Balancing misclassification errors of land cover classification maps using support vector machines and Landsat imagery in the Maipo river basin (Central Chile, 1975-2010)

The ability to carry out land cover change analyses based on Land Use/Cover Classification (LUCC) maps from remote sensing data depends on the quality of the mapping method. Land cover areas obtained from unadjusted classifiers with unbalanced misclassification between different classes could result in erroneously identifying trends. The aim of our work is to describe a novel approach to obtaining LUCC maps with balanced misclassification errors and therefore unbiased predicted areas for each class. We achieve this by numerically minimizing the differences between area proportions obtained with unbiased statistical reference estimates, which is measured by a quantity we refer to as the Sum of Squared Class Unbalancedness (SSCU). We assess the proposed methods in the context of land cover classification with support vector machine classifiers at four points in time between 1975 and 2010 in the Maipo river basin (Central Chile) based on Landsat imagery. In this study, the optimization reduced the SSCU (θ) by 94% on average compared to unadjusted classification. The classifier adjustment also slightly increased the accuracy of the resulting LUCC maps. The amount of bias in classified land cover area and the degree of unbalancedness of misclassification errors differed among the land cover classes. Agricultural land showed the largest reduction in mean relative differences from 27% to 2% compared to the unbiased statistical area estimates. The greatest increase in User's Accuracy was obtained for urban land cover in 1999, where an increase from 56% to 85% was achieved. Qualitative improvements in the classification were visible in difficult classification areas such as dry floodplains. The proposed method is especially recommended for studies that aim to provide multitemporal comparisons.