Comparison of segmentation methods on images of energy plants obtained by UAVs

Bioenergy crops are widely used as a form of renewable energy. The special form of it is the energy forestry that includes short rotation coppice plantations in which a fast-growing species of tree or woody shrub is grown (e.g. willow, poplar). It is important to accurately estimate the yield before harvest in order to maximize the profit and reduce the costs of production [1]. The accurate prediction of forest biomass and volume can be used for the evaluation of plant breeding efficiency as well. Since energy forestries often contain different trees for estimating their volume it is essential to find segments containing the same tree species in the image. In recent times, the use of Unmanned Areal Vehicles (UAV) became more and more popular in precision agriculture [2], [3]. We investigated the applicability of a low cost UAV in the field of agricultural image segmentation that is the first stage of biomass estimation [4]. This paper compares 3D reconstruction methods that can be performed on aerial photographs of energy plants and focuses on the segmentation that can be applied to find different tree species that form the energy plantation. In order to this, several available segmentation algorithms were evaluated, such as the widely used eCognition and other free software tools like the Orfeo Toolbox. Matlab implementations of segmentation algorithms were also evaluated, such as the Mean-shift segmentation [5], Statistical region merging [6] and Link-based clustering ensembles [7]. The best segmentation accuracy according to the Dice similarity coefficient was reached by the supervised process of the Multi-resolution segmentation of the eCognition by 74.85%. The best accuracy obtained by an automated segmentation method, the Mean-shift segmentation, was 66.12 %.