Development of Machine Learning Model for in-field Detection and Counting of Whiteflies.

Abstract

Cassava production has been greatly affected by whiteflies, which are a key pest and vector. When left unchecked, whiteflies can reduce crop yields, degrade cassava quality, and, in extreme situations, result in total crop loss. As a result, knowing the number of whiteflies is crucial for detecting, identifying, and preventing their spread. Visual examination and manual counting of whiteflies are the current methods for identifying and counting whiteflies; nevertheless, these traditional ways of counting whiteflies have drawbacks such as inefficiency and time consumption. In the last decade, there have been significant advancements in the science of machine learning, particularly in the field of computer vision for application such as self-driving cars, video analytics, edge computing and better triangulation of sensor data. In this project, we employ current advancements in computer vision to assist in the detection and counting of whiteflies in order to help farmers in Sub-Saharan Africa reduce food shortages and increase food security. Research has been done by various people on the topic of detecting and count- ing whiteflies. Faster R-CNN and Haar-Cascade, image processing techniques with MATLAB and YOLOv3 have all been employed as approaches in detecting and counting of whiteflies. In this project we used the YOLOv5 which is an object detection algorithm that came out in 2020 to tackle the same problem. During training we managed to achieve an mAP of 0.842 @0.5IOU. The results show a 20% improvement when compared to the current state-of-art models. The approach proposed in this project utilized tiling operation on the image data to enhance the performance of the YOLOv5 model on the whitefly dataset.