Land cover of Greece, 2010: a semi-automated classification using random forests

Information about land cover (LC) and land use is fundamental in various areas of research regarding the Earth's surface. However, field campaigns are costly and time consuming while existing data sets have strong limitations. Classification of LC by remote sensing, although considered a technically and methodologically challenging task, can facilitate mapping initiatives at various scales. This study suggests an efficient and robust methodology of LC classification with minimal user requirements. The study site is Greece which faces a lack of up to date LC maps at national scale. In this context we employed Landsat imagery, open source software and the random forest classification algorithm to produce a high resolution national LC map for 2010. The algorithm was trained semi-automatically, extracting information from available data sets. The results are promising, achieving an overall accuracy of 83%. The methodology presented minimizes many obstacles that lead to data deficiencies and can act as a baseline for future LC mapping initiatives.