed close agreement in classification accuracy and class probabilities across architectures, with maximum differences of 4.8% (mosaic) and 9.7% (wilting). The estimated visual symptom severity was generally consistent with expert visual assessment under the conditions examined in this study. The proposed method requires only minimal annotation and uses standard model outputs, making it simple, interpretable, and potentially applicable to other visually assessed traits.