Selecting estimation parameters for the Finnish multisource National Forest Inventory

Abstract

The paper examines the selection of parameters for the nonparametric k-NN estimation method that is used in the Finnish multisource National Forest Inventory (MS-NFI). The MS-NFI utilises NFI field plot data, optical area satellite images and digital maps and produces forest variable estimates from the single pixel level up to the national level. The most important parameters to be selected are: the distance metric, the number of the nearest neighbours, k, parameters related to the digital elevation model, stratification of the image data, as well as the width of the moving geographical horizontal and vertical reference areas (HRAs and VRAs). The root mean square errors (RMSEs) and significance of biases at pixel level were evaluated in order to find optimal parameters. A leave-one-out cross-validation method was applied. The emphasis is placed on the search for moving geographical HRAs and VRAs, as well as in the stratification of the field plots and the satellite images on the basis of auxiliary data. Stratification reduces the bias of the estimates significantly within each strata. With the current sampling intensity of the Finnish national forest inventory, a geographical HRA with a radius of 40–50 km was found optimal for the total volume estimates and for volumes by tree species in the mineral land map stratum. On the average, there was a sufficient number of field plots to cover the variation of forest variables within the image area to be analysed. The inclusion of field plot data beyond this area introduced bias to the estimates. For the peatland strata, a wider reference area, 60–90 km, was needed. A VRA, together with topographic correction of the digital values of images, reduced the standard error of the volume estimates in Northern Finland.