Crop Type Assessment
A Phenological Approach for Mapping Crop Types using Hypertemporal MODIS NDVI Data Products
Background
The Production Estimates and Crop Assessment Division (PECAD) of the U.S. Dept. of Agriculture’s (USDA) Foreign Agricultural
Service (FAS) is one of the world’s major operational users of remote sensing data. FAS / PECAD uses a variety of imagery,
including NASA assets, in conducting crop area estimates and assessments and seeks operational improvements. Also, the National
Aeronautics and Space Administration (NASA) Geospatial Information Requirements Focus Area Working Group identified the need to
develop rapid land cover and acreage estimates in areas with limited ground reference data.
Purpose
This project uses Moderate Resolution Imaging Spectroradiometer (MODIS)-derived hypertemporal, curve-fitted vegetation indices
for accurately mapping corn and soybean crops cultivated during the 2004/2005 growing season within portions of Argentina’s
Córdoba and Santa Fe provinces (see Figure 1). Hypertemporal analysis provides a means to map crop types based upon phenological
curves. We selected this study area because it falls within one of the world’s most agriculturally productive regions (the Pampas)
and Argentina is one of the United States’ major competitors.
Research Partners
NASA funded this project through the Mississippi Research Consortium (MRC), a coalition of Mississippi’s four research
universities (Jackson State University, Mississippi State University (MSU), University of Mississippi, and University of Southern
Mississippi) formed in 1986. The Institute for Technology Development (ITD) is supporting researchers affiliated with MSU’s
GeoResources Institute in crop yield modeling efforts. Science Systems and Applications, Incorporated (SSAI) and Computer Sciences
Corporation (CSC) are supporting ITD via the continued development of the below-mentioned Applications Research Toolbox (ART) and
the Time Series Product Tool (TSPT).
Data
Our research team obtained MODIS daily Terra and Aqua image products, including 250m GSD Planetary Surface Reflectance, 1km
Geolocation Angles, and 1km Surface Reflectance Quality Data.
We acquired in situ data as a result of a partnership with the National University of Córdoba’s Center for Surveying
Agriculture and Natural Resources (CREAN). ITD developed an Ipaq Field Data Collection System for CREAN researchers to collect
information during the 2004/2005 growing season. They used this combined hand-held computer, GIS, and GPS to navigate to sample
locations and record field observations on a digital form. The system included imagery for background information.
Figure 1. Map showing the relationship of the study area (grey area) with some of Argentina’s Provinces. Inset map shows the main map’s extent in relation to South America.
Methods
Image Pre-processing
ART was used to re-project MODIS imagery from Sinusoidal to UTM WGS84 Zone 20 South (ART includes the MODIS Re-projection Tool
(MRT) as well as MRT-Swath for radiance data) and to subset imagery to the area of interest. TSPT was then used to derive daily
NDVI images from the MODIS Aqua and Terra data, to fuse Aqua and Terra datasets together for each given day, and to create daily,
curve-fitted Normalized Difference Vegetation Index (NDVI) imagery via median and Savitzky-Golay filtering.
Analysis
We produced 8-day layerstacks from the curve-fitted, daily MODIS NDVI imagery. Individual images matched the ending dates of
the MODIS gridded, 8-day, reflectance products were selected in producing the 8-day layerstacks, beginning with the 5 September
2004 composite and ending with the 18 February 2005 composite. This enables a comparison between the derived hypertemporal and
the standard MODIS data products.
We then performed supervised classifications and conducted accuracy assessments.
This involved randomly selecting training and test samples for corn and soybean fields using buffered versions of a shapefile
that provided field boundary and crop type information for the 2004/2005 growing season (ground truth). Confusion matrices that
included the overall accuracies, producer and user accuracies, the kappa coefficients, and errors of commission and omission were
then produced. Figure 2 shows the results obtained from using the Mahalonobis Distance Classifier. The Minimum Distance Classifier
produced nearly identical results. However, classification approaches that use covariance matrices, e.g. Maximum Likelihood,
failed to produce accurate results, probably because of a singularity caused by the lack of variation between NDVI composites.
Conclusions and Future Research
The research results suggest that FAS / PECAD would benefit from using 8-day hypertemporal NDVI composites derived from daily
reflectance data currently acquired by MODIS at a spatial resolution of 250m. Since conducting the above-described experiment, the
team is making improvements to the ART and TSPT computer codes and developing standards for creating metadata files. Code
improvements include developing weighting schemes related to image quality, e.g. cloud-free pixels would be more heavily weighted
than cloudy pixels, and image geometry.
The team is planning to recreate the analysis described above using improved versions of ART and TSPT. A similar analysis will
be conducted using MODIS radiance data products. MODIS radiance imagery is provided in a raw and un-gridded format, unlike the
reflectance imagery. Because converting the gridded reflectance imagery from the sinusoidal projection to the UTM projection may
introduce error, classification results should improve when using radiance versus reflectance imagery.
Figure 2. Results obtained from using the Mahalonobis Distance Classifier on the MODIS-derived layerstacked 8-day NDVI composites. The producer’s and user’s accuracies for corn (soybean) were 81.3% and 73% (89.8% and 93.4%), respectively.
For more information about this or any other current research, contact us.
|
