Kim Lowell

Utilizing discriminant function analysis with a geographical information system to model ecological succession spatially

Abstract A raster/quadtree geographical information system (GIS) was established For a 932 ha natural area in central Missouri. The GIS contained map layers depicting vegetative cover from 1939 to 1982, soils, topographic aspect and distance from a forest seed source. Discriminant function analysis was used to quantify and describe ecological succession on the area during this period. The calibration of discriminant functions is discussed, as are the statistical and spatial validations of the model. It was concluded that this approach provides a useful technique to examine long-term temporal phenomena in a spatial context.

How to test habitat selection at the home range scale: A resampling random windows technique

Abstract Habitat selection studies at the home range scale involve testing whether an animal selects a particular spatial arrangement of habitats within a landscape. We propose a technique based on comparing the landscape pattern indices (LPIs) of observed home ranges with those of a large sample of random windows. In order to determine the appropriate size of the windows, we examined how this variable influences LPIs in a 138-km2 boreal forest block where clearcuts were prevalent. Square random windows of three different sizes (289, 729, and 1156 ha) yielded similar habitat composition proportions but very different values for many of the other LPIs. When compared with values computed for the entire landscape as a single window, most LPIs except composition were very different in the random windows. An example of a comparison between marten (Martes americana Turton) home ranges (n = 11) and square random windows (n = 100) in our study area is given, showing a strong selection for mosaics containing more forest (> 30 years), fewer open regenerating stands, and a larger amount of core area in forest. We conclude that computing LPI values for the entire landscape as a single window is inappropriate for testing habitat selection at the home range scale, and that windows of similar size as the home ranges should be used instead.

Forest attributes and spatial autocorrelation and interpolation: effects of alternative sampling schemata in the boreal forest

Abstract It was not possible to estimate forest volume surfaces having high local precision using samples from three different relatively dense sampling schemata and two spatial interpolation techniques. At best, volume estimates for a given location from area-stealing interpolation or kriging were no more precise than ±80% with 95% confidence even for a forest sample having an intensity of 16%. This precision is only a slight improvement over using the mean as an estimate for each sample location. Although areas of high volume do tend to cluster, i.e. are positively spatially autocorrelated, the forest did not behave as a continuous surface relative even to the densest sampling schema. This suggests that only conducting an a priori assessment of spatial autocorrelation to determine interpolability may be misleading.

Probabilistic temporal GIS modelling involving more than two map classes

Abstract The development of probability-based spatial-temporal models for more than two classes is demonstrated using discriminant function analysis. This represents a significant departure from other probabilistic approaches which have largely been confined to modelling the outcome of a single binary event. It also represents a significant departure from conventional deterministic overlay analysis. Procedures for calibrating and validating such models are presented and discussed. The validation procedures discussed implicate both aspatial consideration, i.e., the amount of each cover type estimated, and spatial factors, i.e., the location of each cover type and the spatial dependence of model residuals (spatial autocorrelation).

Ground-truth verification of relations between forest basal area and certain ecophysiographic factors using a geographic information system

Abstract A geographic information system (GIS) was used to relate measured basal area and cartographically derived ecophysiographic variables in an effort to produce locally reliable estimates of basal area over a forest rather than at individual point locations. The best regression model produced a relatively low R 2 of 0.50 and a root mean square error (RMSE) of 7.3 m 2 per ha. Moreover, the independent variables of this model were terrain-based data—not cartographic information. Thus this model is incapable of producing basal area estimates for all locations in a forest. It is concluded that, because a considerable amount of work needs to be conducted in a number of disciplines before such a model can improve significantly, the approach adopted will not be a viable alternative for producing locally reliable information for a number of years. This work has wider implications as it demonstrates that, while GIS technology has the capability to apply a statistical model to a spatial framework, the results will not necessarily reflect local ground conditions with precision.

Spatial autocorrelation among forest stands identified from the interpretation of aerial photographs

Abstract In an effort to improve on forest inventory regroupment strategies for forest stands in Quebec, spatial adjacency information was obtained for forest stands which had been identified by the interpretation of aerial photographs of a test forest in Quebec, Canada. The Join–Count statistic was used to measure spatial autocorrelation to determine if strata had an unusual affinity for, or repulsion to, other strata. Spatial autocorrelation was measured using a binary contiguity matrix (‘touching/not touching’) and one which used the length of common boundary as the measure of connectedness. Furthermore, basic strata were grouped in four different ways to produce four additional maps of forest strata at different scales of interpretation. In no case was spatial autocorrelation sufficiently strong to allow a satisfactory forest inventory regroupment of strata to be conducted. It is suggested that the reason for this is that the test forest is a managed forest rather than one which ecological processes have been allowed to evolve undisturbed. Operations (e.g. harvesting and planting) on the forest have destroyed the positive spatial autocorrelation expected among natural ecological communities, but have not been conducted long enough to produce significant negative spatial autocorrelation.

Relationships between ground-based forest data and a multiple-interpretation-based certainty map

Abstract The locations of 482 ground-based forest inventory sample plots were overlaid on a map indicating the certainty of the forest type as identified at each location by aerial photo-interpreters. Plots were then stratified into areas of high and low certainty and the means and coefficients of variation for wood volume calculated for each group. Results indicated that the measured wood volumes from sample plots located on areas of high map certainty were higher and less variable than wood volume estimates from sample plots located on areas of low map certainty. It is suggested that sample plots located on areas of high certainty can be used from one forest inventory to another to provide long-term cost-savings.