Low-cost multi-spectral vegetation classification using an Unmanned Aerial Vehicle

In Precision Agriculture (PA), the decision support system is expected to be able to assist determining the needs of the farm. One of the crucial needs is to specify the amount of different fertilizers to be used, and also distinguished hydration levels of crops. The amount of necessary fertilizers to apply is directly related with the area occupied by a certain type of plant, and with the hydration levels of such area. In order to achieve such a goal, Unmanned Aerial Vehicle (UAV) can be used to scan over an area, while detecting and classifying the type of vegetation in herbaceous crops. The majority of current monitoring technologies are very expensive, or the low cost systems use cameras that will gather information only in the visible spectrum. Therefore, we propose a low-cost multi-spectral system, where an Unmanned Aerial Vehicle (UAV) was equipped with a set of exchangeable filters over a camera, connected to a Raspberry Pi (RPi). Two classifiers were implemented and optimized in order to maximize the true positive rate (TPR) while minimize the false positive rate (FPR). The entire system is automated and the classification output is provided from the RPi to a ground station in real-time, by a Wi-Fi socket connection. The classifiers have shown to be able to distinguish, based on our sensor data, two types of vineyard and tree species of plants. For comparison purposes, we present results showing the performance of both classifiers while using data gathered by our system. The Region Of Interest (ROI) was identified by a thresholding algorithm based on Normalized Difference Vegetation Index (NDVI) measurements.