Quantification of Landscape Indicators/Aquatic Resource Associations in the Savannah River Basin
Two multivariate analyses, canonical correlation and partial least square (PLS) regression, were conducted to study the relationships between landscape metrics and surface water quality variables. Canonical correlation is well known in biological and ecological studies, but to our knowledge, this is the first use of PLS to explore relationships in ecological data. Although PLS is unfamiliar to many statisticians, it has been used extensively in chemistry to describe relationships between chemical structures and activity. We wanted to test PLS as a potentially frugal method for landscape ecologists, faced with issues of collinearity and small sample size, to explore relationships which may be used to assess the quality and vulnerability of an ecosystem. We ran PLS only for the biota landscape data to permit detailed exploration of the relationships.
In canonical correlation analyses, collinearity, missing observations, multinormality, and the ratio of number of variables to number of observations are important issues that need to be dealt with prior to analysis. PLS is less subject to these constraints. For the canonical correlation analyses, landscape metrics were selected based on pairwise correlation and discriminant analyses resulting in a total of four landscape metrics. For the PLS analyses, all landscape variables (26) were initially considered in the model and the most important ones were included based on their VIP (> 0.8) in the final model. VIP provides information not only as to how important each landscape variable is (e.g. Table 7) but how similar the contribution is to that of other variables. For example, landscape metrics such as Crop_slp_mod, Ag_slp_mod, and Ag_mod had the same VIP as Soil_er (Table 13) indicating equal contributions of soil erodibility and crop/agriculture on slopes with moderately erodible soil in predicting biota. Therefore, for this group of landscape variables, one variable (e.g. soil erodibility) alone may be used in any model.
Area on slope > 3 has the largest canonical coefficient in the land-biota analysis and has the largest VIP in PLS. The only other landscape variable common to canonical correlation and the initial PLS model is Pct_bar which is not greatly important in either method. Pct_bar did not make it to the final model because of its low VIP value. In contrast to the canonical model, the PLS model indicated that landscape metrics such as stream density, percent forest and agriculture on moderately erodible soil were the second important landscape variable group. The other intriguing feature, PLS allowed analysis by ecoregion which revealed different landscape metrics that relate to surface water biota based on their spatial association (ecoregion, Table 15). It is evident that the importance of any landscape metric is a function of its spatial location in the study area. The importance of percent forest and percent pasture were the highest in Blue Ridge and decreased consistently across the adjacent ecoregions of the study area; the opposite relationship was found for Crop_slp and soil erodibility. Crops on area with slope > 3% with moderately erodible soil, percent crop, percent of areas on slopes > 3% and soil erodibility are the most important landscape variables in the Piedmont. The relative importance of percent pasture on slopes > 3% increased in the Coastal Plain.
Multivariate Analyses (Canonical Correlation and Partial Least Square, PLS) to Model and Assess the Association of Landscape Metrics to Surface Water Chemical and Biological Properties using Savannah River Basin Data.(4671 KB)