LeafyResNet: Fusarium wilt detection in lettuce using UAV RGB imaging and advanced deep learning model

Lettuce, one of the most consumed leafy greens globally, offers significant health benefits due to its high vitamin, mineral, and fiber content. However, Fusarium wilt, a soil-borne fungus, threatens lettuce yields by reducing both quality and quantity. Traditional disease detection methods, such as manual inspection, are time-consuming and inefficient. This study proposes an Unmanned Aerial Vehicle (UAV)-based- approach for detecting Fusarium wilt in lettuce using high-resolution Red-Green-Blue (RGB) imagery. (1) a high resolution RGB lettuce dataset captured by drones at approximately 10 m altitude in collaboration with the Yuma Center of Excellence for Desert Agriculture (YCEDA), (2) identification of candidate Fusarium-infected regions by evaluating 300x300 pixel image patches for light tan coloration, followed by the application of a customized Residual Neural Network (ResNet), called LeafyResNet, to confirm Fusarium presence, and (3) a method for quantifying Fusarium infection severity, which was validated against an expert-ground truth. In this study, 6,239 high-resolution RGB images (4000 × 3000 pixels) were collected over nine weeks from YCEDA fields and cropped into 675 ROI patches (100 × 100 pixels) representing healthy and Fusarium-infected areas. Data augmentation expanded the training set to 57,000 patches, while the testing set (20%) remained unaugmented to ensure fair evaluation on unseen data. Our approach to detect Fusarium wilt achieves 96.30% accuracy, 94.10% precision, 100% recall, 99.5% mean Average Precision (mAP) and a 97.10% F1-score, with a 4% false positive rate. Disease severity scores showed an overall accuracy of 86%. LeafyResNet has been specifically adapted with ResNet as a base. Overal, the adapted LeafyResNet performed better than the best baseline, ResNet34, with a 0.3% higher accuracy and 0.8% higher F1-score, showing measurable improvements over ResNet18, ResNet34, Inception_v3, and VGG16, and highlighting the benefits of customizing models for agricultural applications. Core of the adaptation is the starting layers and residual blocks, which tailored LeafyResNet to capture subtle wilting symptoms in leaves as multi-scale features in UAV datasets, providing an efficient and scalable solution for precision agriculture. • LeafyResNet achieved an accuracy of 96.3% in detecting Fusarium wilt in lettuce using UAV data. • Light tan intensity analysis improved plant–soil differentiation in UAV images. • Automated Fusarium wilt severity detection matched expert field assessments. • LeafyResNet outperformed the reference models VGG16 and Inception_v3, which are used for Fusarium detection in other crops.