Soil / Crop Segmentation from Remotely Sensed Data Acquired by Unmanned Aerial System

Today Unmanned Aerial Systems (UAS) are widely used for many applications that involve advanced payload as is found to be the case for mounted remote sensing apparatus. Remote sensing from UAS platforms is now common and the use of light and smart multi/hyper-spectral cameras has opened the field to novel applications. These sensors can operate in cloudy conditions ensuring ultra high resolution images while at the same time overcoming the limitations of satellite photography. In this paper we focus on just one such advanced payload application, namely, the segmentation of tree-cover / canopies over soil terrain. This task is mandatory in order to mask-out areas that are not of direct interest. The approaches studied are based on both supervised and unsupervised algorithms which take into account multi-spectral as well as synthetic features derived from the Digital Surface Model (DSM). We process the DSM by testing 2D convolution kernels together with a pseudo-random image slicing that tries to derive/model the ground/soil profile. Global thresholding is not able the segment tree / canopy area over the soil because the terrain slope is subject to significant change over small areas as is often seen to be the case with vineyards. The proposed approach takes into account such local variability to ensure a correct segmentation analysis in presence of slopes or other undulatory terrain variations. The results obtained show that the proposed method enables the segmentation of tree / canopy vs soil with an overall accuracy greater than 95%.