Coping with partially corrupted data

One of the obstacles in data analysis tasks is the variable quality of the data. We investigated ways to automatically deal with corruptions in the data. These include robust measures which avoid over fitting, interpolation-based imputation of missing values, and polishing by which the corrupted elements are fitted with more appropriate values. We applied such methods to a data set of vegetation indices and land cover type assembled from NASA´s Moderate Resolution Imaging Spectroradiometer (MODIS) data collection. The experimental comparison suggested that straight-forward Interpolation, although a commonly used technique for filling in missing values, may not always yield the best results. Robust algorithms and polishing are both viable alternatives for dealing with partially corrupted data, with polishing performing slightly better in our experiments in terms of classification accuracy.