Presentation on theme: "The West Cascades Park City The West Cascades NaFISNationwide Forest Imputation Study."— Presentation transcript:
The West Cascades Park City The West Cascades NaFISNationwide Forest Imputation Study
Gradients in Plant Community Ecology Plant species exhibit distributional patterns that are a reflection of changing environmental conditions.
Hierarchies from landscape ecology “… a system of interconnections wherein the higher levels constrain the lower levels to various degrees...” ( Turner et al. 2001) Broad-scale, factors (e.g., climate) constrain local species pools. Local topography, disturbance, succession and competition determine which species from that pool occupy a given site. Time > Spatial Extent > CLIMATE Disturbance Local Topography
Objective Explore vegetation-environment relations in the context of imputation mapping. Different modeling techniques make different assumptions about the world. –Euclidean Nearest Neighbor. –Gradient Nearest Neighbor. –Random Forest.
Methods –Maps built from: –784 records from our plot database (FIA annual plots) –and 16 mapped explanatory variables. LandsatBands 3,4,5 ClimatePRISM: Means, seasonal variability TopographyElevation, slope, aspect, solar LocationX, Y
study area (2) Place new pixel within feature space (3) find nearest- neighbor plot within feature space (4) impute nearest neighbor’s value to pixel Methods: Euclidean Nearest Neighbor Imputation feature spacegeographic space Elevation Rainfall (1) Place plots within feature space
Advantages –Simplicity. –Quick to run. –Makes no assumptions about how vegetation relates to the environment. Disadvantages –May not represent species-environment relations well. Pros and Cons: Euclidean Imputation
(2) calculate axis scores of pixel from mapped data layers study area (3) find nearest- neighbor plot in gradient space (4) impute nearest neighbor’s value to pixel Methods: Gradient Nearest Neighbor Imputation gradient spacegeographic space CCA Axis 2 (e.g., Temperature, Elevation) CCA Axis 1 (e.g., Rainfall, local topography) (1) conduct gradient analysis of plot data
Advantages to GNN –Shapes environmental space as it relates to forest composition. –Model structure is straightforward, reasonably intuitive. Disadvantage to GNN –Assumes that species show a unimodal response to environmental gradients (Gauch, 1982; ter Braak and Prentice, 1988). Pros and Cons: Gradient Nearest Neighbor Imputation
study area Methods: Random Forest Nearest Neighbor Imputation Random Forest spacegeographic space
Methods: Random Forest One Classification Tree: Elevation < 1244 August Maximum < Temp August Maximum < Temp Summer Mean < Temp Aug. to Dec. Temperature < Differential Elevation < 1625 LANDSAT Band 5 < 24 PSME TSHE PSMETHPL ABAM TSME PSME PIPO High Elevation ( > 1244) High August Temp (> 23.24°C) High reflectance in Band 5 (> 24)
Methods: Random Forest A “Forest” of classification trees. Each tree is built from a random subset of plots and variables.
Methods: Random Forest Imputation
Advantages –Models vegetation-environment relations –Free from distributional assumptions –High accuracy Disadvantages –Computing time –Interpretation is difficult Pros and Cons: Random Forest Imputation
Comparisons Root Mean Square Difference (RMSD) for species basal area. Mapped distribution (presence/absence) –Douglas-fir (Pseudotsuga menziezii) –Sugar Pine (Pinus lambertiana)
Accuracy (scaled RMSD) GNN Advantage Random Forest Advantage Methods Equally Good Euclidean model not shown. Results were comparable, but never best.
Douglas-fir Often dominant. Widespread, early colonizer, long-lived. Only disappears at v. high elevations.
Douglas-fir Range Euclidean 78.0% GNN 59.2% Random Forest 74.3% Estimated Actual Area 79.77%
Douglas-fir Range Euclidean GNN Random Forest (scaled)
Sugar Pine Spotty distribution, wide elevation range, mostly in the South.
Sugar Pine Range Euclidean 5.4% GNN 3.5% Random Forest 4.3% Estimated Actual Area 4.6%
Sugar Pine Range Euclidean GNN Random Forest (scaled)
Conclusions The answer is... YES!! –The world can be seen as a gradient. –But in some cases, the world is better described by a hierarchy.
Conclusions: Which model? Broad-scale patterns are consistently predicted by all 3 model types. GNN works well most of the time. If rare, or quirky species are our focus, however, Random Forest may provide a very useful alternative. Both Random Forest and GNN are an improvement over simple euclidean imputation in terms of RMSD, but euclidean was often less biased in the range-maps.