Pacific Northwest Ecosystem Research Consortium:
Willamette River Basin Mapping Project



Doug R. Oetter, GCSU Assistant Professor of History and Geography
Warren B. Cohen, USDA Forest Service PNW Research Station
Mercedes Berterretche, OSU Forest Science Department
Thomas K. Maiersperger, OSU Forest Science Department
Robert E. Kennedy, OSU Forest Science Department




Overview:

The Pacific Northwest Ecosystem Research Consortium (PNW-ERC), funded by the Environmental Protection Agency, was created "to address specific priority environmental problems in the Northwest, while at the same time developing the ecological understanding and scientific approaches needed to implement ecosystem management on a broad scale." The main goals of the consortium are to:

There are 31 focused projects within the PNW-ERC, addressing a wide variety of ecological research questions over two different ecological provinces. The Willamette River Basin Mapping Project (WRBMP) is an effort to provide a detailed land use/ land cover map of the Willamette River Basin from Landsat Thematic Mapper (TM) data. This map will be used to correlate current land use patterns with existing ecological conditions, and to provide the reference for the generation of future land cover scenarios. In addition, the map will be used by various local planning groups, including the Willamette Valley Livability Forum and the Governor's Watershed Councils.

Previous efforts to map the 29,727 km2 Willamette River Basin include the U. S. Geological Survey's 1974 National Mapping Program (from Landsat MSS and high-altitude photography), Oregon State University's 1976-78 Environmental Remote Sensing Applications Laboratory mapping project for the Oregon Water Resources Department (from Landsat MSS and aerial photography), U. S. Geological Survey's 1992 LULC digital map (from Landsat Thematic Mapper imagery), and Oregon Department of Fish and Wildlife Ecological Applications Laboratory's 1993 Willamette River Basin vegetation inventory (from Color 1:24000 aerial photographs). None of these maps served the consortium's needs, either because the information was outdated or because there were significant land cover types left unmapped. Other science teams within the consortium desired a certain set of land use/ land cover classes, not all of which could be derived from TM data. This class list was based on the Muddy Creek watershed pilot project, which came from interpretation of aerial photographs.

The Willamette River Basin consists of upland forest, lowland forest and other natural cover, agricultural croplands, and urban areas. The upland forest area was mapped as part of a project entitled "Modeling Carbon Dynamics and Their Economic Implications in Two Forested Regions: Pacific Northwestern USA and Northwestern Russia", using a single-date 1988 TM data set. The challenge of the WRBMP was to map lowland classes. For this, we used multi-seasonal 1992 TM data. To interpret the TM data, we constructed a reference data set from a variety of sources, including aerial photographs and Geographic Information System (GIS) coverages. Standard image processing procedures were used to develop the final map, which features 23 classes with an overall accuracy of 73%.

Remote Sensing Data and Preprocessing:

The satellite imagery we selected for this project consisted of five Landsat TM images purchased from the USGS EROS Data Center from 1992 and 1993. Each image was path 46, rows 28, 29, and 30. They were selected to represent one full growing season during 1992, plus a reference from the following year:

Each image was georeferenced to an existing geocoded 1988 TM image mosaic in our library. We used an automated ground control point selection algorithm and a second-order polynomial, nearest neighbor resampling (Kennedy and Cohen 1999). The root mean square error was less than one pixel for all dates.

Radiometric normalization was an important consideration. We used a relative normalization technique, whereby each date was adjusted to a common image. The 7 June 1992 image served as the reference, and the other dates were adjusted to it. First we identified a subset of image pixels containing water, forest, and urban areas which were determined to be no-change pixels throughout all the image dates. The no-change pixels were selected by multiple isodata clustering of a seven-band image (six bands of TM differences plus Band 4 of the reference image). We then subsampled those pixels using a random stratified method to represent equally the three types of land cover. Each band of the uncorrected images was then transformed using a single regression linear equation relating the no-change pixels in the uncorrected image to the same pixels in the reference image.

Given the number of spectral bands across five different dates and the large geographic study area, each image was subset to the first three tassled cap transformation bands to save file space and allow for physical interpretation of band values.

The basin was sub-divided into four mapping units: 1) the forest area previously mapped in the 1988 Western Oregon forest cover project, 2) urban area within the basin as defined by the state's urban growth boundary zoning distinctions, 3) other forest area mainly in the valley floor, and 4) non-forest agricultural and natural areas.


Ground Reference Data:

The ground reference data used to complete this study came primarily from aerial photographs and digital orthophotographs, especially:

The U. S. Forest Service and Bureau of Land Management photographs were used to reference land cover conditions in the forested uplands of the basin, as explained in the Western Oregon Vegetation Mapping Project. The 1988 National High-Altitude Photographs were used solely to identify a small number of testing polygons for the barren (lava) and permanent snow cover classifications. The color digital orthophotograph set, taken from a six mi2 area near Tigard, was used to reference the urban land cover classification. The Spencer Gross digital orthophotographs were used for cross-reference and spot-checking of forest, urban, and agricultural cover types. The 1993 WAC aerial photographs were used to reference forest cover conditions in and along the valley floor. Two hundred and eighty-four forest cover polygons were interpretted for percent canopy closure of conifer and hardwood tree species, particular to oak, orchard, or Christmas trees where discernable. This data set was divided into training and testing sets and used to reference the valley floor forest classification.

The bulk of the non-forest portion of the basin was referenced using the Farm Service Agency 35mm slides. These data are acquired annually by the FSA for their crop compliance program. They are not georeferenced nor subjected to strict photogrammetric requirements, yet proved very useful for the identification of crop types in 1992. Each of 369 slides was scanned into a .tif format at 300 dpi, then imported to Imagine 8.3 and georeferenced to the Landsat TM imagery with a minimum of nine ground control points and an acceptable RMS error of 10m. The slides were then mosaicked into 34 separate study areas, spread throughout all nine counties in the basin, and across all land cover types.

To train our photointerpretation of the FSA slides, five study areas were interpretted for each property tract which participates in the FSA crop compliance program. Our interpretations were then compared to the crop reports filed by the farmers for the study year, which yielded a photographic interpretation error of only 20%. Over 1100 polygons within the FSA study areas were interpretted for 56 different land use/land cover codes. This data set was then divided by class and geographic location into testing and training data sets.

The interpretation of individual plots was sometimes aided by inspection of ancillary GIS coverages, including the Environmental Protection Agency Pudding River riparian assessment, U. S. Fish & Wildlife Service National Wetlands Inventory, U. S. Geological Survey Willamette River Basin pesticide study, Oregon Department of Fish and Wildlife Ecological Application Laboratory natural vegetation inventory, county zoning and taxlot coverages, The Nature Conservancy Willamette River basin wetlands inventory, and Tom Moser's valley field visit database. In addition, inspection of the Landsat TM imagery was used to determine greenness curves throughout the 1992 growing season. Where possible, field visits were conducted during 1998 to examine land use types that may have remained consistent over the six year interval since the photographs were acquired.


Image Processing:

The 29,727 km2 area within the Willamette River Basin was mapped in four stages:

The bulk of the Upland Forest stage was completed as part of the Western Oregon Vegetation Mapping Project. Non-forest lands, such as permanent snow cover and lava fields, were re-mapped in later stages. The continuous predictions of green vegetation cover, conifer cover , and closed conifer age were collapsed to create the classes in the final map.

The Urban Areas were mapped using an isodata clustering (unsupervised) approach on the 15-band tassled cap image (Brightness, Greenness, and Wetness for the five 1992 dates) using four iterations. Originally, eight classes were created in urban areas, however, four of these were later re-examined in the Valley Non-forest stage. For the Valley Forest mapping stage, we created a 15-band principal components image from the five tassled cap images, and used an isodata clustering (unsupervised) method over seven iterations to isolate forest pixels. Once identified, the forest pixels were classed as Closed Conifer, Closed Hardwood, and Closed Mixed using a supervised classification method. The Closed Conifer class was then subdivided into age groups using the 1988 WOV conifer age predictions.

After much experimentation, we classed the remaining Valley Non-Forest pixels by applying a supervised classification method on a 16-band (5 tassled cap dates plus the 30-meter Digital Elevation Model) image. Originally, ten main classes were output, and through many successive revisions, the pixels were further divided into the 26 non-forest classes contained in Version Two (35-class map). Version Three (23-class map) was created by recoding Version Two into 23 more similar classes.

A small portion of the 7 June 1992 image was obscured by cloud and cloud shadow. The mapping of these pixels was accomplished using a 13-band (4 tassled cap dates plus the DEM) image.


Accuracy Assessment:

Error matrixes for both versions are available as Microsoft Excel (v.7) spreadsheets. Each spreadsheet contains two worksheets, one for Majority and one for Point. The Majority matrix assumes that the predicted class for each testing polygon is best represented by the class with the highest number of pixels in that polygon; the Point matrix assigns the predicted class for each polygon to a single pixel within the polygon, identified with the ARC/Info label point of the polygon.


If you should choose to recombine our classes, you should be able to recalculate the map accuracy using the above error matrices. However, if you are interested in subsetting the map area, you may wish to examine the ARC/Info coverages below of the cover and conifer age testing polygons.

For the cover polygon dataset, the observed value is listed as the attribute 'Code' its associated thematic class in the 35-class Version 2 is listed in the attribute 'Value'. The other attributes identify the source of the photo-interpretation and the polygon identification name or number.

The conifer age testing dataset contains the 71 testing polygons we used to assess the accuracy of our continuous age prediction map. Each polygon is identified by the source of the data and the polygon identification name or number. And for each polygon there are three associated ages- the observed mean age of the stand, the model age prediction (obtained by averaging the spectral information for every pixel in the polygon and applying that to the prediction model), and the map age prediction (obtained by averaging the ages of each map pixel within the polygon which has a predicted conifer age, excluding those which were not predicted as having at least 70% green vegetation cover and 70% conifer cover).



References:

Cohen, Warren B.; Spies, Thomas A.; Fiorella, Maria. 1995. Estimating the age and structure of forests in a multi-ownership landscape of western Oregon, U.S.A. International Journal of Remote Sensing 16(4):721-746.

Oetter, Doug R.; Cohen, Warren B.; Berterretche, Mercedes; Maiersperger, Thomas K.; Kennedy, Robert E. 2001. Land cover mapping in an agricultural setting using multi-seasonal Thematic Mapper data. Remote Sensing of Environment 76(2):139-155.

Cohen, Warren B.; Maiersperger, Thomas K.; Spies, Thomas A.; Oetter, Doug R. In Press. Modeling forest cover attributes as continuous variables in a regional context with Thematic Mapper data. International Journal of Remote Sensing.

Cohen, Warren. B.,; Spies, Thomas A.; Alig, Ralph J.; Oetter, Doug R.; Maiersperger, Thomas K.; Fiorella, Maria. In review. Characterizing 23 years (1972-1995) of stand replacement disturbance in western Oregon forests with Landsat imagery. Ecosystems.

Kennedy and Cohen. In preparation. Automated designation of tie-points for multiple-image coregistration.

Contact:

Doug R. Oetter


Download Files: The map is available in two different versions:

Most of the classes are the same in both versions. The only difference between the versions is the further breakout of cover types within Row Crop, Field Crop, and Pasture/Natural. Both versions are available below as Imagine8.3 image files or ARC/Info Grids. Preliminary metadata is available now; the complete metadata set will be posted soon.

To download the image or grid files for Unix, first save the file to your hard drive, then uncompress it using "gunzip <filename.tar.Z>", and then unpack the tar archive using "tar xvf <filename.tar>". To download the files for an NT machine, first save the file to your hard drive, then use WinZip to decompress and extract the directory.


Map Product:

     **Requires Adobe Acrobat v.4**
Land Cover of the Willamette Valley (1992)  Preview  Adobe PDF (19 Mb)

Last updated: 24 May 2001