Improving Diagnostic Transparency in Tomato Crop Health Using Gradient-Weighted Class Activation Mapping (Grad-CAM) and Transfer Learning with Convolutional Neural Networks
Year: 2026
Type: article
Abstract: Timely and precise diagnosis of tomato leaf diseases is crucial for safeguarding crop health, maximizing agricultural yield, and minimizing environmental impacts. Convolutional Neural Networks (CNNs) have emerged as exceptional tools for automated disease monitoring and precision agriculture. This paper presents a comparative analysis of five transfer-learning CNN architectures—MobileNetV3, MobileNetV4, InceptionV3, EfficientNetB1, and ResNet50—for the multiclass categorization of tomato leaf conditions. The dataset encompasses a healthy control alongside nine pathological classes: mosaic virus, target spot, bacterial spot, yellow leaf curl virus, late blight, leaf mold, early blight, spider mites, and septoria leaf spot. Although the more complex EfficientNetB1 yielded the highest classification metrics, statistical validation via Analysis of Variance (ANOVA) and Tukey’s Honestly Significant Difference (HSD) test revealed that the lightweight MobileNetV3 achieved statistically comparable, near-optimal efficacy, with a testing F1-score of 98.76%. Moreover, MobileNetV3 demonstrated superior computational efficiency, utilizing only 4.2 million parameters, completing training epochs in 28.36 seconds, and enabling rapid inference at 1.34 milliseconds per image. Gradient-weighted Class Activation Mapping (Grad-CAM) visually validated that the network accurately isolated relevant pathological leaf lesions, providing clear diagnostic explanations and enhancing interpretability. Leveraging its optimal balance of robust accuracy and computational efficiency, MobileNetV3 was successfully deployed into a responsive web application for real-time crop health monitoring. Ultimately, this framework offers a scalable, efficient solution for optimizing agrochemical usage, mitigating crop loss, and promoting sustainable agriculture.
Source: Engineered Science
DOI: 10.30919/es2505
Institutions: Prince of Songkla University, Walailak University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: diamond
OpenAlex ID: W7204649576