The NLCD 2001 is created by partitioning the U.S. into mapping zones. A total of 66 mapping zones were delineated within the conterminous U.S. based on ecoregion and geographical characteristics, edge matching features and the size requirement of Landsat mosaics. Mapping zone 01 encompasses whole or portions of several states, including the state of Washington. Questions about the NLCD mapping zone 01 can be directed to the NLCD 2001 land cover mapping team at the USGS/EROS, Sioux Falls, SD (605) 594-6151 or mrlc@usgs.gov.
Each class accuracy is as follows: (Errors of Omission/Commission) 0 Background (N/A) 1 Unclassified (Cloud, Shadow, etc)(N/A) 2 High Intensity Developed (88%/98%) 3 Low Intensity Developed (99%/89%) 4 Cultivated Land (81%/79%) 5 Grassland (86%/81%) 6 Deciduous Forest (76%/81%) 7 Evergreen Forest (95%/90%) 8 Mixed Forest (80%/76%) 9 Scrub/Shrub (74%/88%) 10 Palustrine Forested Wetland (77%/78%) 11 Palustrine Scrub/Shrub Wetland (67%/70%) 12 Palustrine Emergent Wetland (83%/79%) 13 Estuarine Forested Wetland (N/A) 14 Estuarine Scrub/Shrub Wetland (N/A) 15 Estuarine Emergent Wetland (79%/79%) 16 Unconsolidated Shore (94%/96%) 17 Bare Land (88%/91%) 18 Water (98%/99%) 19 Palustrine Aquatic Bed (N/A) 20 Estuarine Aquatic Bed (100%/96%) 21 Tundra (N/A) 22 Snow/Ice (96%/100%)
The validation points were both collected in the field and photo interpreted. The accuracy assessment selection methods were developed to minimize spatial autocorrelation between the training and accuracy assessment. The first pool of accuracy assessment sites came from field data and photo interpretation of black and white digital orthophotos and digital color infrared imagery (primarily Emerge and Ikonos data). These sites were collected prior to initial mapping and were collected at the same time as the training data. The sites were selected to capture the physical and spectral diversity of the land cover. After these sites were identified, they were separated into training and accuracy assessment sites by imposing a 1 km x 1 km grid over the study area. Accuracy assessment sites could only be selected from alternate 1 km squares. Only 1 sample per class was allowed from each potential square. After the first criteria was met, the accuracy assessment sites were buffered to see if they fell within 1000 meters of another accuracy assessment site of the same class or within 1000 meters of a training site of the same class. Those that fell within the 1000 meter buffer were eliminated. All sites were to be from a homogeneous 3x3 area.
After an analysis of the point distribution, it became clear that there were not enough samples for every class. The remaining points were selected from the initial draft final classification and had to be a homogeneous 3x3 area. A stratified random sample was used to locate sites. These sites were restricted to the same alternate 1 km x 1 km grid that was used to separate training from AA sites in the initial analysis. Sampling was limited to areas where there was high resolution color infrared imagery. The imagery included the previous Ikonos and Emerge imagery, but also included an additional 60 scenes of Ikonos imagery. The additional Ikonos imagery provided sampling areas across the entire study area. When possible, we tried to identify 50 samples of the uncommon classes and 20 sites of the common classes. Samples were selected for the common classes so that there were samples for classes using this methodology.
In total, an additional 637 additional points to the accuracy assessment analysis for a total of 2208. All classes have a minimum of 50 accuracy assessment points except for estuarine aquatic bed and estuarine emergent. These classes have 24 and 29 sites respectively. These classes are limited in the study area and to some extent in the imagery that was available to sample from.
Also as part of the assessment, NOAA staff field tested the classification to determine a subjective goodness of fit.
Post-Processing Steps: None
Known Problems: None
Spatial Filters: None
Conceptually, the descriptive tree is a classification tree generated by using the final minimum-map- unit land cover product (1 acre) as training data, and Landsat and other ancillary data as predictors. The goal of the descriptive tree is to summarize the effects of boosted trees (10 sequential classification trees) into a single condensed decision tree that can be used as a diagnostic tool for the classification process. This descriptive tree can be used to assess the relative importance of each of the input data sets on each land cover class. Such information may also be useful to customize the minimum-mapping-unit classification to meet a user's specific needs through raster modeling. Descriptive trees usually capture 60 to 80% of the information from the original land cover data.
The leaf or terminal nodes of the descriptive tree are assigned to sequential numbers (called node numbers) and mapped across the entire mapping zone on a pixel-by-pixel basis. These node numbers can then be matched with the various conditional statements associated with each respective terminal node. This spatial layer appears similar to a cluster map, but is the result of a supervised classification - not an unsupervised clustering. This node map can potentially be used as input by users to customize NLCD land cover, by linking the spatial extent of an individual node with the rules of the conditional statement.
The Land Cover spatial classification confidence data layer is provided to users to help determine the per-pixel spatial confidence of the NLCD 2001 land cover prediction from the descriptive tree. The C5 algorithm produces an estimate (a value between 0% and 100%) that indicates the confidence of rule predictions at each node based on the training data. This spatial confidence map should be considered as only one indicator of relative reliability of the land cover classification, rather than a precise estimate. Users should be aware that this estimate is made based on only training data, and is derived from a generalized descriptive decision tree that reproduces the final land cover data. However, this layer provides valuable insight for a user to determine the risk or confidence they choose to place in each pixel of land cover.
A logic statement from a descriptive tree classification describes each classification rule for each classified pixel. An example of the logic statement follows:
IF tasseled-cap wetness > 140 and imperviousness = 0 and canopy density < 4, then classify as Water
This logic file can be used in combination with the spatial node map to identify classification logic and allow modifications of the classification based on user's knowledge and/or additional data sets.
Additional information may be found at <http://www.mrlc.gov/mrlc2k_nlcd.asp>.
Summary- This section outlines the classification procedure for the Oregon C-CAP. The three dates of imagery were first reviewed for image quality and shifts between image dates. Training points were used as the dependent variable in a CART (Classification Analysis by Regression Tree) approach. Ancillary data layers were calculated from the TM data and were used as additional independent variables in the analysis. Different versions of the map were produced using different combinations of independent variables. The rough map represented the output from the CART classification routine. Ancillary data were used in spatial models were applied to the rough map to produce the provisional map. This represented a fully automated product. This product was then altered by hand edits to refine the classification. In addition, a percent impervious data layer developed from TM data using high resolution imagery, was imbedded into the classification to define the developed classes. This produced the final-with-edits version which is the final version of the classification and is the one described here.
Pre-processing steps- Each Landsat TM scene was geo-referenced by USGS (United States Geological Survey) EROS. The Space Imaging staff reviewed the spectral and spatial quality of the imagery. Areas that were greater than 1-2 pixels off were sent back to USGS for reprocessing. The data was geo-referenced to Albers Conical Equal Area, with a spheroid of GRS 1980, and Datum of WGS84. The data units is in meters. The Washington TM data was delivered in the form of USGS zone mosaics. The data included three dates of TM: leaf-on, leaf-off, and spring. For each date of TM, spectral and tasseled cap data were received.
Acquisition dates of Landsat ETM+ (TM) scenes used for land cover classification in zone 01 are as follows:
SPRING-
Index 1 for Path 45/Row 26 on 03/21/01 = Scene_ID 5045026000108010
Index 2 for Path 45/Row 27 on 03/26/00 = Scene_ID 7045027000008650
Index 1 for Path 45/Row 28 on 03/21/01 = Scene_ID 5045028000108010
Index 3 for Path 46/Row 26 on 05/07/01 = Scene_ID 7046026000112750
Index 4 for Path 46/Row 27 on 05/31/01 = Scene_ID 5046027000115110
Index 5 for Path 46/Row 28 on 04/10/00 = Scene_ID 5046028000010110
Index 6 for Path 47/Row 26 on 02/13/00 = Scene_ID 5047026000004410
Index 7 for Path 47/Row 27 on 02/26/02 = Scene_ID 7047027000205750
Index 7 for Path 47/Row 28 on 02/26/02 = Scene_ID 7047028000205750
Index 8 for Path 48/Row 26 on 04/03/01 = Scene_ID 7048026000109350
Index 8 for Path 48/Row 27 on 04/03/01 = Scene_ID 7048027000109350
LEAF ON (Summer)-
Index 1 for Path 45/Row 26 on 07/16/00 = Scene_ID 7045026000019850
Index 2 for Path 45/Row 27 on 07/22/02 = Scene_ID 7045027000220350
Index 1 for Path 45/Row 28 on 07/16/00 = Scene_ID 7045028000019850
Index 3 for Path 46/Row 26 on 08/11/01 = Scene_ID 7046026000122350
Index 4 for Path 46/Row 27 on 07/07/00 = Scene_ID 7046027000018950
Index 4 for Path 46/Row 28 on 07/07/00 = Scene_ID 7046028000018950
Index 5 for Path 47/Row 26 on 07/30/00 = Scene_ID 7047026000021250
Index 5 for Path 47/Row 27 on 07/30/00 = Scene_ID 7047027000021250
Index 6 for Path 47/Row 28 on 07/01/01 = Scene_ID 7047028000118250
Index 7 for Path 48/Row 26 on 07/21/00 = Scene_ID 7048026000020350
Index 8 for Path 48/Row 27 on 06/03/00 = Scene_ID 7048027000015550
LEAF-OFF (Fall)-
Index 1 for Path 45/Row 26 on 10/18/99 = Scene_ID 7045026009929150
Index 2 for Path 45/Row 27 on 10/04/00 = Scene_ID 7045027000027850
Index 3 for Path 45/Row 28 on 08/17/00 = Scene_ID 7045028000023050
Index 4 for Path 46/Row 26 on 09/12/01 = Scene_ID 7046026000125550
Index 5 for Path 46/Row 27 on 09/25/00 = Scene_ID 7046027000026950
Index 5 for Path 46/Row 28 on 09/25/00 = Scene_ID 7046028000026950
Index 6 for Path 47/Row 26 on 10/05/01 = Scene_ID 7047026000127850
Index 7 for Path 47/Row 27 on 11/01/99 = Scene_ID 7047027009930550
Index 8 for Path 47/Row 28 on 10/16/99 = Scene_ID 7047028009928950
Index 9 for Path 48/Row 26 on 09/23/00 = Scene_ID 7048026000026750
Index 9 for Path 48/Row 27 on 09/23/00 = Scene_ID 7048027000026750
Field-Collected Data- The goals of the field data collection were to sample the diversity of the landscape, within the classes, and among image dates. Classes that would be more difficult to collect from air photos were targeted for field data collection. To meet these goals, Space Imaging stratified the image into spectral clusters and located the field sites throughout the study area based on these. In addition to these pre-arranged sites, Space Imaging collected points while driving between locations. Due to limited time and accessibility, not all polygons were assessed in the field. Those that we did not visit on the ground were labeled with digital orthophotographs or Emerge data if it was available. Both training and validation points were collected together. See the accuracy assessment section to see how the points were split into training and validation points.
Space Imaging used laptop computers and GPS (Global Positioning System)to correctly locate field points on the TM imagery. Software downloaded from the Minnesota's Department of Natural Resources (DNR)was used to connect the Garmin GPS to the laptop (<http://www.dnr.state.mn.us/mis/gis/tools/arcview/extensions/DNRGarmin/DNRGarmin.html>) computer and ESRI's ArcView software. Space Imaging's programmer developed an ArcView application that allowed entry of location and field notes with a click of the mouse. These data were stored in a shape file.
The items that were collected were- Land Cover characterization Special conditions and remarks Photograph Number Date/time X,Y location
The data and equipment used for the fieldwork are as follows- Ancillary datasets- TIGER 2000 NLCD NWI - mosaicked into zones State road map and Delorme state atlas www.delorme.com
Hardware- Lap-tops with ArcView and data GARMIN GPS modules and external antennae, redundant data cables Cameras Backup devices (Floppy Drives) Extra batteries (lap-top and GPS) Mobile phones System backup CD's with data and software Compass Binoculars Field notebooks with instructions and road maps with pre-determined routes Wetland and Vegetation Field Guides
Imagery- Multi-spectral data for each zone Initial classifications
Classification- After the field points for training were collected, they were combined with photo-interpreted points and used as the dependent variable in a CART classification approach. Many layers tested as independent layers. They included three dates of spectral and tasseled cap imagery, DEM, slope, aspect, texture, band indices (NDVI, Moisture, NDVI-Green), shape indices fractal dimension, compactness, convexity, and form), Census data (housing and population density). Statistical analyses and visual inspection of the output was used to eliminate data that was redundant or not useful in the classification. Additional training points were added to help reduce some of the confusion between classes. The rough classification was created at the end of this process using only the CART discrete decision-tree software. A provisional classification was produced by applying spatial models using ancillary data to the rough classification. The provisional map was then edited using hand editing techniques while using high resolution imagery as reference data. Independently, of this process, Space Imaging produced percent impervious data layers for Washington. This layer was developed from Regression Tree and used impervious classifications from IKONOS imagery to predict pixel level percent impervious at the TM pixel level. The continuous percent impervious data was thresholded to produce the developed categories and imbedded into the final map.
Attributes for this product are as follows- 0 Background 1 Unclassified (Cloud, Shadow, etc) 2 High Intensity Developed 3 Low Intensity Developed 4 Cultivated Land 5 Grassland 6 Deciduous Forest 7 Evergreen Forest 8 Mixed Forest 9 Scrub/Shrub 10 Palustrine Forested Wetland 11 Palustrine Scrub/Shrub Wetland 12 Palustrine Emergent Wetland 13 Estuarine Forested Wetland 14 Estuarine Scrub/Shrub Wetland 15 Estuarine Emergent Wetland 16 Unconsolidated Shore 17 Bare Land 18 Water 19 Palustrine Aquatic Bed 20 Estuarine Aquatic Bed 21 Tundra 22 Snow/Ice
Ancillary Datasets- Non-TM image datasets used are DEM (Digital Elevation Model), slope, aspect, positional index, NWI, NLCD, TIGER2000, field-collected points, photo-interpreted points, Washington (Gap Analysis Program), Census data (housing and population density), Ecoregions, IVMP (Interagency Vegetation Mapping Program), Washington Coastal Atlas, Washington ShoreZone Inventory Data.
QA/QC Process- There were several QA/QC steps involved in the creation of this product. First, there was an internal QA/QC. This was done by viewing the classification frame- by-frame along with the TM imagery, the classification, and high resolution reference imagery. NOAA staff completed a similar review and provided both general and point comments.
Post-Processing Steps- Both Washington and Oregon zones were classified concurrently but independently. When they were completed, they were edgematched to each other.
The completed single pixel product was then generalized to a 1 acre (approximately 5 ETM+ 30 m pixel patch) minimum mapping unit product using a "smart eliminate" algorithm. This aggregation program subsumes pixels from the single pixel level to a 5- pixel patch using a queens algorithm at doubling intervals. The algorithm consults a weighting matrix to guide merging of cover types by similarity, resulting in a product that preserves land cover logic as much as possible.