Vegetation Classification Using a Small UAV Based on Superpixel Segmentation and Machine Learning

Abstract

This paper describes a low-cost and flexible vegetation classification system using a small and lightweight unmanned aerial vehicle (UAV). Compared with traditional remote sensing systems such as airplanes and satellites, small UAVs can obtain very high (a few cm) resolution aerial images. They are unaffected by cloud cover because they can fly at a very low altitude. In addition, owing to their small size and light weight, UAVs are highly portable, have a low operational cost, and can be operated safely. However vegetation classification based on traditional pixel-based classification approaches cannot work well when aerial images with cm-level resolution are used. Furthermore, the area of the aerial image captured by a UAV is limited to a small region due to its low-altitude flight. We propose a new method for determining vegetation type from aerial images obtained by a small UAV based on superpixel segmentation and machine learning techniques. We developed a method for creating a wide-area high-resolution mosaic image from multiple aerial images obtained by the UAV and onboard sensor data such as GPS and inertial sensors. Superpixel segmentation based on mean shift was then used to divide the mosaic image into small regions. Finally, the vegetation type was classified using the support vector machine (SVM) training and classifying process. We conducted our experiments at Yawata moor in Hiroshima Prefecture to classify moor vegetation. We evaluated the success rate of our proposed method based on the results of these experiments, and concluded that the small UAV system is both effective and useful for vegetation classification.