Forest Tree Extraction And Coverage Estimation Based On Convolutional Neural Network

With the development of the Times,the country pays more and more attention to the urban ecological environment,and gradually shifts its focus to the urban forest community,planting large-scale plantation forests close to natural succession,and monitor the status of forest resources and forest changes by observing changes in forest coverage.The forest area and coverage rate are selected as the main indicator to measure the urban forest resources.Because of the long acquisition cycle and not high enough resolution of the remote sensing image,it is difficult to improve the forest coverage estimation accuracy.In recent years,with the rapid development of unmanned aerial vehicle(UAV),the use of UAV images to estimate forest coverage has become the focus.This paper takes the experimental forest farm in Harbin City of Heilongjiang Province as the research object,and uses the estimation of forest coverage as the research center.UVA forest images are obtained by UAV to estimate the forest coverage.Around the forest coverage estimation center,the research focused on the methods of forest tree extraction in the forest image,in order to achieve accurate forest segmentation of each plant,and estimate the forest area and coverage rate based on the segmentation images.Firstly,several common models of forest images segmentation are studied to select the optimal forest segmentation results.As the problems on forest trees by the optimal FCN-8s network model,the Improved Fully Convolutional Neural Network(IFCN)model is designed.The segmentation ability of the model on forest trees is not affected by small batch sizes,which effectively solves the problem of unstable IFCN model performance.The network structure is changed.The high and low level information of forest trees are fused to improve the learning ability and recognition ability of IFCN network.Edge refinement processing at the back end of the IFCN network is introduced to refine the segmentation result at the edge of the forest.Secondly,as achieving good results in the forest image segmentation task by IFCN model,it is necessary to introduce new methods in the backend.For the problem,a lightweight IFCN network model is designed to segment and extract trees,and the multi-scale features of forests are used to improve the network structure.Compared with the IFCN model,the complexity of the IFCN model is simplified.It does not reduce the IFCN's ability to learn the characteristics of forest trees,which further improve the accuracy of forest identification and segmentation.Finally,the improved IFCN network model was adapted to the estimation of forest coverage in the study area.Then this paper compared and analyzed the effectiveness and practicability of the model by multiple sets of comparative experiments,which provides a reference for forestry applications.