Classification of an agrosilvopastoral system using RGB imagery from an unmanned aerial vehicle

Abstract

This paper explores the usage of unmanned aerial vehicles (UAVs) to acquire remotely sensed very high-resolution imagery for classification of an agrosilvopastoral system in a rural region of Portugal. Aerial data was obtained using a low-cost UAV, equipped with an RGB sensor. Acquired imagery undergone a photogrammetric processing pipeline to obtain different data products: an orthophoto mosaic, a canopy height model (CHM) and vegetation indices (VIs). A superpixel algorithm was then applied to the orthophoto mosaic, dividing the images into different objects. From each object, different features were obtained based in its maximum, mean, minimum and standard deviation. These features were extracted from the different data products: CHM, VIs, and color bands. Classification process – using random forest algorithm – classified objects into five different classes: trees, low vegetation, shrubland, bare soil and infrastructures. Feature importance obtained from the training model showed that CHM-driven features have more importance when comparing to those obtained from VIs or color bands. An overall classification accuracy of 86.4% was obtained.