Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/22/2006
Publication Date: 3/15/2007
Citation: Su, L., Chopping, M.J., Rango, A., Martonchik, J.V., Peters, D.C. 2007. Support vector machines for recognition of semi-arid vegetation types using MISR multi-angle imagery. Remote Sensing of Environment. 107:299-311. Interpretive Summary: Nadir viewing reflectances from satellite sensors such as Landsat and MODIS sometimes have problems adequately classifying community types in desert grassland areas. Multispectral sensors that can view the surface at different angles like the Multi-angle Imaging Spectro-Radiometer (MISR). Surface anisotropy patterns from bidirectional reflectance distribution function models, and support vector machine algorithm were tested to see if the community type classification in desert grasslands can be improved. It was determined that nadir-viewing classifications had a classification accuracy of 45.0%; addition of the MISR data, then surface anisotropy patterns, and finally the support vector machine algorithms improved the classification accuracy to 60.9%, 68.1% and 76.7%, respectively. In the future these additional remote sensing approaches may be useful to wild land managers in different agencies while for now the results are of most interest to remote sensing scientist attempting to develop optimum methods for classification in desert grasslands using all possible approaches as demonstrated here.
Technical Abstract: Mapping accurately community types is one of the main challenges for monitoring arid and semi-arid grasslands with remote sensing. The multi-angle approach has been proven useful for mapping vegetation types in desert grassland. The Multi-angle Imaging Spectro-Radiometer (MISR) provides 4 spectral bands and 9 angular reflectance. In this study, 44 classification experiments have been implemented to find the optimal combination of MISR multi-angular data to mine the information carried by MISR data as effectively as possible. These experiments show the following findings: 1) The combination of MISR's 4 spectral bands at nadir and red and near infrared bands in the C, B, and A cameras observing off-nadir can obtain the best vegetation type differentiation at the community level in New Mexico desert grasslands. 2) The k parameter at red band of Modified-Rahman-Pinty-Verstraete (MRPV) model and the structural scattering index (SSI) can bring useful additional information to land cover classification. The information carried by these two parameters, however, is less than that carried by surface anisotropy patterns described by the MRPV model and a linear semi-empirical erneldriven bidirectional reflectance distribution function model, the RossThin-iSparseModis (RTnLS) model. These experiments prove that: 1) multi-angular reflectance raise overall classification accuracy from 45.8% for nadir-only reflectance to 60.9%. 2) With surface anisotropy patterns derived from MRPV and RTnLS, an accuracy of 68.1% can be obtained when maximum likelihood algorithms are used. 3) Support Vector Machine (SVM) algorithms can raise the classification accuracy to 76.7%. This research shows that multi-angular reflectance, surface anisotropy patterns and SVM algorithms can improve desert vegetation type differentiation importantly.