Filtering mislabeled data for improving time series classification

The supervised classification of optical image time series allow the production of accurate land cover maps over large areas. However, the precision yielded by learning algorithms strongly depends on the quality of the reference data. The reference databases covering a large geographical area usually contain noisy data with an important number of mislabeled instances. These labeling errors result in longer training time, less accurate classifiers, and ultimately poorer results. To address this issue, we proposed a new iterative learning framework that removes mislabeled data from the training set. Specifically, a preliminary outlier rejection procedure based on the well-known Random Forest algorithm is proposed. The presented strategy is tested with the classification of Sentinel-2 image time series acquired on 2016 by using an out-of-date reference dataset collected on 2014.