VMap_Base

email preferred

<15 meters

These are the four channel NAIP image data.

These are the six channel TM image data that have been calibrated to exo-atmospheric reflectance to account for between scene variation in sun angle and solar elevation.  These data provided the base imagery for the image segmentation, vegetation indices, and Kauth-Thomas Tassel-Cap transformations.  This was the "peak greenness" imagery, upon which, the change detection was based.

The TC is a linear transformation of the reflectance calculated TM data that rotates the data structure such that the majority of the information contained in the 6 bands will occupy 3 dimensions that are directly related to the on-the-ground physical scene characteristics.  These dimensions define planes of soils (brightness), vegetation (greenness), and a transitional zone that relates to canopy and soil moisture (wetness).  These three dimensions capture 97%+ of the data variation in the 6 TM bands and can enable the discernment of key forest attributes (i.e., species, age, and structure).

The Normalized Difference Vegetation Index (NDVI) is calculated as the normalized difference between the NIR and the Red bands (NIR - R)/(NIR + R). The NDVI is probably the most widely used vegetation index and has been shown to be related to a number of different biomass variables. Simple vegetation indices such as NDVI, however, provide an inadequate representation of complex vegetation cover as they are related only to the total amount of above-ground green leaf biomass, and give no indication of the types of vegetation present. Vegetated areas will generally yield a higher NDVI value than rock, which will have values greater than that of clouds, snow, and water.  The 5meter NAIP imagery was used for the NDVI where applicable.   In other cases, NDVI was calculated for the Landsat TM imagery was used.

TRASP was used as a biophysical predictor to create a Whitebark pine probability surface for input into eCognition for the Gallatin NF portion of the VMap and also for some of the non-forest classes (see Random Forest.)  The circular aspect variable is transformed to a radiation index (TRASP.)  This transformation assigns a value of zero to land oriented in a north-northeast direction, (typically the coolest and wettest orientation), and a value of one on the hotter, dryer south-southwesterly slopes. The result is a continuous variable between 0 - 1 (Robert and Cooper 1989).TRASP=(1 - cos((pi / 180)(aspect - 30)))/2

The Compound Topographic Index (CTI) was used as a biophysical predictor to create a Whitebark pine probability surface for input into eCognition for the Gallatin NF portion of the VMap and also for some of the non-forest classes (see Random Forest.)  CTI is a steady state wetness index. The CTI is a function of both the slope and the upstream contributing area per unit width orthigonal to the flow direction. CTI was desigined for hillslope catenas. Accumulation numbers in flat areas will be very large and CTI will not be a relevant variable. CTI is highly correlated with several soil attributes such as horizon depth(r=0.55), silt percentage(r=0.61), organic matter content(r=0.57), and phosphorus(r=0.53) (Moore et al. 1993).The implementation of CTI can be shown as: CTI = ln (As / (tan (beta)) where As = Area Value calculated as(flow accumulation + 1 ) * (pixel area in m2)and beta is the slope expressed in radians.  The ArcInfo approach to calculating Flow Direction uses the D8 algorithm producing very unrealistic results. Several other methods are available for calculating flow directions. One of the more robust approaches is the D infinity algorithm (Tarboton 1997). There is a freeware download and documentation for TARDEM http://www.engineering.usu.edu/dtarb/ or TAUDEM http://moose.cee.usu.edu/taudem/taudem.html.  The derivative of the FLOWDIRECTION calculation from either of these two programs can be used in the CTI AML or you can calculate FLOWDIRECTION in GRID.

This layer contains the elevation information for each sub-path model. Illumination, slope and aspect were derived from the DEM.

Derived from the 10 meter DEM using ArcGISs Spatial Analyst function. The raster created is the global radiation or total amount of incoming solar insolation (direct and diffuse) calculated for each location of the input DEM for one year.   The output has unit watt hours per square meter (WH/m2).  This surface was used as a biophysical predictor to create a Whitebark pine probability surface for input into eCognition for the Gallatin NF portion of the VMap and also for some of the non-forest classes (see Random Forest).

In machine learning, a random forest is a classifier that consists of many decision trees and outputs the class that is the mode of the classes output by individual trees. The algorithm for inducing a random forest was developed by Leo Breiman and Adele Cutler, and "Random Forests" is their trademark. The term came from random decision forests that was first proposed by Tin Kam Ho of Bell Labs in 1995. The method combines Breiman's "bagging" idea and Ho's "random subspace method" to construct a collection of decision trees with controlled variations.  The Random Forests classifier is part of the open source statistical package R.  The USDA Remote Sensing Applications Center created a CartTools Pythogn program script (contact Bonnie Ruefenacht for the latest version of this script 801-975-3828) which accesses Rs Random Forest to create prediction surfaces. This script was used to create additional non-forest class predictions.  The Mid-level non-forest classes predicted with Random Forests include grass-bunch, grass-single-stem; the litter classes litter>90%, litter 60-89.9%, litter < 60%; and the xeric-shrub canopy classes xeric shrub10-24.9%, and xeric shrub > 25%.

In statistics, the generalized additive model (or GAM) is a statistical model developed by Trevor Hastie and Rob Tibshirani for blending properties of generalized linear models with additive models. This surface was used as a biophysical predictor to create a Whitebark pine probability surface for input into eCognition for the Gallatin NF portion of the VMap. The Whitebark probability surface was created using the statistical package R version 2.72 with a General Additive Model (GAM).

These layers are transformations of the DEM derivatives of aspect and percent slope that combines the information into single files for east/west (ewslp) and north/south (nsslp), respectively. These data have had a zonal majority calculated based on the z-grid for each sub-path model so that there is a unique value retained for each image object.

A series of calculations of texture were created from the color infrared NAIP imagery and used in the eCognition segmentation and map classification. Texture calculates a variance (minimum, mean) from an adaptive window around each pixel as its measure of texture. The resulting texture image or band is a composite of minimum variance values calculated for each pixel.  Two sets of three banded texture images were created using these focal windows and parameters:  The first three banded image was created from 1m NAIP using a minimum variance and focal windows of (3x3), (5X5), and (9X9), then resampled back to 5meters; the second three banded texture image was created from 5m NAIP using a mean variance and focal windows of (3X3), (5X5) and (9x9).

VMap models for processing are based on general ecological or management units.  They are restricted in size for better mapping precision and also to keep within the size limit restrictions of eCognition software.

Through the segmentation process, eCognition calculates a number of transformations and derivatives on the input data layers. For each image object a mean and standard deviation, of the values of the pixels contained within the boundaries that image object, is calculated for each input data layer. There are also three (3) eCognition features that are calculated on a subset of the input data layers, in this case the six (6) (CH 1-5, 7) channels of the "peak greenness" TM scene and three (3) (CH 1-3) of the color infrared NAIP imagery. The first feature is termed "BRIGHTNESS", which is the sum of the mean values of those six (6) layers divided by their quantity computed for an image object. The second feature is "Maximim Difference" (MAX DIFF.) , which is the sum of the mean values of those six (6) layers minus (-) the derived BRIGHTNESS value.

This is the photogrammetrically interepreted and ground survey based "ground truth" data employed to model the classification membership functions and to drive the Nearest Neighbor analysis.

The path-level LandSAT 5 TM data were ortho-rectified to the 1 meter color infrared NAIP imagery 2009 using the Ortho-Rectification Module and the Landsat orbit model in ERDAS Imagine 9.2 as well as 10 meter digital elevation models. A minimum of at least 50 ground control points (GCP) throughout each of the unrectified images. Actual rectification involved the Cubic convolution algorithm and a 30m pixel size. The resulting Root Mean Square (RMS) error was less than one pixel or 30 m.

Path-level data subset to 13 map area regions based on the dissolved boundaries. Resample the 30m path level TM data to 5m resolution using ERDAS Imagine Cubic Convolution resampling technique.

Calculate minimum and mean texture images from NAIP imagery for each sub-model.

Create an eCognition "project" using the source data layers. Load the 12 TM layers, 12 tassel cap transformations and derivatives for the fall Landsat scenes, 12 Principal Component analysis 6 texture image derivatives the DEM, pnv layer, NDVI image, subpath model mask, illumination mask, and the aspect/slope combination layer.

Course level multiresolution segmentation using color infrared 5 meter NAIP imagery and texture band in eCognition software. Multiresolution segmentation is essentially a heuristic optimization procedure, which locally minimizes the average heterogeneity of image objects for a given resolution over the whole scene. Multiresolution segmentation is a method of generating image objects. It produces highly homogeneous segments in any chosen resolution, fitting your purpose. The resulting image segmentation, whose individual elements are referred to as image objects, can be universally applied to almost all data types. The image objects themselves, contain feature information based on the values of the pixels contained within the borders of each image object. These image object values are then used in the classification process, either through the use of fuzzy logic based membership functions or a Nearest neighbor analysis.

Sample data for each class in the classification schema is then loaded, or digitized, into the eCognition project. eCognition will then return a histogram for each feature and each sample base, which displays the spectral distribution of the samples over the range of the data feature chosen.

Compute a hierearchical classification based on fuzzy membership values calculated using the image object values computed through the multiresolution segmentation. The classification scheme first divides out water from non-water; then vegetated from non-vegetated; tree from herbaceous and shrub; shrub from herbaceous; 4 tree canopy cover classes from the tree dominated lifeform; 4 tree size classes from the tree dominated lifeform (using Nearest Neighbor analysis); and 8-12 dominance types (using a Nearest Neighbor analysis).

Compute classification of the Dry Grass type into two grass types (Bunch Grass and Single Stem Grass) using Random Forest alogrithm with field sample data and NAIP, Landsat, and Topographic variables.

Compute classification of the Dry Shrub type into two canopy cover classes using Random Forest alogrithm with field sample data and NAIP, Landsat, and Topographic variables.

Compute classification of Grass and Shrub lifeform types into two ground litter cover classes using Random Forest alogrithm with field sample data and NAIP, Landsat, and Topographic variables.

Final forest base level database was created by combining the 13 map area databases together and two map area databases from the 2009 Vmap m2901 and m2902 datasets.

Dissolved versions of the forest wide base level database were created for lifeform, tree canopy cover, tree diameter, and tree dominance type.  A dissolve of tree canopy cover, tree diameter, tree dominance type combined was also created.

An accuracy assessment was provided for the four primary map products to quantify accuracy following four distinct lines of analytical logic. VMap accuracy assessment data are those polygons associated with the Forest Inventory Analysis (FIA) plot data.   Summaries of the FIA plot data provide a means to achieve the most reliable dominance type and size determinations for each assessment reference polygon and can assist to some degree with canopy cover. The VMap accuracy assessment includes an area-weighted error matrix, which is based on the aerial extent of each class. The nature of errors in the classified map can, thus, be derived from the error matrix. A relatively recent innovation in accuracy assessment is the use of fuzzy sets for accuracy assessments. Fuzzy logic is designed to handle ambiguity and, therefore, constitutes the basis for part of the VMap accuracy assessment. Instead of assessing a site as correct/incorrect as in a traditional assessment, an assessment using fuzzy sets can rate a site as absolutely wrong, reasonable or acceptable match, good match, or absolutely right. The resulting accuracy assessment can then rate the seriousness of errors as well as absolute correctness/incorrectness. For these reasons, the VMap accuracy assessment includes a fuzzy set-based error matrix and an area-weighted fuzzy set-based error matrix.

Metadata imported.

Metadata imported.

Metadata imported.

Metadata imported.

email preferred