Study On Tree Species Identification Of Remote Sensing Data Based On Three-dimensional Residual Convolutional Neural Network Algorithm

Aiming at the problem of low precision of remote sensing tree species identification under complex forest canopy structure and high stand density.East Anhui forest region was selected as the research area,and the multi-source information such as GF-5 AHSI,GF-6 PMS image,digital elevation model and forest resource survey data,are used to extract classification feature factors.JM distance and linear discriminant analysis were used to select the best feature factors for tree species identification.The 3D residual convolutional neural network(3D-RCNN)tree species identification model was proposed and constructed,and the accuracy was verified by combining the feature combination scheme to evaluate the performance of the 3D-RCNN model,and the influence of feature factors on the tree species identification accuracy was explored.This essay studies the results as below:(1)Through the analysis of inter-class separability and linear discrimination,51 features are selected from 461 classification features as tree species identification best feature factor,and the number of spectral features,texture features,vegetation index,and topographic features are 26,12,10 and 3.In the spectral feature set,11 belong to the near-infrared long wave band,9 belong to the near-infrared short wave band,and 6 belong to the visible light band;4 mean values in texture feature set,3 second-order angle moments and entropy,1 each for uniformity and correlation;in the feature set of vegetation index,there were 3 indexes for leaf pigment and 3 indexes for canopy water content,2 indexes for greenness and 1 index for carbon attenuation and light use efficiency;terrain features are elevation,slope,and aspect.(2)The 3D-RCNN algorithm to build tree species recognition model are as below:the pixel block size is composed of 17 × 17 × L input layer(L is the spectral dimension of the input sample),5 3D residual convolution units,a fully connected layer and a softmax classifier.Among them,the size of the convolution kernel is 5 × 5 × 5,and the standard step size is 2.The number of convolution kernels varies with the number of 3D residual units in order of 8,16,32,64,128.The convolution kernel uses L2 regularization,and the maximum pooling size is 2 × 2 × 2.BN and Dropout strategies are used after the fully connected layer,where the dropout coefficient is set to 0.4.Using the back-propagation algorithm,the batch-size is set to 128,the loss function is optimized using the RMSprop optimizer,and the learning rate is 0.0004.(3)The 3D-RCNN algorithm model was cross-validated with support vector machine(sVM)and random forest(RF).Tree species identification and accuracy evaluation were carried out for different feature combination schemes.The study shows:1.the 3D-RCNN model can effectively extract high-dimensional data features,and performs well in the classification of feature factors.It improves the overall accuracy of traditional machine learning algorithms from 85.22%to 91.72%.2.Compared with a single feature factor,the combination of multiple feature factor tree species has higher recognition accuracy.The highest recognition accuracy of the single feature obtained by the spectral feature is 70.33%,and the best recognition accuracy of the multiple feature combination obtained by the combination of the spectral feature,the texture feature and the vegetation index feature is 91.72%.With the increase of feature factors,the recognition accuracy also improves accordingly.Among them,the vegetation index features improve the accuracy more than the texture features.The terrain features are affected by altitude and human factors,and their effects are minimal.3.The overall accuracy of the 3D-RCNN model is 91.72%,and the average relative accuracy of the tree species survey is 92.92%.It can well identify the subtropical forest tree species with complex forest composition.The model performs well when the training samples are sufficient.This shows that as the sample set is enriched,the model still has great optimization potential.It is also proved that the domestic GF-5 AHIS and GF-6 PMS data can complement each other well and effectively identify subtropical deciduous broadleaved forests and evergreen coniferous forests.