Assessment of potato late blight from UAV-based multispectral imagery

Abstract

This paper evaluates a method for assessment and detection of potato late blight using UAV-based multispectral imagery. Traditional methods of detection and mapping of late blight are time consuming, require large human effort and, in many cases, are subjective. The approach evaluated integrates morphological operations and evaluates the performance of five Machine Learning (ML) algorithms: Random forest, Gradient Boosting Classifier, Support Vector Classifier, Linear Support Vector Classifier and K-Nearest Neighbours Classifier to detect zones of late blight occurrence. The main components of the proposed approach are: (i) radiometric and geometric correction of raw images; (ii) soil and weed removal by application of a thresholding technique; (iii) a supervised classification procedure using ML algorithms; and (iv) use of trained models to classify a new data set. The performance of the method is evaluated on two dates in an experimental potato field. Results showed that the Linear Support Vector Classifier and Random Forest algorithms had the best performance in terms of both accuracy metrics and run time. The study showed that the proposed method allows the detection of late blight with little human intervention.