Open access publications in the SLU publication database

Abstract

Modern technology such as the Global Positioning System (GPS) and Geographical Information Systems (GIS) provide new opportunities for forest inventory. These technologies allow representation of forest variables using rasters with cell sizes on the order of 25 m. Such rasters can be estimated from remotely sensed data using models of the relationship between the image’s digital number and the forest variables. This thesis investigates the possibility of using estimation methods incorporating remotely sensed data as well as spatial similarity of neighbouring field measurements, to improve prediction accuracy compared to using only remotely sensed data. Two new spatial prediction methods are presented and evaluated: ordinary kriging using information about edges detected in remotely sensed images, and prediction using Markov Chain Monte Carlo (MCMC) simulation of a new Bayesian state-space model. In addition, ordinary kriging, stratified ordinary kriging, ordinary cokriging, collocated ordinary cokriging, simple kriging with varying local means, and spatial regression using the autoregressive response model, are also evaluated. The methods are applied to predict forest stem volume per hectare in boreal forest in northern Sweden (Lat. 64°14’N, Long. 19°40’E) using Landsat TM data and a large field sampled dataset. Prediction accuracy, as well as practical aspects of the methods, is evaluated. In particular, accuracy is compared with Ordinary Least Squares regression (OLS) using remotely sensed data. Spatial prediction was, with a few exceptions, more accurate than OLS regression. The largest improvement, 49% lower root mean square error (RMSE), was obtained for plot-level predictions by ordinary kriging using information of edges detected in remotely sensed images, although the method is dependent on densely sampled field data. Promising results were also obtained by simple kriging with varying local means. This method performed well (26% lower RMSE than OLS regression for stand-level predictions), is rather straight-forward to apply in practice, and not as dependent on densely sampled field data. The Bayesian state-space model did not provide improved predictions compared to OLS regression. However, Bayesian modelling is promising for application of spatial models of higher complexity than possible with the other methods.