Development of a high-throughput plant disease symptom severity assessment tool using machine learning image analysis and integrated geolocation

Tomato spotted wilt virus (TSWV) has the potential to cause severe yield losses in peanut (Arachis hypogaea L.), an important annual legume grown around the world. The most effective approach to manage the disease caused by TSWV is to grow disease resistant peanut varieties. One of the key challenges to breeding for disease resistance is to develop an accurate, reproducible and efficient disease assessment method. Accurate field-based assessment of disease incidence and severity is technically challenging and time-consuming. To address this challenge, a field-based, high-throughput assessment tool was developed to quantify the spatial distribution of disease symptoms over experimental peanut plots using a Real Time Kinematic Global Positioning System (RTK-GPS), consumer-grade cameras, a microcontroller, and an open-source machine learning software. A field experiment was designed to establish a range of disease incidence and severity scenarios. This field experiment was imaged for two seasons to develop and validate the tool. Using transfer learning, an existing Convolutional Neural Network (CNN) was trained from supervised training imagery to classify and quantify areas within the plot-level imagery as, symptomatic, asymptomatic, or ground. Multiple images were assessed by the machine learning model and georeferenced to individual experimental plots using RTK-GPS data. The CNN model trained to detect the symptom, ?stunting and mottling?, was evaluated using Receiver Operating Characteristic (ROC) curve analysis and yielded an Area Under the Curve (AUC) of 0.97, sensitivity of 0.77, and specificity of 0.98 on the test set. Results from the disease assessment tool were compared with results from visual disease assessments, conducted by a trained plant pathologist. Field plot level means from CNN-based assessment of stunting and mottling correlated with plot level means from visual assessment of severity (r = 0.78; P < 0.0001). To further validate the CNN-based method, the TSWV field experiment was analyzed using linear mixed models with both visual severity and CNN-based severity assessments used as responses. Both models (visual or CNN-based assessment) identified the same main effects as being significant and post hoc analysis resulted in the same separation of varieties for their severity of TSWV symptoms. The results of this study demonstrate the successful application of this tool for high-throughput disease severity assessment in peanut under field conditions.