Signature extension through space for northern landcover classification: A comparison of radiometric correction methods

Abstract

Northern landcover mapping for climate change and carbon modeling requires greater detail than what is available from coarse resolution data. Mapping landcover with medium resolution data from Landsat presents challenges due to differences in time and space between scene acquisitions required for full coverage. These differences cause landcover signatures to vary due to haze, solar geometry and phenology, among other factors. One way to circumvent this problem is to have an image interpreter classify each scene independently, however, this is not an optimal solution in the north due to a lack of spatially extensive reference data and resources required to label scenes individually. Another possible approach is to stabilize signatures in space and time so that they may be extracted from one scene and extended to others, thereby reducing the amount of reference data and user input required for mapping large areas. A radiometric normalization approach was developed that exploits the high temporal frequency with which coarse resolution data are acquired and the high spatial frequency of medium resolution data. The current paper compares this radiometric correction methodology with an established absolute calibration methodology for signature extension for landcover classification and explores factors that affect extension performance to recommend how and when signature extension can be applied. Overall, the new normalization method produced better extension and classification results than absolute calibration. Results also showed that extension performance was affected more by geographical distance than by differences in anniversary dates between acquisitions for the range of data examined. Geographical distance in the north–south direction leads to poorer extension performance than distance in the east–west direction due in part to differences in vegetation composition assigned the same class label in the latitudinal direction. While extension performance was somewhat variable and in some cases did not produce a best classification result by itself, it provided an initial best guess of landcover that can subsequently be refined by an expert image interpreter.