banner literature about projects tools meetings search links BBS GAP home GAP home
home

 

Previous | First | Next

LAND COVER

Collection and Processing of a High Volume of Field Data for
the Missouri Land Use/Land Cover Mapping Project

DAVID D. DIAMOND, TAISIA M. GORDON, AND KAN HE

Missouri Resource Assessment Partnership (MoRAP), University of Missouri, Columbia

Introduction

The Missouri Resource Assessment Partnership (MoRAP), a cooperative of natural resource
agencies, mapped land cover for Missouri using Thematic Mapper (TM) satellite data. We
developed a field sampling scheme based on spatial segmentation of individual TM scenes by
ecological subsection that was designed to derive the maximum relevant information possible.
Incoming data quality varied due to the diversity of individuals involved, but the spatially
segmented sampling scheme, large data volume, and final expert review ensured that the biases
of any individual had negligible impact on the final land cover product. In order to efficiently
process the large volume of incoming data, we developed the Missouri Land Cover Ground
Verification Tool–a Microsoft Visual Basic application with data management, analysis, and
mapping functions that bridges Microsoft Access and ESRI's ARC/INFO software.

Classification Approach

An unsupervised multispectral image classification for Landsat TM data was used as the core
land cover mapping approach. We first merged two dates for each of 15 scenes and used three
bands from each date to derive 60 spectral classes per scene using ISODATA (Tou and Gonzalez
1977, Jensen 1996). MoRAP then produced two distinct versions of Missouri land cover based
on the level of ground verification. Phase I involved the assignment of 60 spectral classes per
scene to land cover classes using NAP photos and ecological expertise. Phase II incorporates the
information from more than 11,500 on-site visits to polygons that represent large clusters of 30-
meter pixels within different, individual spectral classes. Whereas these two stages are logically
connected and are mentioned for context consistency, this paper will emphasize the details of the
field verification process.

Sampling Design and Preparation of Field Materials

Efficient collection and management of accurate on-site land use/land cover information is
critical to the success of a land cover classification from remotely sensed data. MoRAP designed
an ambitious sampling scheme with the goal of collecting several on-site data points for each
spectral class derived from each TM scene within every ecological subsection that the scene
intersected. Ecological subsections were used because land cover is more homogeneous within
versus among subsections.

The on-site field data collection effort involved more than 150 individuals, and we planned to get
information from more than 11,500 polygons within one year. Hence, we developed a set of  field materials that were easy to understand and interpret and were relatively easy to produce.
After a trial run, we settled on a final set of materials consisting of:
1. the Missouri Phase II Land Cover Classification Scheme;
2. a field data recording sheet (Table 1);
3. a brief explanation of the materials and instructions for filling out the data form;
4. a Land Cover Quadrangle Map with numbered polygons (Figure 1); and
5. a USGS Quadrangle Map with numbered polygons (Figure 2).

Development of field materials for ground verification involved three major steps. First, we
clipped each TM scene by ecological subsection or, in some cases, land type association (Bailey
et al. 1994, Keys 1995, Schroeder et al. 1998) (Figure 3).

Figure 3A. A Classifed TM scene overlaid with ecological subsection polygons.

Figure 3B. Part of the classified TM scene clipped by ecological subesction.

In this way, Missouri was spatially subdivided into 38 major ecological subunit polygons. Second, we selected target USGS 7.5-minute quadrangles based on (1) their spatial position in relation to the boundary of TM scenes and ecological polygons, and (2) visual inspection of the diversity of spectral clusters within the quadrangle. Diverse quadrangles were selected over more uniform ones, and quadrangles entirely within a single TM scene or ecological polygon were selected over those that were not. Finally, we created and numbered sampling polygons that represented large spatial clusters of 30-meter pixels of uniform ISODATA spectral classes (Figure 4). 

Figure 4. A polygon is created around pixels representing the same spectral class. The X and Y coordinates (circled) of the center of the polygon are recorded for data analysis.

The minimum sampling polygon was approximately one hectare (11 pixels), and we attempted to choose polygons that were accessible via public roads or public lands. These numbered polygons were in turn outlined on two maps based on the boundary of a USGS 7.5-minute quadrangle, one with classified land cover and one Digital Raster Graph of the same quadrangle. These maps were then printed from a large-format plotter. Field personnel used the familiar Digital Raster Graph to navigate to sites, which had clearly marked and numbered polygons to visit. The numbers on the polygons (Figure 1 and 2) corresponded with numbers on the field data collection form (Table 1).

Table 1. Field data recording sheet

Principal Data Collector: __________________ Contact Phone Numbers: __________________
Organization: _______________________ Collecting Date: ___________________

USGS 1:24,000 Quadrangle Name and location : Tiff Quadrangle, Washington & Jefferson Counties, Missouri

Site ID Land Cover Type on map Verification
Method*
Ground verified classification and site description
01 

Deciduous

Forest

Agree? Y / N

site visit___
windshield ___
experience___
other______
More Detailed or Alternative
Class(es)**_________________________
Dominant Species/Structure___________________________

Changes since 1991-1993?____________________________

Other comments:
02

Eastern Red

Cedar Forest

Agree? Y / N

site visit___
windshield ___
experience___
other_______
More Detailed or Alternative
Class(es)**_________________________
Dominant Species/Structure___________________________

Changes since 1991-1993?____________________________

Other comments:
03

Eastern Red

Cedar Forest

Agree? Y / N

site visit___
windshield ___
experience___
other_______

More Detailed or Alternative
Class(es)**_________________________
Dominant Species/Structure___________________________

Changes since 1991-1993?____________________________

Other comments:

*Please check one. A field level observation of the site is highly preferred, but close inspection of the entire site is not
needed. If you cannot gain access, but can see the site remotely or have experience on-site, please indicate.
**From Phase II Land Cover Scheme.

Capture and Processing of Incoming Data

We developed a Microsoft Access database designed to facilitate initial data entry and analysis.
The Access database contains 21 fields that uniquely identify and describe each polygon visited
on the ground (Figure 5). Among those 21 fields, 15 have one-to-one relationships between the
record and record item, whereas six have a one-to-many relationship. For the sake of speed and
efficiency, we divided the database into four linked tables. The master table contains all one-to-
one relationship fields, and three ancillary tables hold fields with one-to-many relationships.

Figure 5. Microsoft Access tables demonstrating one-to-one and one-to-many relationships.

We developed a Microsoft Visual Basic application, the Missouri Land Cover Verification Tool
(Figure 6). 

Figure 6. Initial interface of Missouri Land Cover Ground Verification Tool.

Visual Basic's built-in Data Access Object (DAO) allows easy management of the
Access database (Microsoft Corporation 1998). ARC/INFO's Open Development Environment
(ODE) allows access to GIS functionality through a custom-created Visual Basic interface (Potts
1998). Through the data-centric application development environment of Visual Basic, the
Missouri Land Cover Verification Tool provides fast, comprehensive data access and data
analysis, visualization, and mapping capabilities. Hence, the overall complexity of data
management and analysis is much reduced from using Access and ARC/INFO independently.
Integrity of the field data was ensured with this tool by making data transcribed from field forms
read-only after initial entry. All ground verification data were summarized per spectral class
within at least three spatial contexts: TM scene-based, ecological subsection-based, and USGS
7.5-minute quadrangle-based. These types of summaries facilitate the analysis of each spectral
class within each TM scene, ecological subsection, or small region (quadrangle). Ecologists
from MoRAP partner agencies used these summaries to complete in-lab expert reviews of all
data in order to detect obvious errors. The tool's built-in GIS functions were used to intersect
GIS data layer values (ancillary data) with spectral class data from the field by TM scene and by
ecological subsection. The use of ancillary data was thus restricted to small areas or to single
spectral classes. Hence, the final product was improved where possible, but any errors related to
the overapplication of ancillary data were avoided.

Results and Discussion

More than 11,500 ground points were visited, the data recorded and processed, and summary
statistics calculated. Initial accuracy of the classification ranged from 64% to 88%, and visual
examination of the revised classification (e.g., no “seams” are apparent between ecological
subsection polygons) indicates significant improvement for the final product. Validity and
accuracy of ground data is always a central focus in attempts to limit error propagation in land
cover mapping.

A large volume of sampling data has a better chance of truthfully representing the target
population (Cochran 1977). In Missouri, the opportunity to collect a large volume of field data
came with the caveat that a large number of people would be involved. We elected to pursue this
course rather than collect a much smaller volume of data in a more “controlled” way. We then
needed to overcome the problems of high data volume and variation in the quality of field data.
We were successful in developing a field form and protocol that was used with little apparent
difficulty by more than 150 field biologists. These forms were delivered in person, with
accompanying verbal explanations, to 15 different natural resource agency field offices. Fewer
than 10 calls for clarification were received from the field. Likewise, the development of the
Missouri Land Cover Ground Verification Tool allowed us to successfully deal with the high
volume of data generated.

Involvement of many different people with different backgrounds and dispositions directly
impacted the quality of our field data. In this regard, data were used for an unsupervised
classification in which TM scenes were subdivided by ecological section. Therefore, errors
associated with any individual were limited to relatively small areas (generally less than 15% of
one TM scene) on the final map. In addition, some field checking errors were detected and
corrected based on (1) review of the field data records by ecologists in the office and (2) review
of “suspect” information using NAP aerial photos. For example, several biologists in one region
reported that grassland patches were all warm-season grasses. This result seemed in error
according to ecologists involved in review of the final map product, and quick field revisits
confirmed the error: they were all cool-season grasslands.

Perhaps the most compelling reason to involve a large number of people in collection of a high
volume of ground data is related to buy-in and practical use of the finished land cover map.
Results of the statewide classification and mapping should be used by natural resource agencies
and planners at all levels of government. Policy decisions such as where to locate new
development, new parks, highways, conservation areas, etc. can be facilitated with this land
cover data layer in combination with other information. However, decision makers will only use the information if they have confidence in its accuracy and validity. Involvement of many
people from different agencies helps ensure that as many people as possible have ownership and
confidence in the results of Missouri's land cover mapping effort. Use of the Missouri land cover
map applied towards conservation issues is our primary measure of success, and we believe that
involvement of so many people in its development will ensure the widest possible use.

Literature Cited

Bailey, R.G., P.E. Avers, T. King, and W.H. McNab, editors. 1994. Ecoregions and subregions
of the United States. USDA Forest Service, 1:7,500,000 map.

Cochran, W.G. 1977. Sampling techniques. Wiley Eastern Limited, New Delhi, India. 428 pp.

Jensen, J.R. 1996. Introduction to digital image processing: A remote sensing perspective. Prentice Hall, Upper Saddle River, New Jersey. 316 pp.

Keys, J.E., Jr. 1995. Ecological units of the eastern United States. USDA Forest Service, Washington, D.C.

Microsoft Corporation. 1998. Microsoft Visual Basic 6.0 Evaluation Guide. Microsoft Corporation, Redmond, Washington. 54 pp.

Potts, A. 1998. What is the open development environment? ArcUser , April-June: 40-41. Schroeder, W., T. Nigh, and L. Richards. 1998. Ecological sections and subsections in Missouri.

USDA Forest Service North Central Experiment Station, General Technical Report, Columbia, Missouri.

Tou, J.T., and R.C. Gonzalez. 1977. Pattern recognition principles. Addison-Wesley, Reading, Massachusetts. 377 pp.


<- Previous | First | Next ->