Forest attribution using K-NN methods with Landsat 8 imagery and forest field plots

This study presents an approach to integrate Landsat satellite imagery and forest monitoring field plots, to produce forest attribute maps across the state of Victoria, Australia. Over 450 field plots, sampled from a stratified systematic random framework, were measured to characterise woody and non-woody forest attributes. Field plot data were applied in various k-NN procedures using Landsat8 data to map biomass, stems per hectare and species diversity. The study investigated four k-NN distance metrics (Mahalanobis Nearest Neighbour, Most Similar Neighbour, Gradient Nearest Neighbour and Random Forest Nearest Neighbour), as well as five k values representing number of Neighbours used in the prediction model (1, 2, 5, 10 and 20). Model accuracy was assessed in various dimensions, including plot (root mean square difference) and regional level (Area comparison of design-based (plots) vs. model-based (map) estimates). Results are used to provide guidance for implementing k-NN in a large area operational setting.