Abstract:

The phenomenon of data loss often occurs in forest resource sampling investigation.So it is necessary to study the filling method of missing data in order to improve the accuracy of the data analysis.Linan County located in Zhejiang Province was chosen as the case study area.Landsat-5 TM image in 1996 and County-level fixed plot data of forest resources continuous detection in the same period were used as the main information,and the average DBH(Diameter at Breast Height) of trees in sample plot as the missing factor to make spatial filling,non-spatial filling,model filling of remote sensing estimation for missing data.And 10 fold cross-validation method on the basis of spatial autocorrelation analysis of the average DBH of trees in sample plot was employed to make accuracy evaluation.The results show that:(1) The Moran’I coefficient of the average DBH of sample plot trees in study area is 0.21 and its spatial distribution shows strong spatial autocorrelation;(2)The filling accuracy of K-Nearest Neighbor of remote sensing estimation models is the highest,the second is Random Forest followed by the Kriging Interpolation of spatial filling.However,the filling accuracy of expectation maximization algorithm of non-spatial fillings is the lowest;(3)Among four semi-variance models of Kriging interpolation,the filling accuracy of spherical model is higher than any other models.Its correlation coefficient constitutes 0.632 5,the mean absolute error makes up 2.049 3 centimeters and the root mean square error accounts for 3.809 3 centimeters;(4)According to the order of filling accuracy from high to low,four priority filling methods of missing data includes:K-Nearest Neighbor,Random Forest,Kriging Interpolation and Inverse Distance Weighting.It is the K-Nearest Neighbor that is most suitable for filling missing data of the average DBH of sample plot trees in Linan with complex topography and great different altitudes.

Key words: forest resource sampling investigation, DBH, missing data, filling methods, Linan