You are here:
The Use of MODIS NDVI Data for Characterizing Cropland Across the Great Lakes Basin
Citation:
Shao, Y. AND R. S. LUNETTA. The Use of MODIS NDVI Data for Characterizing Cropland Across the Great Lakes Basin. Presented at International Geoscience and Remote Sensing Symposium, Boston, MA, July 07 - 11, 2008.
Impact/Purpose:
Presentation
Description:
The Moderate Resolution Imaging Spectroradiometer (MODIS) provides new opportunities for characterizing land-cover (LC) to support monitoring and assessment studies at watershed, regional and global scales. This research evaluated the potential for using the MODIS Normalized Difference Vegetation Index (NDVI) 16-day composite 250 m product (MOD13Q) time-series data to develop a cropland mask and identify four major crop types (corn, soybean, hay, and wheat), throughout the entire 480,000 km2 Great Lakes Basin (GLB). The objective of this research was to evaluate the impacts of scale on the performance time-trajectory analytical approaches for LC classifications. MODIS-NDVI data were first acquired from the USGS EROS Data Center for calendar year 2002 and subsequently preprocessed (anomalous data removed and replaces with estimated values), to provide a high quality uninterrupted data stream to support multi-temporal (phenology-based) analysis. LC classifications were then performed for the entire GLB (n=1) and for individual GLB ecoregions (n=11). For the GLB scale analysis, training samples of agricultural and non-agricultural land were collected across the entire area to support a single regional scale classification. In the latter approach (ecoregion-stratified), the GLB was first stratified into 11 ecoregions and both the training sample collections and classifications were conducted on an individual basis for each ecological unit. Landsat panchromatic imageries from 2000−2002 were used as the primary reference datasets for identifying training samples. Also, a variety of image classification algorithms were examined, including (i) supervised statistical classification, (ii) principle component analysis, and (iii) non-parametric techniques such as neural networks and decision tree. A validation of 2002 NDVI-derived crop mask was conducted using both a pixel-wise and county-aggregated approaches. For the pixel-wise accuracy assessment, testing sites were generated using a stratified random sample approach. The testing sites were visually interpreted from Landsat panchromatic images. Addition, the agricultural pixels derived from NDVI image were aggregated to the county level and compared to statistics from the National Agricultural Statistics Service (NASS). The pixel-level and county-level comparisons provided a multi-level accuracy assessment. The crop mask with the highest accuracy was used as a baseline dataset for subsequent classification of individual crop types. The agricultural lands were classified into five major crop types (corn, hay, soybean, wheat, and other) using three classification algorithms; which were evaluated using the county level agricultural statistics to perform an accuracy assessment. The results from this research indicated that the ecoregion-stratified approach generated superior crop mask compare to the GLB-wide classification. The stratification of the study area reduced the confusions between agricultural and non-agricultural pixels with regard to temporal and/or phenological information. Also, the accuracy assessments results indicated that both neural network and decision tree classifiers performed better than a statistical maximum likelihood classifier. The primary reason was that neural networks and decision tree made no assumptions about the input data distribution and they are also less sensitive to the correlations among input features. Due to the lack of reference data, no pixel-wise accuracy assessment was conducted for individual crop identification (i.e., corn, hay, soybean, and wheat), however, the accuracies of crop acreage estimates at county level were acceptable compare to the NASS agricultural statistics.