You are here:
A PIXEL COMPOSITION-BASED REFERENCE DATA SET FOR THEMATIC ACCURACY ASSESSMENT
Citation:
Knight, J, R S. Lunetta, S. Khorram, H. Cakir, AND B. Hester. A PIXEL COMPOSITION-BASED REFERENCE DATA SET FOR THEMATIC ACCURACY ASSESSMENT. Presented at ASPRS 2005 Annual Conference, Baltimore, MD, March 7-11, 2004.
Impact/Purpose:
Our research objectives are to: (a) develop new methods using satellite remote sensor data for the rapid characterization of LC condition and change at regional to national scales; (b) evaluate the utility of the new NASA-EOS MODIS (Moderate Resolution Imaging Spectrometer) leaf area index (LAI) measurements for regional scale application with landscape process models (e.g., biogenic emissions and atmospheric deposition); (c) provide remote sensor derived measurement data to advance the development of the next generation of distributed landscape process-based models to provide a predictive modeling capability for important ecosystem processes (e.g., nutrients, sedimentation, pathogens, etc.); and (d) integrate in situ monitoring measurement networks with UAV and satellite based remote sensor data to provide a continuous environmental monitoring capability.
Description:
Developing reference data sets for accuracy assessment of land-cover classifications derived from coarse spatial resolution sensors such as MODIS can be difficult due to the large resolution differences between the image data and available reference data sources. Ideally, the spatial resolution of the reference data would be sufficient to resolve and identify major pixel inclusions; however, excessive resolution introduces additional complexity that can overwhelm the interpreter. Commonly used reference data sources for accuracy assessment include aerial photographs, on site sampling, image data from another sensor, and ancillary GIS data. The different resolutions of these data sources can create problems in developing a way to scale the resolution of the reference data to that of the image data, which is a subject of active research.
This presentation describes a method for creating a land-cover reference appropriate for comparison with 250 m resolution of MODIS pixels. A 250 m fishnet vector coverage corresponding to MODIS pixel resolution was first overlain on randomly selected USGS Digital Ortho Quarter Quads (DOQQs) across the study area. Next, a dot grid comprised of l00 dots was displayed for each 250 m cell. Each grid cell was then visually interpreted to determine the percent of each land-cover class of interest contained within the cell. Finally, a simple decision tree classifier is used to determine a class label for each cell based on its primary class composition and major inclusions.
This method, while labor intensive, resulted in a reference data set that can be compared on a pixel-wise basis with a MODIS derived land-cover classification. This method could be extended to incorporate different classification types, reference data sources, and image data.