Transcript waller
Overcoming Chance Agreement in Classification Tree Modeling: Predictor Variables, Training Data, and Spatial Autocorrelation Considerations Eric Waller Colorado Division of Wildlife Southwest Regional GAP Project Arizona, Colorado, Nevada, New Mexico, Utah US-IALE 2004, Las Vegas, Nevada: Transdisciplinary Challenges in Landscape Ecology SWReGAP Approach • Land Cover Mapping: Classification Trees – Satellite Image Classification? – Ecological Modeling? – Hybrid? (Be wary of ancillary variables!) Hybrid Implementation – Iterative: Ecological modeling within spectral (NLCD) strata? – Single Model: “Kitchen Sink” approach? • Captures nuances / avoids error in spectral strata Classification Trees “Need piles of data” Pseudoreplication • Pseudoreplication (multiple sampling within a polygon) - Recommended by EROS Data Center (EDC) Pseudoreplication: Pros and Cons Exacerbates overfitting? – Classification tree strategies • Boosting, Cross-validation / pruning – Rely on independent data – Pseudoreplication = Non-independent data Benefits of additional data? - Swamp anomalous data Autocorrelation • Satellite imagery? • Related to land cover? • DEM-derived variables? • When combined with pseudoreplication, the explanatory power of those predictor variables is inflated. – Even advanced classification tree techniques will be fooled. Chance Agreement Autocorrelation in predictor variables that lack strong explanatory power, when combined with pseudoreplication and a lack of sampling of the range of classes for a given value or combination of those predictor variables, leads to repeated chance agreement between the "bad" predictor variables and the land cover that fools the classification tree model. Sampling - Only one class for a given slope/aspect combination? Slope Mountain mahogany Aspect • Need to sample the range of classes (e.g. ponderosa pine, mountain mahogany, etc.) that occur for any particular combination of predictor variables • Limit predictor variables used in modeling (for those not strongly correlated with land cover) Example Over Golden, Colorado Landsat ETM+ Image Path 34, Row 32 Summer, 2000 DEM and DEM-Derived Variables Elevation Slope Aspect Landform autocorrelated autocorrelated autocorrelated autocorrelated Classification Comparison Foothill Grass Invasive Grass Foothill Shrub Ponderosa Pine Mixed Conifer Wooded Riparian Residential Agriculture Urban Mining Water Classification With DEMDerived Variables Classification Without DEM-Derived Variables Matrix Overlay – Aerial Photography - A matrix can be used to highlight areas of disagreement. - Air photos can be used to resolve discrepancies. Conclusions • This presentation demonstrated a method for dealing with problems associated with using DEM-derived variables in classification tree modeling. • The approach does not guarantee improvement upon a more spectrally derived land cover product. • Future efforts may want to establish a sampling strategy that ensures, a priori, that training data represent the range of classes across landscape variability, especially if DEM-derived variables are to be used.