tds gap

Slide1: 

Mapping plant communities in the Mojave Desert using classification trees (CART): a preliminary assessment of mapping NVCS systems and alliances   Todd D. Sajwaj 1, William G. Kepner 2, and David F. Bradford 2   1 U.S. Army Corps of Engineers, co-located at U.S. EPA, Las Vegas NV 2 U.S. Environmental Protection Agency, Las Vegas NV 2002 National Gap Analysis Program Meeting July 31 – August 4, 2002 Sheperdstown, West Virginia

Slide2: 

Introduction: SW ReGAP Region and the Mojave Map Zone Mojave Map Zone Attributes Size: 45,500 sq. km. Climate: Extremely Arid (4” precip./year) Isolated mountains interspersed with wide basins A range of environments found in the Mojave; strongly linked to elevation and precipitation Outside of Las Vegas, human population is exceedingly sparse

Slide3: 

Mojave Map Zone Spring Mtns. Sheep Mtns Lake Mead City of Evil St. George UT Lake Mojave

Slide4: 

Playa

Slide5: 

Creosotebush – Bursage (Larrea tridentata – Ambrosia dumosa)

Slide6: 

Joshua Trees (Yucca brevifolia)

Slide7: 

Blackbrush (Coleogyne ramosissima)

Slide8: 

Pinyon – Juniper Woodlands (Pinus monophylla – Juniperus osteosperma)

Slide9: 

Ponderosa and White Pine (Pinus ponderosa and P. concolor)

Slide10: 

Methods: Field Data Collection Selection: sites were selected opportunistically in the field based on Uniformity of dominant canopy vegetation Spectral homogeneity from Landsat TM imagery Topographic homogeneity from DEMs Large size (e.g. 90 x 90 meters); riparian sites were captured at smaller sizes Goal: to capture a minimum of 30 sites/community to capture geographic and spectral variability Characterization: Estimates of dominant canopy cover by species (dominant herbaceous, shrub and tree components) Abiotic cover elements (e.g. bedrock, vegetative litter, rock fragments, etc.) Topographic attributes (Slope, Aspect, Elevation, Landform) Delineation: Once a site is identified, it is located and delineated (hand-digitized) on Landsat Thematic Mapper imagery. Documentation: Photographs are collected to aid in site classification within the NVCS. Our sampling methodology was developed to maximize speed in order to collected as much data as possible.

Slide11: 

Methods: Classification System and Site Labeling The National Vegetation Classification System (NVCS) is the basis for classification The “alliance” level of the NVCS was the initial goal of the SW ReGAP landcover maps. Since then, “ecological systems” or “systems” (aggregation of plant communities that occur in similar ecological settings), have become the focus of SW ReGAP. The suggested mapping methodology uses hierarchical classification scheme: Alliance: distinct plant communities based on dominant species and ecological setting “Systems”: aggregations of alliances A coarse “NLCD-like” level (e.g. agriculture, developed, forests, shrublands, herbaceous) Problem 1: The NVCS classification is not complete, nor does it have a systematic methodology for assigning the proper alliance label to a site. Problem 2: units at lower levels do not always fit neatly into a single unit at a higher level Problem 3: Ecological systems are based on 2 aggregation strategies: Compositional groups: alliances found in similar environments and possessing similar spectral signatures Ecological complexes: sets of alliances found in similar environments

Slide12: 

Methods: Digital Geospatial Data Each field site is intersected through a total of 60+ spectral, topographic and edaphic variables. Landsat 7 Enhanced Thematic Mapper (ETM+) Imagery - 3 imagery dates (spring, summer, fall) Raw bands Tasseled Cap (Kauth-Thomas transformation) Fractional Vegetation Texture Digital Elevation Models (DEMs): Slope, aspect, elevation Landform and relative moisture index STATSGO soils data: Soil Texture and Depth Soil Type (soil order and characteristic soil horizons) DAYMET climate data: Derived from weather station point samples interpolated over a coarse DEM Precipitation Average Min./Max. temperature Annual Radiation Map of Nevada Geology (1:250,000)

Slide13: 

Methods: Classification Trees (CART) – General Description Classification trees recursively partition a dataset into increasingly “pure” subsets based on a multitude of predictor variables. A CART starts with a binary split of a single predictor variable that produces the most homogeneous daughter nodes. Splitting stops when an acceptable level of node “purity” has been reached. CARTs tend to produce overfit models that are revised by a process called “pruning”. In the case of SW ReGAP, the “pure” subsets are groups of field sites that belong to the same alliance or system. The predictor variables are the 60+ geospatial data types. The output of a classification tree is a set of decision rules. We attempted a classification at each level NLCD Ecological system Alliance

Slide14: 

Results: Dataset Description – Field Data Summary Field Sites Summary (2750 field sites): ~1100 sites were collected by by NV GAP crews ~450 sites were obtained from the Spring Mountain dataset ~900 sites were acquired from the Nevada Test Site classification ~300 sites were digitized from TM imagery (e.g. agriculture, water) Field sites characterize: 13 NLCD classes (11 modeled) 24 ecological systems (20 modeled) 79 alliances (35 modeled) “Modeled” here means there was a minimum amount of data for a classification tree to TRY to model, it does not mean decision rules were formulated by CART

Slide15: 

Results: NLCD - CART Misclassification error rate: 0.1576 = 342 / 2170 Number of terminal nodes: 67 CART produced decision rules for: Desert Shrublands Montane Shrublands Montane Woodlands/Forests Subalpine Woodlands/Forests Sparse Vegetation Water Important Variables (Ranked): Elevation Spring Tasseled Cap – Wetness Band Spring Red Band Slope Fall Fractional Vegetation

Slide16: 

NLCD Level Classification Overall Accuracy 381/519 = 73.4% Desert Shrublands 270/329 = 82.1% Water 20/21 = 95.2% Montane Shrublands 20/31 = 64.5% Montane Woodlands 30/55 = 54.5% Subalpine Woodlands 4/10 = 40.0% Sparse Vegetation 37/73 = 50.6%

Slide17: 

Results: Systems - CART Important Variable (Ranked): Fall Fractional Vegetation Elevation Spring Tasseled Cap – Soil Bright. Slope Fall Blue Band Misclassification error rate: 0.3503 = 757 / 2161 Number of terminal nodes: 88 CART produced decision rules for: Interior Riparian Shrubland Blackbrush Mojave Desert Wash Montane Forest PJ Woodlands Subalpine Woodlands Sagebrush Saltbush Scrub Creosotebush Clifflands Desert Playas

Slide18: 

Systems Level Classification Overall Accuracy 241/519 = 46.4% Creosote 50/127 = 39.4% Playas 10/21 = 47.6% Saltbush Scrub 31/45 = 68.9% Blackbrush 72/150 = 48.0% PJ Woodlands 22/50 = 44.0% Ponderosa – White Pine 7/19 = 36.8% Bristlecone Pine 2/12 = 16.7%

Slide19: 

Results: Alliances - CART Important Variables (Ranked): Elevation Spring Tasseled Cap – Soil Bright. Slope Summer Fractional Vegetation Fall Texture Middle Infrared Band Misclassification error rate: 0.4402 = 938 / 2131 Number of terminal nodes: 78 CART produced decision rules for 20 alliances including: Creosotebush Blackbrush Joshua Tree Woodlands Big Sagebrush Manzanita PJ Woodlands Bristlecone Pine Woodlands

Slide20: 

Overall Accuracy 107/560 = 19.1% Creosote 46/126 = 36.5% Shadscale Scrub 4/22 = 18.2% Blackbrush 24/90 = 26.7% PJ Woodlands 7/38 = 18.4% Ponderosa – White Pine 3/14 = 21.4% Bristlecone Pine 0/11 = 0.0% Alliance Level Classification

Slide21: 

Preliminary Assessment Maps of NLCD and systems levels are encouraging As the level of ecological detail increases, the degree of accuracy decreases The use of a divide-and-conquer strategy (applied to NLCD classes) will likely require greater amounts of sampling than the methods described here Cross validation “accuracy” is over optimistic; a real accuracy assessment is needed to gauge the success of a map Use of these preliminary models and accuracy assessments as guides to a secondary round of field sampling for the Mojave map zone

Slide22: 

Suggestions Directions for Nevada ReGAP Increased interaction with NatureServe to stratify field site selection To stratify field sampling To clarify ambiguities of the labeling process The use of “random forests” To identify outliers Produce more accurate maps Get field crews collecting data Eastern Great Basin Revisit Mojave map zone

Slide23: 

Parting Shot