Estimation Of Forest Volume Based On Variable Optimization And Deep Learning

Forest stock is one of the important indexes reflecting the quantity of forest resources.Based on the variable optimization and depth neural network algorithm,this study takes small class as the research unit to estimate the forest resources per mu in some areas of Longquan City,which provides a new method and idea for the estimation of forest reserves at the county level.Based on the secondary survey data of forest resources,high-resolution two(GF-2)remote sensing image data and digital elevation model data in the study area,multi-source data is extracted to form the original feature set.The whole variable data set and the three data sets obtained by the variable optimization method(boruta feature selection,Pearson correlation analysis,principal component analysis)are used for modeling and accuracy test by the partial least square algorithm to determine the best data set for subsequent experiments.Two kinds of deep learning models are used: the improved deep neural network(LMBP-DNN)based on Levenberg Marquard algorithm,and the integrated model based on convolutional neural network and support vector machine regression Regression,CNN-SVR),two shallow machine learning models: Levenberg Marquard back propagation neural network(LMBP-NN),support vector Regression(SVR)is used to train the estimation model of forest volume,and the modeling effect is tested by the methods of 10 fold cross validation,grid search and experiment to determine the optimal parameters of the model,and the fitting and generalization performance of the four models are compared and analyzed.The results show that:(1)The data set obtained by the boruta feature selection method is the best.Six features are retained after selection: soil thickness,age,canopy density,altitude,slope and aspect;(2)Ten fold cross validation combined with grid search can quickly and accurately determine the optimal combination parameters of the model and improve the accuracy of the model.(3)The generalization ability of LMBP-DNN and CNN-SVR forest volume estimation models have their own advantages and disadvantages.The root mean square error(RMSE),correlation coefficient(R),mean absolute error(MAE)of LMBP-DNN are better,and the average accuracy and overall accuracy of cnn-svr are better.(4)The modeling results and generalization performance of the two deep learning models are better than that of the shallow machine learning model.The average estimation accuracy is over 81%,and the overall accuracy is over 97%.The improvement of the shallow machine model based on the deep learning algorithm can improve the performance of the model,which has certain advantages and research value in the field of forest storage estimation...