Land Cover Classification Using Features Generated From Annual Time-Series Landsat Data

The objective of this letter is to investigate and demonstrate potential accuracy improvements in land cover classification using Landsat data, by integrating time-series features with a specially designed classification method. We present a new framework, mapping land cover types using annual time-series Landsat data (LandUTime), which adopts a two-stage approach. First, to generate pattern features, regression analysis is conducted on annual time-series Landsat data at the pixel level. Second, all features are packed into “blocks” and delivered to “UniBagging,” which organizes base classifiers in separated sets according to the feature subspace, and conducts classification by integrating the results of many base classifiers. To evaluate the effectiveness of LandUTime, we performed a series of experiments on six Landsat bands. The results show marked improvements in both overall classification accuracy and Cohen's kappa coefficient. LandUTime is particularly promising for distinguishing objects whose spectral characteristics show temporal differences, such as deciduous or evergreen forest and cultivated crops.