Armyworm detection using machine learning

Abstract

Uganda and Africa at large largely depend on agriculture. However, currently, Uganda's agriculture sector has faced a problem of armyworms that destroy people's crops and in the long run affect the country's food security. This project aims at making use of a combination of an algorithm design (machine learning) and tensorflow for armyworm detection. We did this through data collection. The first step is to collect a large dataset of positive training examples (images containing the armyworm). In this case, we collected 1000 images from Namulonge. We did data labeling and cleaning, where we used the labeling tool to label the data, a process called annotation. The majority of practical machine learning uses supervised learning; our problem falls under the classification task. In classification, we used convolutional neural networks, and for object detection in our case, we used the single-shot multibox detector. We split the dataset into two sets, training and testing. The training folder consisted of images that the algorithm would learn from while the testing folder consisted of images that the algorithm would test its performance on. We used tensorflow to implement the SSD algorithm, using its Convolutional Neural Network API and its object detection API. After the training process, we exported the inference graph. The inference graph is a frozen version of the model with the appropriate weights to perform predictions. We call our inference model within jupyter Notebook, where we can see the model in action. We aim to reduce the amount of money farmers spend on pest control and the destruction of crops by pests and thus improve the nation's food security.