Brian Schwind, Remote Sensing Specialist

Ralph Warbington, Section Head - Plans and Inventory

Chris Curlis, Image Analyst

Pacific Meridian Resources

In Residence at:

Pacific Southwest Region Remote Sensing Lab

1920 20th Street

Sacramento, CA 95814

ccurlis/r5_rsl@fs.fed.us

Sabrina Daniel, Mapping Ecologist

North State Resources

In Residence at:

Pacific Southwest Region Remote Sensing Lab

1920 20th Street

Sacramento, CA 95814

sdaniel/r5_rsl@fs.fed.us

Since 1988, the Pacific Southwest Region of the Forest Service has had an active and scheduled program for creating and maintaining existing vegetation data layers of National Forest lands. These layers have been developed by the classification and modeling of a variety of remotely sensed and ancillary data. This program has resulted in a library of vegetation layers developed in a consistent manner that are highly suited for Forest and Regional planning, assessments, and monitoring. Additionally, other agencies and private entities have developed vegetation databases to meet a variety of analysis needs outside Forest Service administrative boundaries. With the advent of Federally mandated monitoring of forest conditions in the Northwest, a need arose to combine Forest Service data with comparable data sets on adjacent lands into a spatially uniform layer that met standards recommended by the Regional Ecosystem Office (REO) and Interorganizational Resource Information Coordinating Council (IRICC). These groups have released data standards and implementation recommendations for vegetation information in areas of the Presidents Forest Plan (Interagency Vegetation Information, Data Standards and Implementation Recommendations 1995). In California, the strategy selected for developing this comprehensive existing vegetation database involved the aggregation of the component data sets to a common spatial and thematic level. Subsequently, classification of Landsat TM imagery and GIS modeling techniques were applied to add the remaining floristic and structural detail necessary to meet the defined standards. Ground-based field observations, existing vegetation samples, aerial photography, digital ortho photography, SPOT imagery and field review of draft maps were all used in validating and correcting classification and modeling errors where observed.

The Pacific Northwest has been a major focal point of natural resource issues over the past decade, driving the development and implementation of planning and monitoring policy on Federal Lands. At the forefront of current Federal policy for forestland resources is the Northwest Forest Plan (NFP), mandated by the Clinton Administration in 1993. Critical to the success of the NFP is the development of an effectiveness monitoring program that evaluates the condition and trends of key resources, including Late Successional Old Growth Habitat, Northern Spotted Owl, Marbled Murrelet, and Riparian and Aquatic Resources (Effectiveness Monitoring Team 1997). Direction for monitoring was originally outlined by the Federal Ecosystem Management Assessment Monitoring Team (FEMAT) , defined in an interagency report titled "Interagency Framework for Monitoring the President's Forest Ecosystem Plan" (March 1994). The Regional Interagency Executive Committee (RIEC), part of an Intergovernmental Advisory Council that oversees the management of the NFP, assigned responsibility of monitoring plan development to an interagency Effectiveness Monitoring Team (EMT). The EMT has produced and series of reports laying the scientific foundation of a monitoring program and outlining the resource needs of such a program (EMT 1997). Included in those resource needs are the data necessary to answer questions about important forest resources.

The mandate of an effectiveness monitoring plan has resulted in the need to develop consistent, regional vegetation databases depicting forest landscape conditions across multiple ownerships. In order to meet time frames imposed by the NFP and associated committees, agencies responsible for database development have relied on the classification of digital imagery in conjunction with existing ancillary and ground-based data sets to describe forestland conditions. Many of the significant vegetative characteristics identified by the monitoring plans can be efficiently and cost effectively derived from digital remote sensor data, while vegetative characteristics not effectively derived from remotely sensed data are better captured through plot-based sampling procedures. Statistical relationships between mapped vegetation attributes and sample-based measurements can effectively be used to describe florisitics and the structural condition of vegetated landscapes at regional scales (Miller et al. 1994, Riemann and Alerich 1998).

Development of the Northwest Forest Plan vegetation database occurred in two zones, based on the physiographic provinces of western Washington and Oregon and northwestern California. The Pacific Southwest Region (Region 5) of the USDA Forest Service (USFS) was given lead responsibility for providing the required vegetation data for the nearly 18 million acres of the Northwestern California zone. This paper will focus on the methodologies used by Region 5 to integrate its vegetation mapping efforts with other map products developed within the project area. Over half of this area existed outside the administrated boundaries of the Forest Service and had not previously been mapped under the Region 5 vegetation mapping program or to the NFP interagency standards. Utilizing Region 5 mapping strategies in conjunction with vegetation layers previously developed for lands adjacent to the National Forests, the components for a comprehensive, consistent, and standardized vegetation database were produced in a cost effective and timely manner.

The vegetation data standards for the monitoring database were the result of work accomplished by a Vegetation Strike Team and a Data Coordination Team under the oversight of the Regional Ecosystem Office (REO) and the Interorganizational Resource Information Coordinating Council (IRICC). A Vegetation Strike Team of interagency specialists in remote sensing and vegetation inventory was organized by the IRICC to develop standards and implementation recommendations. The standards were developed from a needs assessment of requirements for watershed analysis as well as bioregional monitoring. Care was taken to make sure vegetation maps developed under these standards could later be crosswalked into the FGDC (Federal Geographic Data Committee) vegetation standards that were simultaneously being developed at the National level. These standards and recommendations were adopted by the REO for vegetation information in areas of the Presidents Plan (Interagency Vegetation Information, Data Standards and Implementation Recommendations, 1995).

Table 1 illustrates the existing vegetation map attributes and data standards defined by the Vegetation Strike Team and Data Standards Team. Standards focus on the tree lifeform type.

Table 1 - IRICC Recommended Data Standards for Existing Vegetation Data

Two categories of ground-based plot data were assembled for this project. The first included data used in the development of the component data sets, existing plot data collected for unrelated purposes, and plot data collected specifically for this project. These data were used to evaluate the consistency of differing map products in areas of overlap and to provide the ecological basis for spatial modeling of vegetation types.

The second category of plot data was obtained from the USFS Region 5 Forest Inventory Analysis (FIA) program (USFS lands) and from the USFS Pacific Northwest Research Station FIA program (non-USFS lands). Collectively, these programs establish a systematic random set of permanent sample points across all ownerships in California. These data inventory detailed vegetation characteristics which are then statistically linked to map labels to aid in the description of mapped attributes. Significant vegetative characteristics, not feasibly captured from remotely sensed data, can be statistically described on a broad continuum. They also serve to inform the map user of floristic and/or structural variance associated with a map label. As map labels impose a narrow definition on what is often a highly varied condition, plot data are necessary to explain variation inherent within the landscape.

The FIA data also served as a valuable and convenient independent reference source for the assessment of map accuracy. All FIA plot data were kept independent of the mapping process in order to avoid autocorrelation between the map-based and sample-based data sets. Map accuracy results generated from methods currently used by the RSL (Milliken et al. 1998) will be published at future date.

The methods used to integrate multiple classifications and map products into a consistent data base were based on the mapping approaches employed at the USFS Region 5 RSL. These methods have been used to develop standardized vegetation databases for all the National Forests in California. Furthermore, Region 5 vegetation data standards closely match those defined by the REO, IRICC, and EMT. Central to this approach was the derivation of stand-based polygons from a single data source and independent of map attributes. These polygons were used to provide both spatial consistency and a means of efficiently fitting multiple data sources to the spatial map standard . Stand delineations or regions, meeting the specified minimum mapping unit (mmu), were systematically derived from Landsat TM imagery. The result was a layer of uniquely identified stands or regions that corresponded to intuitively recognizable landscape patterns. Classification and modeling of thematic attributes were performed separately and hierarchically for each attribute. Landcover features were classified first and subsequently used as a stratification for classification and modeling of more detailed floristic and structural vegetative characteristics. Using the image derived stand layer, each pixel-based thematic layer was then regionalized, creating a stand-based thematic layer. Regionalization was accomplished by evaluating the thematic variance of map class pixels within each unique stand. With a user specified set of parameters, the image analysts controlled the sensitivity of each class in the labeling process. This was an important control measure for evaluating a varied set of vegetation classifications that were developed under differing classification systems.

Independent stand polygons were generated using an image segmentation module developed for the software Image Processing Workbench (Frew, 1990). This module uses a centroid-linkage region growing algorithm for multiple-pass, slow growth region formation within a user specified spectral threshold (Woodcock and Harward 1992). An additional set of auxiliary passes forces polygons to meet the user specified minimum polygon size. Formation of stand polygons was spatially constrained within processing regions derived from Miles and Goudey (1997) ecological section and subsection boundaries. The purpose of this was to 1) establish an ecologically based tiling system within the project area, and 2) limit the number of uniquely identified polygons formed to less than 65,535 (16 bit). Without constraining the number of polygons, an unreasonably large number of unique polygons would have been generated, resulting in exceedingly poor hardware and software performance.

Inputs to the first stage of the image segmentation module were Landsat TM bands 3, 4 and a texture derivative of band 4. Inclusion of the texture band as an input to the segmentation algorithm has been shown to be an important spatial additive (Ryherd and Woodcock 1996). Differences in vegetative characteristics, structure, canopy closure, stand fragmentation, and land use patterns, are often reflected in changes of image texture. These are some of the same characteristics important to the analysis of key forestland resources highlighted by the effectiveness monitoring plans.

In addition to developing a spatially consistent basis for integrating multiple map products, a level of thematic commonality was established. Of the component data sets evaluated for potential integration, all were based on classification systems containing enough floristic detail to achieve a minimum of Anderson level II classes (Anderson et al. 1976) for non-developed landcover types. The Hardwood and Klamath Bioregional Assessment data sets were aggregated to a lifeform level (table 3) to achieve the lowest common denominator of map classes between all data sets. Additional floristic and structural detail contained within these data sets was not directly integrated, used instead as ancillary information to subsequent tree size and density classifications. This was done due to significant differences in classification methods of vegetation type, tree size, and crown density between the three principal data sets. The aggregated classifications were then regionalized against the independently derived stand polygons to create a stand-based lifeform layer. Decision rules for assigning labels to polygons were based on two factors, 1) the lifeform classification logic, and 2) the idiosyncratic (subjective) nature of a given classification. This process required detailed evaluation of the source layers in order to minimize the risk of thematic inconsistency between sources and ultimately in the final database. Lifeform classes crosswalked from the source data were reviewed against both aerial photography and field data to ensure that labeling parameters resulted in attributed polygons consistent with classification logic. A final review of labeled polygons was performed on-screen and anomalous errors were manually edited to the appropriate label. This last step, while labor intensive and costly, was considered necessary for minimizing errors resulting from spectral confusion between map classes. This step also served to reduce the differences between input data sources that remained following the aggregation to the lifeform level .

Table 3 - Classification Logic for Lifeform Classes

Labeled and edited lifeform polygons were used as a stratification for the classification of more detailed vegetation attributes. This hierarchical approach was used to minimize potential error caused by simultaneously mapping forest characteristics that are not mutually exclusive. Tree size, tree canopy closure, and stand structure occur in numerous combinations under varying terrain conditions. Correlating a statistically unique spectral footprint with each of these combinations is unrealistic in the view and experience of the authors. Furthermore, specific techniques have been utilized by the USFS Region 5 RSL to capture tree size and tree density individually (Miller et al. 1994). These techniques were used to ensure consistency of size and density mapping across the project area.

Vegetation Type - Prior to classifying the imagery for tree size and density, vegetation types based on the CALVEG (Classification and Assessment with Landsat of Visible Ecological Groupings, USDA 1981) classification system were spatially modeled within each tree and shrub lifeform. The CALVEG classification system describes dominant vegetation types similar to an alliance level classification. While this level of floristic detail was not required for the regional scale of analysis proposed by the effectiveness monitoring plan, it was captured to 1) meet the level of thematic resolution within the existing USFS vegetation layers, and 2) add an additional level of stratification for tree size and density mapping. Models were developed within ecologically similar units based on observed correlations of dominant vegetation type to terrain variables, principally elevation, slope, and aspect. These terrain variables have been shown to be useful indicators of vegetation associations (Macomber et al. 1991). Matrices depicting the distribution trends of vegetation types across elevation and slope/aspect classes were built. Concurrently, other digital data layers were evaluated for potential to further discriminate among vegetation types. Soils, geology, precipitation, watershed, older vegetation maps, and land use layers were all found to have utility specific to certain vegetation types and geographic modeling zones. Predictive decision rules were then written and subsequent outputs reviewed in the field by mapping technicians and project cooperators. Anomalous map error was manually corrected, on-screen, based on field observations and plot data.

The remaining structural attributes, specifically canopy closure, tree size, and stand structure, were mapped by vegetation type or group of physiologically similar types. The purpose of this was to minimize confusion between spectrally similar CALVEG classes that exhibit differing crown geometries or size profiles.

Tree Crown Closure - Canopy closure was modeled using the Li-Strahler canopy model (Strahler and Li 1981, Li and Strahler 1985). This is a geometric optical model that employs the varied geometry of tree crown structure and four spectral components: sunlit crown, shaded crown, sunlit background and shaded background to estimate canopy characteristics. The model is typically calibrated from field data collected to build regression equations relating crown geometry to stand density. The regressions derive crown geometry ratios for the forest types in the project which are used as inputs into the canopy model. Training stand characteristics were obtained from fieldwork or, where insufficient stands existed for a particular type, photo interpretation. Each type group was evaluated spectrally and plotted in brightness/greenness feature space. Signatures for each type group were estimated by iteratively comparing the fit of model derived components to the relationship of crown and spectral background components of the field measurements. Once the best fit signatures are estimated for each type they are used in conjunction with the remaining model inputs; stand based slope aspect, brightness and greenness images to label each stand by its mean spectral statistics. The Li-Strahler model then inverts this data using the type signatures and imagery characteristics (solar zenith and azimuth) and derives a variety of canopy characteristics. Among the outputs the inversion generates is a continuum of crown closure values for each stand ranging from 0 to 100 percent. Canopy closure outputs were evaluated against plot data and aerial photography to determine if a positive relationship existed between observed and modeled crown closure. Difficulties with the model output were encountered when there was minimal differentiation between the spectral response of tree crowns and background vegetation. This was most evident in low tree density stands with dense montane shrub understories. The tendency has been for the model to over-predict canopy in these conditions. In these cases it was necessary to manipulate the histograms of canopy closure values. Histogram stretches were very effective in improving overall classification accuracy.

Overstory Tree Size - Tree size was classified as a measure of crown width using iterative unsupervised classification of Landsat TM bands 1-5, 7, and a spatial texture image derived from TM band 4. Relationships of crown width to stem diameter have been established for major forest types in California. These relationships were used to infer stem diameter classes from mapped crown width classes for each tree type. Subsequent size classifications were compared to ancillary GIS layers of land use and ownership. Land ownership and corresponding land use were found to be good indicators of tree size for certain vegetation types. Most notably, the National and State Park lands in the coastal redwood belt contained extensive stands of trees greater than 50" dbh. Particular attention was paid to classification results where these unique and significant stand conditions were known to exist. Tree size classification was prepared using the Northwest Forest Plan size scheme of 10" dbh classes and then crosswalked to the Region 5 size classification standards for maximum utility.

Canopy Structure - Canopy structure mapping was based on a predictive modeling approach using floristic and structural characteristics. While Landsat TM data is considered a viable data source for the classification of single and multi-structured stands (Lannom et al. 1998), only mapped features were used for canopy structure mapping. A two class structure attribute was spatially modeled using vegetation type, tree size, and canopy closure as inputs. Existing plot data were used to evaluate trends in forest canopy structure as it relates to tree type, size, and density. Decision rules were written based on these observed trends and applied as a raster-based model. Future structure modeling work will consider Landsat derived textural and/or structural classes as an additional input.

The advent of federally mandated monitoring policy for the forests of the Pacific Northwest has created the need to develop consistent regional scale vegetation databases. Concurrently, numerous vegetation map products have been derived from remotely sensed data to meet a variety of natural resource management information needs. Most often these data sets have been limited in extent or were designed to meet differing map standards. Logically, these products should be evaluated for their potential to be integrated into more extensive planning and monitoring databases. To date, however, little has been done to take advantage of the cost savings and processing efficiency that these data sets potentially provide. Perhaps the risk of thematic and spatial inconsistency that occurs when assembling unrelated map products has prevented data producers from considering such an approach.

This project has sought to take advantage of current data availability while avoiding the incorporation of inconsistencies inherent between independently produced map products. The use of image segmentation to derive a stand-based spatial layer was found to be effective for establishing a uniform foundation for map integration. Map components based on hierarchical classification systems were easily aggregated to a common landcover classification level, preserving core map integrity while enabling the assembly of varied map products. Fitting thematically aggregated map components to a uniform spatial base resulted in a quick and efficient assembly of a project extent landcover classification. The relatively quick production of a uniform and common base layer allowed the authors to allocate more resources towards the classification of required vegetation attributes not consistently available in the component layers.

A note of thanks to all the cooperators, data stewards, and data producers who contributed data, time, and expertise to this project. Particular thanks go to Dr. Larry Fox and Humboldt State University for providing much of the data analyzed in this project. Also, to Mark Rosenberg of the California Dept. of Forestry and Fire Protection for coordinating product reviews and to Hazel Gordon for her modeling and reviewing skills.

Allen, B. H., D. Diaz, and R-5 Regional Advisory Committee. (1986). Draft Ecosystem Classification Handbook. USDA Forest Service. Region 5 Regional Office, San Francisco, CA. 102 pp.

Airola, D. A. (1988). Guide to the California Wildlife Habitat Relationships System. State of California Resources Agency Department of Fish and Game. 74 pp.

Andersop, J.R., Hardy, E.E., Roach, J.T. and Witmer, W.E. (1976). A land use and land cover classification system for use with remote sensor data (USGS Professional Paper 964). Reston, VA: U.S. Geological Survey.

EMT (1997). Effectiveness Monitoring for the Northwest Forest Plan. Final Report by the Effectiveness Monitoring Team. USDA Forest Service, USDI Bureau of Land Management, USDI Fish and Wildlife Service, and USEPA. Portland, OR

Federal Geographic Data Committee. (1997). Content standard for digital geospatial data. Washington, DC. www.fgdc.gov

FEMAT. (1993). Forest Ecosystem Management: an ecological, economic and social assessment. Report of the Forest Ecosystem Management Team. USDA Forest Service, USDC National Marine Fisheries Service, USDI Bureau of Land Management, USDI National Park Service, and USEPA. Portland, OR.

Frew, J.E. Jr. (1990). Image Processing Workbench. Ph.D. Dissertation. University of California, Santa Barbara. Geography. 305 pp.

Jimerson, et. al. (1996). A Field Guide to the Tanoak and the Douglas-fir Plant Associations in Northwestern California. Pacific Southwest Region R5-ECOL-TP-009. 550 pp.

Lannom, K., A.P. Guay, P.E. Hardwick, S. Bain, H. Lachowski, T. Bobbe. (1998) Implementation of Remote Sensing for Ecosystem Management - Recommendations. Natural Resources Management Using Remote Sensing and GIS: Proceedings of the Seventh Forest Service Remote Sensing Applications Conference. 1998. Nassau Bay, TX. pp. 230-239.

Li, X., A.H. Strahler. (1985). Geometric-Optical Modeling of a Conifer Forest Canopy. IEEE Transactions on Geoscience and Remote Sensing. 46(12): 1563-1573.

Macomber, S., C. Woodcock, R. Warbington, K. Casey (1991). Modeling Species Associations for Vegetation Maps Using Terrain Rules. Proceedings of the GRASS 1991 Users Conference. Berkeley, California. 11 pp.

Miles, S.R., C. Goudey. (1997). Ecological Subregions of Califronia - Section and Subsection Descriptions. USDA Forest Service. Pacific Soutwest Region. R-EM-TP-005. San Farncisco, CA.

Miller, S., H. Eng, M. Byrne, J. Milliken, M. Rosenberg. (1994). Northeastern California Vegetation Mapping: A Joint Agency Effort. Remote Sensing and Ecosystem Management: Proceedings of the Fifth Forest Service Remote Sensing Applications Conference. 1994. pp. 115-125. ASPRS, Bethesda, MA.

Milliken, J., D. Beardsley, S. Gill, R. Warbington. (1998). Accuracy Assessment of a Vegetation Map of Northeastern California Using Permanent Plots and Fuzzy Sets. Natural Resources Management Using Remote Sensing and GIS: Proceedings of the Seventh Forest Service Remote Sensing Applications Conference. 1998. Nassau Bay, TX. pp. 218-229.

Riemann, R., D. Alerich. (1998). The Remote Sensing Band - Bringing Together Remote Sensing Application and Research within the Forest Inventory and Analysis Program. Natural Resources Management Using Remote Sensing and GIS: Proceedings of the Seventh Forest Service Remote Sensing Applications Conference. 1998. Nassau Bay, TX. pp. 201-209.

Ryherd, S., C. Woodcock. (1996). Combining Spectral and Texture Data in the Segmentation of Remotely Sensed Images. Photogrammetric Engineering & Remote Sensing. v62(2). pp. 181-194.

Strahler, A.H., X. Li. (1981). An Invertible Coniferous Forest Canopy Reflectance Model. Proceedings of the 15th Int Symp. Remote Sensing Environment. pp. 1237-1244.

USDA, U.S. Forest Service - Regional Ecology Group. (1981) CALVEG: A Classification of California Vegetation. San Francisco, CA. 168 pp.

Woodcock, C., V.J. Harward. (1992). Nested-hierarchical Scene Models and Image Segmentation. International Journal of Remote Sensing. v13(16). pp. 3167-3187.