Global discrimination of land cover types from metrics derived from AVHRR pathfinder data

Abstract

Seasonal changes in the greenness of vegetation, described in remotely sensed data as changes in the normalized difference vegetation index (NDVI) throughout the year, have been the basis for discriminating between cover types in previous attempts to derive land cover from AVHRR data at global and continental scales. Several researchers have suggested and applied the use of metrics, such as maximum NDVI or length of growing season derived from a temporal profile of 10-day or monthly NDVI values, as an alternative to classifying cover types from the monthly NDVI values directly. This study examines the use of metrics derived from the NDVI temporal profile, as well as metrics derived from observations in red, infrared, and thermal bands, to improve discrimination between 12 cover types on a global scale. According to separability measures calculated from Bhattacharya distances, average separabilities improved by using 12 of the 16 metrics tested (1.97) compared to separabilities using 12 monthly NDVI values alone (1.88). Separabilities improved from poor to good in 20 out of 25 pairs of cover types with poor separability. Percentage of pixels correctly classified in a maximum likelihood classifications also improved by using the metrics from 76% to 86%. Overall, the most robust metrics for discriminating between cover types were: mean NDVI, maximum NDVI, NDVI amplitude, AVHRR Band 2 (near-infrared reflectance) and Band 1 (red reflectance) corresponding to the time of maximum NDVI, and maximum land surface temperature. Deciduous and evergreen vegetation can be distinguished by mean NDVI, maximum NDVI, NDVI amplitude, and maximum land surface temperature. Needleleaf and broadleaf vegetation can be distinguished by either mean NDVI and NDVI amplitude or maximum NDVI and NDVI amplitude.