Epsilon Archive for Student Projects

Abstract

Spatial patterns is of core interest for landscape ecology, and tracking its temporal
evolvement helps to attain a better understanding of the ecological effects of current
ecosystems. The presence and recent digitization of the first Swedish National Forest
Inventory data and the occurrence of a concurrent historical map offers a unique chance to
take on this very challenging task. The objective was to describe and test a statistical model
of the historical forest landscape for a study area in the county of Halland, Sweden during
the 1920s by utilizing different spatial data sources in the model building towards a
plausible methodological application of the model. Data from a detailed digital elevation
model, thematic maps of soil type, and topographic maps were introduced into the
modeling. Both deterministic and stochastic parts of response variables were extracted by
combining Partial Least Squares and logistic regressions with ordinary Kriging.
Conventional cross-validating was applied to judge the performance of the modeling, so
were two compatible estimations, one made by applying k-Nearest Neighbor method, and
the other origins from quartic kernel function. Both of them indicate the performance of
the modeling is fairly good for the vegetation type part (with an accuracy no less than 0.75),
and barely acceptable for the forest stand age and openness part (Root Mean Square Error
is no bigger than 18.45 years for stand age, and 0.17 for stand openness). The modeling
results agree poorly with information extracted from historical map. The nature of the data
available is considered to be the determinant causal factor behind the results. Certainly, the
modeling has the potential to be further improved methodologically.