Automatic classification of healthy / diseased plants using multispectral images

The advancements in drone technology in recent years have led their use to be widespread for different applications, ranging from search and rescue, weather monitoring, mapping and surveying to multipurpose videography. In agriculture, one of the most prominent uses for drone imagery is for monitoring plantation health. As the size of plantations expands so follows the complexity of monitoring the health status of every single plant. This can be attributed to the fact that the traditional evaluation of plants is based on visual checks that go along possibly laboratory analyses which can prove costly in not only in terms of time but also economics terms. This thesis carries on the Dronuts Project which proposes a software and hardware??based solution in order to provide plant classification and monitoring via remote sensing with the aim of assessing the health status of each individual plant relying on collected multispectral drone images. Different software approaches will be applied in for the evaluation, analysis and processing of the drone shots taken over multiple dates to classify independently the health status of a hazelnut plant. We will discuss at first what are multispectral images, how they are acquired and what type of information they carry in the agricultural sector. Subsequently, we will introduce the different Vegetative Indices used in this project with a brief description about each one of them. These indices will allow us to study and assess the health conditions of any crop. Based on these Vegetative Indices we will extract some metrics, which numerically describe the characteristics of the plant on which will be used in later chapters to build and train machine learning algorithms that will allow a classification of the plant in healthy or diseased. Later we will explore the dataset available, its characteristics and possible inadequacies and what approaches can we use to remedy its criticalities. We will proceed by subdividing the images of the dataset into smaller regions using two different methods: Grid subdivision and k-means clustering. Subsequently vegetative indices will be calculated, and labels will be applied so we can proceed to the next chapter. Then we will investigate the applied classical machine learning algorithms like k Nearest Neighbors, Logistic Regression and Random Forest that will be used to classify the plants, we will highlight their training process, and evidently assess the performances they have achieved. Finally, we will look into the application of Convolutional Neural Networks to our dataset. This different approach will be independent of the vegetative indices analysis, using only RGB images as input to the network.