ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

LeafVision: On-Device Powdery Mildew Detection in Watermelon Using a Reconfigurable Edge Platform

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600178.(doi:10.13031/aim.202600178)
Authors:   Shekhar Suman Borah, Rahul Kumar, Prajwal Mahesha Sethurani, Samvada Jain Manel, Chandrasekar Kousik, Prabha Sundaravadivel
Keywords:   Watermelon powdery mildew, edge AI, greenhouse disease detection, deep learning, transfer learning, thermal imaging, embedded systems, precision agriculture, plant disease monitoring, TensorFlow Lite, Android application.

Abstract. Early detection of powdery mildew is essential for reducing yield loss and improving disease management in watermelon greenhouse production. This paper presents LeafVision, a reconfigurable edge platform for on-device watermelon powdery mildew detection using RGB imaging, thermal sensing, and embedded artificial intelligence. A baseline classification study was first conducted using 95 labeled watermelon seedling images, including 42 healthy and 53 infected samples. Three transfer-learning models, MobileNetV2, EfficientNetB0, and FastAI ResNet34, were evaluated using a hold-out split and 5-fold cross-validation. All models achieved 100% accuracy on the 19-image hold-out test set. Cross-validation results showed that FastAI ResNet34 achieved the highest average accuracy at 96.8% ± 2.6%, followed by EfficientNetB0 at 93.7% ± 5.2% and MobileNetV2 at 91.6% ± 9.8%. Although ResNet34 achieved the best classification performance, EfficientNetB0 was selected for embedded deployment because of its smaller size and reliable TensorFlow Lite conversion. The exported FP16 TensorFlow Lite model (8.81 MB) was deployed on an NVIDIA Jetson Nano with an ELP 12 MP RGB camera and a FLIR Lepton 3.5 thermal camera for real-time on-device inference. Thermal anomaly detection was included as a supplementary decision-support feature, and a lightweight Android application, Plant Doctor, was developed for offline disease classification using TensorFlow Lite and HSV-based fallback segmentation. The results demonstrate the feasibility of lightweight edge AI for portable plant disease monitoring in precision agriculture.

(Download PDF)    (Export to EndNotes)