INVENTORY ASSESSMENT OF GRASSLANDS IN IRELAND USING HYPERTEMPORAL OPTICAL DATA

摘要

Grassland is the dominant land-cover in Ireland, but so far no image based operational land-cover classification system is available for the discrimination of the variety of grassland types. The resulting knowledge gap on grassland coverand -management figures in Ireland needs to be addressed if accurate terrestrial carbon estimates are to be derived. For this purpose we present a methodology for the classification of general landcover and different grassland management regimes. MODIS composite time-series data for the years 2002 to 2010, with an emphasis on 2008, were classified with different machine learning methods, namely Support Vector Machine, Random Forest and Extremely Randomized Trees in two study areas in Ireland and compared with the output from the Maximum Likelihood classifier. In homogenous areas (pure pixels), the 5-fold-crossvalidated classification accuracies exceeded 96 % for all machine learning classifiers, with SVM reaching 99.1 % in 2010 (κ=0.989). However, due to the low spatial resolution of the sensor (250 m) and the high fragmentation of the Irish landscape, mixed pixels are a common occurrence, but can to some extent be analysed using a class probability approach. However, for detailed land cover and use, it is necessary to incorporate additional higher spatial resolution datasets by data-fusion.