Plant Stress Classification on Thermal Images Using MobileNetV3

ABSTRACT Early detection of plant stress is crucial for minimizing crop loss and promoting sustainable food production. Traditional methods often fail to identify stress indicators before visible symptoms emerge. Normal images primarily capture visible signs of stress, which become apparent only after significant damage has occurred, limiting timely intervention and leading to lower accuracy in early detection due to their inability to capture hidden stress markers. In contrast, thermal images reveal temperature variations that indicate stress at an earlier stage, even before it becomes visible to the human eye, allowing for improved accuracy by identifying subtle physiological changes. The proposed system involves collecting and preprocessing thermal images of plants under varying stress conditions, enabling the detection of underlying stress through these temperature variations. A MobileNetV3 model, known for its lightweight architecture, speed, and efficiency, is trained on these thermal images to classify them into stress and non‐stress categories. The experimental results compare three deep learning models—MobileNetV3, DenseNet, and VGG16 for plant stress classification using thermal images. MobileNetV3 achieved the highest accuracy, with an average F1‐score of 0.67, significantly outperforming DenseNet and VGG16. MobileNetV3 strikes an optimal balance between accuracy and computational efficiency, outperforming more complex models while maintaining lower processing demands. This makes it particularly suitable for real‐time, on‐device applications. The proposed system harnesses these advantages to provide farmers and agronomists with an automated, non‐invasive solution for real‐time plant health monitoring.