Evaluation of Natural Forest Canopy Segmentation Based on CA-Net for Unmanned Aerial Vehicle Imagery

Taking Wuyi Mountain National Park in Fujian Province as the study area and three types of natural forests, including Pinus massoniana,Fir and broadleaf mixed forests, CA-Net, an automatic canopy extraction method based on deep learning combined with coordinated attention mechanism was proposed.The results show that:（1） the Watershed algorithm was poorly applicable in the natural forest dataset, whereas the mIoU,mPA and Accuracy of CA-Net is improved comPAred with U-Net, improved by 4.75%,3.27% and 2.84%,respectively.（2） On the background category, CA-Net′s IoU,Recall,and Precision reached 69.58%,84.17% and 80.06%,respectively; whereas on the canopy category, CA-Net′s IoU,Recall and Precision were 83.85%,90.11% and 92.35%,respectively.（3） CA-Net had the highest segmentation Accuracy on mixed fir-broadleaf forests, and the segmentation Accuracy mIoU increased by 1.83% and 3.12%,mPA increased by 6.93% and 2.95%,Accuracy increased by 3.57% and 1.83%,respectively, comPAred with that of U-Net for mixed Masson pine-fir forests, and mixed Masson pine-fir-broadleaf forests.It can be seen that CA-Net can effectively overcome the problems of background interference and difficulties in fine feature extraction to improve the accuracy and efficiency of canopy segmentation in complex natural forests.