Research On Individual Tree Segmentation And Tree Species Classification Method With Adaptive Metric

Research on individual tree segmentation(ITS)and tree species classification is a basic task in precision agriculture and forestry construction.It is essential for sustainable forest management,ecosystem service assessment,biodiversity monitoring and ecological environment protection.With the rise of machine vision technology and the convenience of data acquisition,the application prospect of precision research based on vision is broad.However,in the field of tree segmentation and tree species classification,taking the two-dimensional aerial view as an example,the current technical difficulties are mainly manifested in the following aspects: 1)The complexity of the problem,such as the diversity of tree species,the uncertainty of tree crowns(gradient colors and uncertain shapes),the variability of forest scenes,and the overlap between tree crowns,etc.,pose new challenges to visual technology research;2)Existing methods,such as image processing or feature construction methods,inevitably weaken the generalization ability of models because of their cumbersome steps.In the research of such problems,this paper mainly completes the following work:1.An adaptive metric based on supervised clustering is proposed.In order to overcome the isotropy problem of Euclidean distance metric,this paper introduces an adaptive similarity metric,i.e.,from a small number of marked crown “clusters” provided by users,the similarity between crown individuals is learned and applied to work 2 and 3.Experiments are conducted on artificial and public datasets to verify the effectiveness of the learned metric.2.For complex forest environment in aerial images,a supervised clustering-based crown segmentation framework is proposed for tree counting.Its advantages are: 1)Use a pixel classifier for identifying tree pixels or superpixels,replacing complex image preprocessing or feature extraction techniques;2)The similarity measure used by proposed supervised clustering method is learned from the labels provided by users,instead of requiring pre-specified or grid-searched optimal parameters like existing methods;3)Individual trees can be accurately segmented even in crown overlapping areas.This method can accurately obtain crown segmentation boundaries and tree counting from 2D forest aerial images.3.A multi-scale information aggregation method is proposed.Based on segmentation results obtained in work 2,the obtained individual tree images are substituted into trained deep model to predict tree species.Aiming at the problem that individual scales of trees in aerial images vary greatly,the algorithm uses multi-scale aggregation information to obtain the context information of the target and enhance the classification accuracy of tree species.At the same time,by using artificial recognition to segment tree images at multiple scales,a dataset of related tree species images was established to allow the model to learn multi-scale contextual information during training.