Tree detection using UAV based imagery system based on Random Forest classification

Recent advances in remote sensing have provided access to UAV aerial imagery, among different sensors mountable on UAVs optical cameras are the most cost-effective ones. Therefore, in this study, an optical camera is used to extract the class of trees from other urban features. Two sets of data are used, the first area has an area of ​​3.07 hectares and a GSD of 2.23cm, and the second one has an area of ​​104.5 hectares and a GSD of 8.75cm. The classification is done in two cases using Random Forest, in the first case we use only the visible image and the feature vectors extracted from it. In the second scenario, considering that both study sites are related to urban areas and the land is almost flat, in addition to the image, DSM is also used as a feature vector. The results show that with the addition of the digital surface model (DSM) the total accuracy of the classification compared to only visible data in the two study areas increases by about 16 and 12 percent, respectively, and reaches 93 and 90 percent. In the second part of the article, we examine the effect of reducing feature space using PCA and compare it with the situation in which all features are used for classification. According to the results, if all the extracted features are used, the processing time will increase and the total accuracy will decrease due to the dependencies in some feature spaces.