Remote sensing methods to detect land-use/cover changes in New Zealand's 'indigenous' grasslands

摘要

We compared four remote sensing methods to detect changes in New Zealand's grasslands (image differencing, normalised difference vegetation index (NDVI) differencing post-classification and visual interpretation). The visual interpretation resulted in the best classification results, with a 98% overall accuracy when compared with ground-truthed data. The tests on automatic classification (image differencing, NDVI differencing) and post classification had much lower accuracies, ranging from 47% to 56%. In the New Zealand grassland landscape, automatic detection methods were not able to differentiate between variations of soil moisture and vegetation phenology from variations in land-use change. This, in combination with topographic effects, which have hampered the automated mapping of vegetation, is the main reason why visual interpretation of high-resolution imagery is still needed.