INDIVIDUAL TREE CROWN DELINEATION BASED ON DEEP LEARNING FOR ARID AREAS USING HIGH-RESOLUTION SATELLITE IMAGERY

In this work, we explored the potential of the individual tree crown (ITC) delineation and individual tree species classification using deep learning using a Worldview-3 scene covering the Wushen Banner area, Ordos City, where is a typical arid region located in the hinterland of the Mu Us Desert. Combined with field survey data, an ITC delineation sample set and an individual tree species sample set were established. The Mask R-CNN model was employed for ITC delineation, and ResNet-18, GoogLeNet, and DenseNet-40 network models were considered for individual tree species classification. The results showed that the highest accuracy of ITC delineation using MASK R-CNN reached 78.9%. The training accuracies of individual tree species classification models were above 90%, and ResNet-18 achieved the highest classification accuracy. The trained ResNet-18 model was employed to classify tree species from the ITC delineation results of two testing images, and the overall accuracies were higher than 90%.