Epsilon Archive for Student Projects

Abstract

Landsat TM satellite image data and field data from the National Forest Inventory
(NFI) were combined using the kNN method. Wood volume and basal area
estimations were done for an area in western Sweden. For each pixel in the inventory
area, the distance in feature space was calculated to each reference (NFI) plot. The
estimation for each pixel in the inventory area could be determined by choosing
attribute values for the k closest neighbours in feature space and weighting them with
the inverse squared distances to the estimation pixel. Two distance functions were
used; Mahalanobis distance and prediction difference. The pixels were aggregated
into 296 validation compartments and an average for wood volume and basal area was
derived. Validation data for the compartments was acquired from a field inventory
conducted by STORA Forest and Timber. The main tree species were Scots pine
(Pinus sylvestris), Norway spruce (Picea abies) and birch (Betula spp.). Volume and
basal area could be estimated with a standard error of 21 % by combining satellite
image and NFI data for compartments in the study area that had a volume between
100 and 300
m3 /ha. The standard error was high and wood volume was strongly
overestimated for volumes less then 100 m3 /ha. Wood volume was underestimated by
33 % for volumes greater then 300m3 /ha. Neither correcting of reflectance for
topography nor adding temperature sum improved wood volume estimation accuracy
in the validation compartments. However, wood volume and basal area estimation
accuracy improved if site index, age and mean tree height were included. The
standard error was then 12% for compartments with a volume between 100 and 300
m3
/ha. Using Prediction difference reduced the RMSE, compared to using
Mahalanobis distance, if mean tree height was included as predicting variable.