Open access publications in the SLU publication database

Abstract

Forest information for management planning is today gathered through a combination of field inventories and remote sensing, but the available flow of remote sensing data over time is not yet utilized for continuously improving predictions of forest variables. In the thesis, the utility of data assimilation, in particular the Extended Kalman filter, for forest variable prediction is investigated. This is an iterative algorithm, where data are repeatedly merged and forecasted.

The test site was a forest estate in southern Sweden (Lat. 58°N Long. 13°E). Data assimilation of remote sensing predictions of canopy surface models from digital aerial photogrammetry in paper I and predictions based on interferometric synthetic aperture radar in paper II provided a marginally improved accuracy. This gain was, however, far from the theoretical potential of data assimilation. The reason for this was suggested to be correlation of errors of subsequent predictions across time, i.e. residuals from different predictions over a certain forest area had a similar size and sign. In paper III these error correlations were quantified, and an example of the importance of considering them was given. In paper IV, it was shown that classical calibration could be applied to counteract regression toward the mean, and thus reduce the error correlations. In paper V, it was shown that data assimilation applied to a time series of data from various remote sensing sensors could be used to, over time, improve initial predictions based on aerial laser scanning data. It was also shown how the combination of classical calibration and a suggested modified version of the extended Kalman filter, that accounted for error correlations, contributed to these promising results.