Predicting tree species composition at pixel-scale

Detailed data on tree species drive models that predict risk of insect and disease mortality in forest stands and simulation models for future stand conditions. Application of such models in geospatial analyses requires these data for millions of remotely sensed pixels. However, the vast majority of remotely sensed thematic maps predict a few categories of stand conditions, such as forest type and stage of stand development. Even if the inherent inaccuracies in remotely sensed predictions are ignored, there is considerable variability in tree composition within each category. There is growing interest in k-Nearest Neighbor (k-NN) imputation as an alternative to supervised classification of remotely sensed data. k-NN starts with a set of training sites j, 10<j<10, within the target geographic area. One attractive set of training sites is the field plots measured by the USDA Forest Service’s Forest Inventory and Analysis program. For each pixel i, 10<i<10, which are outside of the training set, k-NN finds 1≤k training sites that are “close” to the i pixel within a feature space formed from remotely sensed and other geospatial data. Then detailed field measurements from those k training sites are used to impute, or predict, the same type of detailed field data for that i pixel. This imputation is separately repeated for each and every pixel in the full target area. The outcome is detailed predictions of tree species, tree size composition and other field measurements for each pixel. The accuracy of k-NN predictions strongly depends upon the distance metric used to measure “closeness” in this feature space, and there are numerous alternatives for that measure. This paper presents and evaluates a new measure that transforms a high-dimensional remotely-sensed feature space into a new space that is optimized to fit a high-dimensional response space, namely tree-level composition at the pixel scale. The advantages of this approach include highly efficient prediction algorithms for risks to forest health and forecasts of future conditions at the pixel scale.