Study On Estimation Method Of Forest Biomass Based On Neural Network Model

In this paper,we discussed one of the three neural network models(common BP neural network,Erf-BP neural network and improved BP neural network model-BP neural network model based on small-batch gradient descent)respectively.By using MODIS remote sensing images,surface survey Data,topographic data,and forest cover to estimate three types of forest aboveground biomass.For each type of forest,the model factors were selected based on the single factor correlation analysis of aboveground biomass,remote sensing data,topographical data,vegetation index and forest coverage.The model factors were introduced into three neural network models,For each neural network model,the optimal parameters of the three neural network models were trained on the training set and the verification set respectively.The optimal network structure was used as the final neural network model.Three training models were performed on the test set Test,and compared with the stepwise regression model.The main results are as follows:(1)Remote sensing estimation model of coniferous forest biomassPearson correlation analysis was conducted with each independent variable and coniferous forest biomass.The results showed that there was a significant correlation between nine variables and biomass of coniferous forest:NDVI,EVI,B1,slope,B4,B7,forest coverage,B3 and B2.NDVI,EVI explained the highest variation of the biomass of coniferous forest.The prediction of R2 on the same test set is 0.427,0.764,0.818 and 0.835 respectively for the stepwise regression model,ordinary BP neural network model,Erf-BP neural network model and BP neural network model based on small-batch gradient descent,the model training time respectively is 0.427 second,123 Seconds,6433 seconds,65 seconds.Comprehensive comparison of model training time and model prediction accuracy,BP neural network model based on small-batch gradient descent is the most suitable model to estimate the biomass of forest for national forest in North Carolina,followed by Erf-BP neural network model,ordinary BP neural network model,the stepwise regression model is the worst.(2)Remote sensing estimation model of broad-leaved forest biomassOf the 13 remote sensing variables,a total of 8 variables had significant correlations with the biomass of broad-leaved forestland,and the order of their absolute values was B4,NDVI,B1,forest coverage,EVI,B2,B3 and B7.The prediction of R2 on the same test set is 0.427,0.764,0.818 and 0.835 respectively for the stepwise regression model,ordinary BP neural network model,Erf-BP neural network model and small batch gradient descent neural network model,the model training time respectively is 0.264,0.706,0.694,0.720 respectively for ordinary BP neural network model,Erf-BP neural network model and BP neural network model based on small-batch gradient descent.The model training time is respectively 0.346 second,69 seconds,2700 seconds,45 seconds.Comprehensive comparison of model training time and model prediction accuracy,BP neural network model based on small-batch gradient descent is the most suitable model to estimate the biomass of forest for national forest in North Carolina,followed by Erf-BP neural network model,ordinary BP neural network model,the stepwise regression model is the worst.(3)Remote sensing estimation model of mixed forest biomassA total of six variable factors were significantly correlated with the biomass of mixed forest in the plot.The order of their correlation coefficients was as follows:forest coverage,EVI,NDVI,B4,B2 and B1.The prediction of R2 on the same test set is 0.182,0.378,0.511,0.523 for the ordinary BP neural network model,Erf-BP neural network model and BP neural network model based on small-batch gradient descent;the model training time is respectively 0.512 second,97 seconds,3245 seconds,47 seconds.Comprehensive comparison of model training time and model prediction accuracy,BP neural network model based on small-batch gradient descent is the most suitable model to estimate the biomass of forest for national forest in North Carolina,followed by BP neural network model,ordinary Erf-BP neural network model,the stepwise regression model is the worst.The BP neural network model based on small-batch gradient descent can be used to quantitatively study the structural parameters of natural forests.This model overcomes the low precision of traditional regression methods based on the faster model training speed and the prediction speed of traditional regression methods Therefore,the model has the potential to be used for real-time monitoring of natural forest biomass because of its high accuracy and real-time performance.