Single tree crown extraction based on gray gradient image segmentation

Tree crown is an important part of trees. It is of significance to extract tree crown information based on remotely sensed images for forest resource inventory and monitoring. However, it's difficult to extract the individual tree crown shape accurately. High spatial resolution image has an abundance of texture and spectral information, which provides a potentially efficient approach to delineate individual tree crown for forest resource inventory. However, with its abundance of information, the object-oriented image segmentation based on the original high resolution image has lower efficiency because of the large calculation and poor robustness since it needs setting spectrum or texture threshold manually. The method of image enhancement highlights or suppresses certain image features selectively by changing the image structure, so effective image enhancement can improve the accuracy and efficiency of the individual tree crown segmentation. In this article, a new gray-gradient image segmentation method was proposed to realize rapid and high accurate extraction of the individual tree crown. For the comparative analyses, we selected conventional Roberts and Laplacian operator, along with the proposed modified mathematical morphology operator as alternatives, subsequently, it was confirmed that the optimal operator was the modified mathematical morphology operator by combining visual interpretation and gradation histogram analysis. Furthermore, the modified mathematical morphology operator combined with object-oriented multiscale segmentation classification method was used for simplifying background information of raw image and extracting large-scaled single-tree crown information rapidly. To validate the efficiency of the method, CCD image of airborne laser radar in Dayekou forest region in Zhangye, Gansu Province of northwestern China was used to extract individual tree crown. The results showed that by using high spatial resolution gray gradient images, the location accuracy of tree crown was 83.19%, and the shape accuracy of crown was 88.62%, both of which were superior to the individual tree crown segmentation based on the original high spatial resolution image. The crown edges are drawn fast, efficiently and relatively accurate.