A Novel Technique Based on Machine Learning for Detecting and Segmenting Trees in Very High Resolution Digital Images from Unmanned Aerial Vehicles

Loukas Kouvaras, George P. Petropoulos

The present study proposes a technique for automated tree crown detection and segmentation in digital images derived from unmanned aerial vehicles (UAVs) using a machine learning (ML) algorithm named Detectron2. The technique, which was developed in the python programming language, receives as input images with object boundary information. After training on sets of data, it is able to set its own object boundaries. In the present study, the algorithm was trained for tree crown detection and segmentation. The test bed consisted of UAV imagery of an agricultural field of tangerine trees in the city of Palermo in Sicily, Italy. The algorithm’s output was the accurate boundary of each tree. The output from the developed algorithm was compared against the results of tree boundary segmentation generated by the Support Vector Machine (SVM) supervised classifier, which has proven to be a very promising object segmentation method. The results from the two methods were compared with the most accurate yet time-consuming method, direct digitalization. For accuracy assessment purposes, the detected area efficiency, skipped area rate, and false area rate were estimated for both methods. The results showed that the Detectron2 algorithm is more efficient in segmenting the relevant data when compared to the SVM model in two out of the three indices. Specifically, the Detectron2 algorithm exhibited a 0.959% and 0.041% fidelity rate on the common detected and skipped area rate, respectively, when compared with the digitalization method. The SVM exhibited 0.902% and 0.097%, respectively. On the other hand, the SVM classification generated better false detected area results, with 0.035% accuracy, compared to the Detectron2 algorithm’s 0.056%. Having an accurate estimation of the tree boundaries from the Detectron2 algorithm, the tree health assessment was evaluated last. For this to happen, three different vegetation indices were produced (NDVI, GLI and VARI). All those indices showed tree health as average. All in all, the results demonstrated the ability of the technique to detect and segment trees from UAV imagery.