CHLOROPHYLL FLUORESCENCE TO SCORE FINGER MILLET BLAST DISEASE USING MACHINE LEARNING MODEL IN EASTERN AFRICA

Abstract

Finger millet blast disease caused by filamentous fungus Magnaporthe oryzae is the most economically significant constraint to finger millet production worldwide causing up to 90% of yield losses. Visual identification of blast disease is time-consuming and requires highly skilled personnel to understand, analyze the symptoms and score the disease. The limitation of visual scoring has necessitated an urgent need to deploy machine learning techniques that can detect and classify the disease at an early stage. Early diagnosis and accurate identification of blast disease can control the spread of infection ensuring its efficient production, which underpins food security in semi-arid areas of the world. To address this challenge, this study aimed to deploy a machine-learning model based on chlorophyll fluorescence (Chl-F) and blast-infected images to aid in accurate blast detection. One hundred and forty-three finger millet varieties from eastern Africa were planted at Maseno University, Kenya. Chlorophyll fluorescence for healthy, control and infected finger millet varieties was measured using a handheld device MultispeQ. Images (795) of healthy and diseased leaves were captured using a smartphone and uploaded to PhotosynQ app. Convolutional Neural Networks (CNN), Support Vector Machines (SVM) and Random Forests machine learning models were used for blast disease prediction. Chl-F levels reduced as the blast disease progress on finger millet leaves. Modeling the leaf parameters like leaf thickness, relative chlorophyll, leaf angle, leaf ambient temperatures and chlorophyll fluorescence showed that SVM had an F1-score of 76 percent in predicting blast scores compared to other supervised models. CNN classified the images as healthy with an F1 score of 89 percent, blast score of one (0.00perent), blast score of two (25 percent), blast score of three (25 percent), blast score of four (29 percent), and a blast of score five (44 percent). The findings from the study demonstrate the effectiveness of Chl-F as a key feature in the prediction of blast disease score levels on finger millet leaves. The study further demonstrated the necessity to adopt technology such as deep learning models to aid in detecting and classifying the presence of blast disease in finger millet leaves.