Watershed-scale crop type classification using seasonal trends in remote sensing-derived vegetation indices.

摘要

The objective of this study was to compare two methods using multiple satellite remote sensing datasets to differentiate land cover, including crop type, for the Salt River/Mark Twain Lake basin in northeast Missouri, USA. Method 1 involved unsupervised classification of Landsat visible and near-infrared satellite images obtained at multiple dates in the growing season, followed by traditional, manual class identification. Method 2, developed in this research, employed the same unsupervised classification but also used normalized difference vegetation index (NDVI) maps obtained on a 16-day cycle from MODIS satellite images as ancillary data to derive seasonal NDVI trends for each class in the classification map. Tree analysis was applied to the NDVI trend data to group similar classes into clusters, and crop type for each cluster was determined from ground-truth data. Additional ground-truth data were used to assess the accuracy of the procedure, and crop acreage estimates were compared to county-level statistics. The crop types identified during the study were: maize, soyabeans, wheat and grasses. The overall classification accuracy of Method 2 was 3% higher than that of Method 1. Method 2 was also more efficient in terms of analyst time and ground-truth data requirements. Therefore, this method, employing variations in seasonal NDVI trends, is suggested for differentiation of crop type. The 30-m resolution crop type maps developed using this process will be useful as input data to environmental analysis models.