Improving MODIS land cover classification by combining MODIS spectral and angular signatures in a Canadian boreal forest

Abstract

This study explores the use of reflectance anisotropy as described by the Bidirectional Reflectance Distribution Function (BRDF) as an additional source of information to improve land surface classification accuracies in a Canadian boreal forest region through the use of a decision tree classifier (C4.5). This effort primarily uses a daily rolling version of the operational algorithm developed for Direct Broadcast to generate 500 m 16-day daily rolling data sets in the study region. Descriptive statistic and statistically rigorous techniques are used to assess classification accuracies based on confusion matrices and a 10-fold cross-validation method. The results show that the inclusion of additional 7-band model anisotropic parameter group (volumetric (VOL) plus geometric (GEO)) with spectral feature group (nadir BRDF-adjusted reflectance (NABR) plus Enhanced Vegetation Index (EVI)) is most useful in classification, increasing overall accuracies by 5.68%. The most improvements of per-class accuracies are seen for Wetland shrub class with user's and producer's accuracies increasing by up to 17.7% and 11.3%, respectively. Increases on the order of 5% to 15% are encountered for the classes of Wetland herb, Wetland tree, Coniferous dense, and Coniferous open with no detriments to other candidate classes. The inclusion of the 2-band BRDF shape indicator group in the classification is, however, not as useful as inclusion of the 7-band model anisotropic parameter group in improving the classification accuracies. A further investigation of the classification accuracies regarding reflectance anisotropy for the sampling pixels within each class shows that land cover types that are dominated by geometric-optical scattering type or a mixture of scattering types are relatively difficult to be classified with spectral feature group alone, and the inclusion of additional BRDF features can significantly improve classification accuracies for these land cover types. However, despite their use as ancillary data, this study also confirms that the spectral feature group provided with NBAR and EVI captures the major information content regarding land cover types, exceeding the information content contained in the model anisotropic parameter group provided with the 7-band VOL and GEO parameters of RossThick-LiSparse-Reciprocal (RTLSR) models.