Drone Imagery in Support of Orchards Trees Vegetation Assessment Based on Spectral Indices and Deep Learning

Over the years, the detection and classification of crown trees raised much interest for the scientists from the forest and environmental sciences, due to their essential role for landscape ecology and forestry management. Besides crown trees delineation and trees classification, an essential part of the trees health assessment is the extraction and estimation of vegetation indices (VI). These VI are exploiting the differences between the visible spectrum (RGB) and the near-infrared spectrum (NIR). Approximately a decade ago, detection of individual trees was focused on Lidar coupled with high-resolution ortho imagery or hyperspectral images, but with the increasing availability of unmanned aerial vehicles this gradual shifted towards Structure from Motion application. Building on the advantages of drone technology and the latest deep learning algorithms, the present study aims at assessing the combined use of ML techniques with spectral VIs derived from visible cameras mounted on drones, to be used as a proxy to characterise vegetation health of individual trees in an orchard field. To accomplish the study objectives, several image processing methods were implemented to the acquired drone data. The tree’s crown was extracted the crown trees by applying a deep learning object instance segmentation method. For mapping, the vegetation health, VI from visible spectrum (Red, Green and Blue) were used. Very good results were obtained in the case of the plum, apricot and walnut trees, mostly because these trees have the leaves oriented towards the camera and the spaces between leaves and branches are much smaller in comparison with the olives trees. Less reliable results were obtained for olive trees crown delineation because of its specific texture with small leaves and large spaces between branches. The signal received from the ground had an essential influence in the assessment of the vegetation health status by increasing the (Green Leaf Index) GLI mean values with a small fraction. Overall, the study demonstrates the real potential of drone applications and deep learning methods for spatial and temporal rapid assessment of trees vegetation heath.