Improving Land Cover Maps in Areas of Disagreement of Existing Products using NDVI Time Series of MODIS - Example for Europe

Regional to global scale land cover (LC) information is one of the most important inputs to various models related to global climate change studies, natural resource use and environmental assessment. This paper presents a methodology to derive land cover maps using time series of moderate-resolution imaging spectroradiometer (MODIS) 250 m normalized difference vegetation index (NDVI). An example for Europe is produced using the random forest (RF) classifier. For the accuracy assessment, the overall performance of our classification product (BOKU, Universitat fur Bodenkultur) is compared to the one of three existing LC maps namely GlobCover 2009, MODIS land cover 2009 (using the IGBP classification scheme) and GLC2000. Considered GlobCover and IGBP, the assessment is further detailed for areas where these two maps agree or disagree. The BOKU map reported an overall accuracy of 71%. Classification accuracies ranged from 78% where IGBP and GlobCover agreed to 63% for areas of disagreement. Results confirm that existing LC products are as accurate as the BOKU map in areas of agreement (with little margin for improvements), while classification accuracy is substantially better for the BOKU map in areas of disagreement. Two pixel-based measures of confidence of classification were derived, which showed a strong correlation with classification accuracy. The study also confirmed that RF provides an unbiased estimation of the error (out-of-bag) and therefore eliminates the need for an independent validation dataset.