Predicting forest stand parameters using the k-NN approach

Remote sensing (RS) data is a key source of auxiliary information on forest stand parameters used in forest inventory. Among the modern techniques that allow coupling ground truth and RS derived information, the nonparametric k-Nearest Neighbors (k-NN) approach has been widely used for imputing missing values in forest inventory data. In order to evaluate the effectiveness of the approach using different predictors and input parameters, we used experimental data collected in local forest inventory conducted in middle part of Kyiv region (Ukraine). The data consist of 141 sample plots distributed systematically within the forested area of about 56 km2. Sample trees were selected using two different approaches – using sampling with probability proportional to the tree sizes and stands density. As an auxiliary dataset, we incorporated three sources of spectral information that represents different types of RS: 1) the seasonal mosaics of Landsat 8 OLI imagery; 2) one scene of SPOT 7 image; 3) time series of PlanetScope observations. All spectral data were orthorectified and radiometrically calibrated to top of atmosphere reflectance. We concluded that the accuracy of prediction of basal areas for all types of RS is higher for angle-counting plots. We refer this fact to the technique of tree selection since trees of bigger sizes are likely selected by relascope. These trees reflect parameters of forest overstory scanned by RS sensors. Among used satellite imagery, dense time series of Earth observations showed better agreement between the reference and predicted values of basal area per hectare. Thus, we concluded that pixel size is less important in case of predicting forest parameters compared to the temporal resolution of RS data. According to our findings, the k = 1 implementation of the k-NN method works better when preservation of covariance between variables is important while increasing the k value reduces their ranges. It was found that random forest was the most accurate nearest neighbors search routine for the k-NN method.