Abstract:

This paper aims to investigate an automated approach for extracting spatial distribution and quantities of planted trees in plain areas by using of visible light imagery obtained from unmanned aerial vehicles.Specifically,the study focuses on analyzing afforestation patches located in Huailai,Zhangjiakou,Hebei Province.To achieve this,a manually constructed single-tree crown dataset was utilized to train and predict Mask R-CNN models with varying backbone networks(R50-FPN-1x,R50-FPN-3x,R101-FPN-3x),and their accuracies were compared on an independent test set.Results indicate that all backbone models were able to segment tree crowns in the afforested areas,with the R101-FPN-3x model achieving the highest accuracy of 75.33% and a recall rate of 73.23%.The results showed that the single tree segmentation method based on Mask R-CNN from UAV images could quickly and accurately detect single tree targets in afforestation areas automatically,and effectively segment single tree crown to map the crown contour,which could meet the needs of efficient investigation and monitoring in small and medium-sized plain afforestation areas.

Key words: UAV, Single-Tree Crown Segmentation, Mask R-CNN, afforestation evaluation