Forest Health Assessment Of Western Hunan Based On SPOT-5 Image Spectral And Textural Features

Forest health assessment is an important basis for forest management.However,traditional forest health research mainly relies on field investigation,which is time-consuming and labor-intensive in the data collection process,and it is difficult to achieve wide-area effective monitoring.In recent years,a large number of studies have shown that remote sensing data has a good correlation with the quantity,structure and function of forest resources.Based on the correct understanding of the connotation of forest health,the use of remote sensing technology with macroscopic,timeliness and high efficiency is an ideal means for forest health research.The main purpose of this paper is to establish a set of forest health quantitative index system,and analyze the effect of forest health index based on SPOT-5 remote sensing image spectral and texture features to provide a theoretical basis for forest health survey.Taking the seventh forest resource inventory data and SPOT-5 remote sensing data of Hunan province as the data source,and taking the image coverage area as the research object.According to the characteristics of the research area and the health characteristics of the forest ecosystem.Selected by average diameter at breast height,average basal area of breast height,stand density,the stand volume,Simpson index,Shannon-Wiener index,Pielou index,mixed degree,the Gini coefficient,the standard deviation of diameter at breast height,angle scale,soil thickness and crown density forest quantitative indicators to evaluate the health of forest property.The factor analysis method was used to determine the weights of each evaluation index.Finally,four common factors were extracted,which represented species diversity,forest stand competition status,tree size diversity,forest soil,and relative weights were F1=0.337,F2=0.314,F3=0.247,F4=0.102.The calculated forest health index follows a normal distribution after normal testing,and use normal isometric analysis to classify forest health into four levels:healthy,sub-healthy,moderately healthy,and unhealthy.The results showed that in the study area,unhealthy forests accounted for 11.8%;middle-health forests accounted for 44.1%;sub-health forests accounted for 30.5%;healthy forests accounted for 13.6%.On the whole,the forest health in the study area was not high,and the healthy and sub-healthy forest stands accounted for the majority.Based on the remote sensing image preprocessing,the SPOT-5 image spectral information,vegetation index and eight GLCM texture feature variables under nine different windows were extracted using ENVI software.Pearson correlation analysis of texture feature variables and forest health index under different windows.It was found that most of the correlation coefficients reach the peak in the 13 × 13 window.Therefore,the texture information extracted under the 13 x 13 window was selected to construct the forest health remote sensing estimation model.The 21 remote sensing feature variables extracted from the image were filtered by LASSO regression algorithm.It was found that the mean square error was the smallest when the number of variables was seven.The optimal combination of the selected independent variables was the surface average reflectance of the panchromatic band(mean_pan),the mean surface reflectance of the short infrared band(mean_swir),the water stress index(MSI),the simple ratio vegetation index(SR),and the mean,variance,homogeneity in the GLCM texture characteristic variables.The overall sample data was randomly divided into 40 training sample samples and 17 test set samples according to the 7:3 ratio.Using multiple linear regression,partial least squares regression,vector machine and random forest to establish four forest health estimation models.The decision coefficient(R2),root mean square error(RMSE),error mean(AE)and estimation accuracy(EA)were used as the basic indicators for evaluating the accuracy of the forest health model.It can be seen that the estimation model established by the support vector machine algorithm has the highest estimation accuracy,reaching 76.94%,the decision coefficient R2 is 0.796,the root mean square error is 1.225,and the error average is 0.995.Secondly,the random forest algorithm determines the coefficient R2 is 0.732,the root mean square error is 1.397,the error average is 1.172,and the estimation accuracy is 73.71%.The partial least squares regression model determines the coefficient R2 is 0.597,the root mean square error is 1.713,and the error average is 1.464.The accuracy is 67.75%;the multiple linear regression is relatively poor in the four algorithms.The decision coefficient R2 is 0.528,the root mean square error is 1.854,the error average is 1.586,and the estimation accuracy is 65.09%.It is indicated that the forest health estimation model established by the support vector machine algorithm has a good application prospect.The two traditional methods are compared with the modeling results of machine learning methods,and machine learning is superior to the traditional regression method.