Research On Classification Method Of Forest Tree Species Based On BCResNet

Forest resource information is an important data and main basis for forestry planning,management and decision-making.Research on automatic classification methods of forest tree species based on remote sensing images is of great significance for inventory of forest resources,retrieval of forest carbon storage and sustainable management of forest resources,and is also an important link to achieve accurate monitoring and management of forest resources.Sentinel-1A radar images contain forest vertical structure information,Sentinel-2A multispectral images contain forest horizontal structure information and are sensitive to vegetation.Digital Orthophoto Map(DOM)has high resolution and rich information,so it is urgent to fuse multi-source remote sensing images to extract forest spectral,vegetation index,and texture features for automatic classification of forest resources.This thesis takes Pizhou City as the research area.Pizhou City is located in the northern part of Jiangsu Province,with a predominantly plain terrain and abundant forest resources,with a coverage rate of 30.8%.Based on Res Net50 network model,this thesis optimizes its network backbone structure,activation function and loss function,and designs BCResNet(Bilinear Revolution Res Net)network model for automatic classification of tree species.The research work of this thesis is mainly divided into the following two aspects:(1)Construction of a forest tree species classification feature set based on multi-source data fusion.Based on the phenological patterns of forests,remote sensing reflectance experiments were conducted using images from different months,and representative remote sensing images were selected as data sources;Preprocess various remote sensing images,unify geographic coordinate system and map projection;Based on the bilinear interpolation method,the image data is resampled to the resolution of 1 meter,and normalized to eliminate dimensional differences.The sample feature set of the model is constructed to facilitate the sample feature extraction of the model.Design different feature combination classification experiments to find the optimal feature combination for classification.(2)A forest tree species classification method based on the BCResNet network model.Based on the Res Net50 network model,Swish activation function is selected to optimize Re Lu activation function to solve the problem that too many dead zone neurons reduce the utilization of model parameters;A method combining smooth label with Focal Loss loss function is proposed to solve the imbalance problem of sample classification data and reduce the possibility of overfitting in the network;Designed a bilinear self-attention mechanism and introduced a dual Res Net50 network with shared weights to form a twin network,constructing a BCResNet network model.The ablation experiment was designed and the hyperparameter design of the ablation experiment was completed.By calculating the confusion matrix of forest tree species classification results of BCResNet network model and three ablation control models,obtain Kappa coefficient,overall classification accuracy,Produce’s Accuracy,and User’s accuracy.Comparison shows that the BCResNet network model proposed in this thesis has the highest Kappa coefficient and accuracy for forest tree species classification results,with a Kappa coefficient of 0.924.Compared with the three ablation control models Res Net50,BCResNet(Re Lu),and BCResNet(CE Loss),the Kappa coefficient increases by 0.115,0.026,and 0.047,respectively;The accuracy of forest tree species classification is 94.24%,which is 10.71%,2.12%,and 3.59% higher than the three ablation control models,respectively.The BCResNet network model has strong reliability in forest tree species classification research.