Monitoring of banana's moko disease (Ralstonia Solanacearum) using multispectral imagery

Abstract

As one of the most important cash crop in the world, banana cropping is constantly threatened by a lot of diseases: Panana disease, Yellow Sigatoka, Black Sigatoka, Moko disease, etc. Not all diseases are treatable by phytopharmaceutical products. Aside from prophylaxis, their early detection is becoming a priority for large scale banana growers. Remote sensing is a possible efficient tool to achieve this goal. It has shown significant results throughout the whole agricultural industry for plant disease monitoring. The aim of this study is to create a classifier of Moko disease (Ralstonia Solanacearum) occurrence five days before the appearance of the symptoms in visible light for Food and Agriculture Industries‘ (FAI) banana plantation of Nieuw Nickerie, Surinam. To this end, ground data were collected every five days and drone flights were conducted. The drone held a multispectral camera (red, green, rededge, near infrared). Around 1000 diseased plants were manually geolocalised. Different steps were followed. Firstly, a banana/non banana classification classifier was created to segmentate the raster images and create a mask of the banana plants. This classifier gave a 96% global accuracy on the randomly generated validation set. A pixel-based and an object-based approach were tested for the classification of the raster image, respectively with a level of 95% and 73% of global accuracy. Secondly, another classifier for the occurrence of the Moko disease was created. It showed a 98% global accuracy applied on the randomly generated validation set. However, it gave mediocre results when the classifier was applied to rasters by pixel-based (53%) and object-based approach (57%). Multiple machine learning algorithms are tested to create the classifier. This study shows that local maxima algorithm for object detection does not allow the user to trust classifiers on high overlapping culture such as banana on an individual based approach.