Forest Land Change Detection Based On Remote Sensing Images Using Robust Linear Regression

Dynamic monitoring of forest resources is the core content of forest resource management and an important information support for achieving the goal of increasing forest area and volume.China has initially established a map of the national forest land covering the mainland and Hainan Island,and stocking volume and forest land data are updated every year.An important part of the data updating is the annual national forest land change survey.This survey is based on remote sensing images,automatic computer identification,visual interpretation and field investigation.With the continuous enrichment of remote sensing data and the continuous progress of computer technology,the application of computer automatic identification technology in change monitoring has become more and more important.For the remote sensing data of the early and late stages,the areas that change insignificantly accounted for the majority,and change obviously account are in minority.Statistically,the data of change is equivalent to “outlier”.So the core of change detection is to monitoring these anomalous data.Robust regression is an important method of monitoring outliers.In this paper,robust linear regression technology is applied to the monitoring of forest land type changes.Taking Lin 'an District of Hangzhou City,Zhejiang Province as an example,the monitoring of forest land type changes is mainly based on three parts:Data preprocessing is performed by three methods: no processing,principal component transformation,and canonical correlation analysis.Based on the data preprocessing,the change information is obtained by three methods: robust linear regression,ordinary least squares regression and image difference.Based on the change information,four kinds of supervised classification algorithms,such as binomial logistic regression,multiple linear regression,support vector machine and BP neural network,are used to divide the land types into two categories: change and non-change,and finally obtaining the information of forest land type change.The results show that compared with the ordinary least square(OLS),robust linear regression(RLR)has better robustness in estimating regression parameters in the presence of outliers.RLR is not sensitive to data preprocessing methods,followed by OLS,and image difference(IM)is the most sensitive,especially for data transformed by principal components analysis(PCA).The average total precision,average user precision,average producer precision and average Kappa coefficient of 12 results based on RLR are all higher than those of OLS and IM.The average total precision and Kappa coefficient reach 98.65% and 0.972 respectively.In data preprocessing methods,canonical correlation transformation(CCT)is helpful to improve the monitoring accuracy of forest land type changes.Among the 4 discrimination methods,9 results of binomial logistic regression(BLR)have the highest average precision,support vector machine(SVM)is the most stable,BP neural network(BPNN)has the same precision as SVM,and multiple linear regression(MLR)has the worst performance.