A comparative study of terra ASTER, Landsat TM, and SPOT HRG data for land cover classification in the Brazilian Amazon

Landsat Thematic Mapper (TM) data have been extensively used for land cover classification, but Terra ASTER and SPOT High Resolution Geometric (HRG) data applications are just beginning. This paper compares the capabilities of TM, ASTER, and HRG in land cover classification in the Amazon basin. Maximum likelihood classification was used for selected multi-sensor image classification. This research indicates that different sensor data have their own merits for land cover classification and no single sensor data or image processing routine provide the best classification accuracy for all land cover classes. The SPOT data fusion result with its 5 in spatial resolution provides the best overall classification accuracies, with Kappa coefficients of 61.8% and 56.3% for 13 land cover classes and 9 vegetation classes, respectively. This is about 3% higher than the second best classification results using SPOT multispectral data with 10 in spatial resolution and ASTER data; and about 4% higher than TM data for 9 vegetation classes. The major errors are due to the confusion between successional stages, agroforestry, and degraded pasture. For the six land cover classification system, the SPOT data fusion provides the best classification accuracy, with overall classification accuracy of 80.4% and kappa coefficient of 75.4%. This research indicates the importance of short-wave infrared bands in land cover classification. Also, higher spatial resolution images provide better classification accuracy when the spectral wavelengths are similar.