Machine Learning to Improve Land Cover Classifications from Multisensor and Multitemporal Data

Principal Investigator: Dr. DeFries

The use of multitemporal and multisensor data in regional and global land cover classifications offers a richness of information and potentially improved classification accuracies. As larger and larger volumes of data become available in the future from MODIS and other sensors, it is imperative to develop automated and reproducible methods to extract land cover information from such data. However, there are several sources of error that are introduced when using multisensor and multitemporal data, including errors from misregistration, bidirectional reflectances, difference in atmospheric properties, radiometric inconsistencies, and actual changes in land surface properties. The machine learning community has developed mature techniques applicable to these problems, but such techniques generally have not been exploited by the remote sensing community. These techniques include, for example, decision tree classification algorithms, boosting, identification of outliers, automatic feature selection and automated extraction of high order features.

In this project, we are applying techniques from machine learning with the aim of improving classification accuracies as well as increasing the degree to which the classification procedure can be automated. We are applying these techniques to regional and global scale land cover classifications using multitemporal, multisensor data. For multitemporal data, we are using the NOAA/NASA 8km Pathfinder Land data set available for 1982 to 1994 and the AVHRR 1 km data set which will be available from 1992 to the present. For multisensor data, we are using a training data set already developed at the University of Maryland based on Landsat and AVHRR data, a data set from Boston University based on analysis of Thematic Mapper data over Central America, and the IGBP Validation data which is currently being generated from high resolution data at the University of Santa Barbara. In the latter years of the project, we expect to apply the techniques developed in this project to data collected from sensors on board the EOS AM platform, primarily MODIS data.