Using coarse scale forest variables as ancillary information and weighting of variables in k-NN estimation: a genetic algorithm approach

The non-parametric k-nearest neighbour (k-NN) multi-source estimation method is commonly employed in forest inventories that use satellite images and field data. The method presumes the selection of a few estimation parameters. An important decision is the choice of the pixel-dependent geographical area from which the nearest field plots in the spectral space for each pixel are selected, the problem being that one spectral vector may correspond to several different ground data vectors. The weighting of different spectral components is an obvious problem when defining the distance metric in the spectral space. The paper presents a new method. The first innovation is that the large-scale variation of forest variables is used as ancillary data that are added to the variables of the multi-source k-NN estimation. These data are assigned weights in a way similar to the spectral information of satellite images when defining the applied distance metric. The second innovation is that “optimal” weights for spectral data, as well as ancillary data, are computed by means of a genetic algorithm. Tests with practical forest inventory data show that the method performs noticeably better than other applications of k-NN estimation methods in forest inventories, and that the problem of biases in the species volume predictions can for example, almost completely be overcome with this new approach.