Abstract:To improve the accuracy of classification with a small amount of training data, this paper presents a self-learning approach that defines class labels from sequential patterns using a series of past land-cover maps. By stacking past land-cover maps, unique sequence rule information from sequential change patterns of land-covers is first generated, and a rule-based class label image is then prepared for a given time. After the most informative pixels with high uncertainty are selected from the initial classification, rule-based class labels are assigned to the selected pixels. These newly labeled pixels are added to training data, which then undergo an iterative classification process until a stopping criterion is reached. Time-series MODIS NDVI data sets and cropland data layers (CDLs) from the past five years are used for the classification of various crop types in Kansas. From the experiment results, it is found that once the rule-based labels are derived from past CDLs, the labeled informative pixels could be properly defined without analyst intervention. Regardless of different combinations of past CDLs, adding these labeled informative pixels to training data increased classification accuracy and the maximum improvement of 8.34 percentage points in overall accuracy was achieved when using three CDLs, compared to the initial classification result using a small amount of training data. Using more than three consecutive CDLs showed slightly better classification accuracy than when using two CDLs (minimum and maximum increases were 1.56 and 2.82 percentage points, respectively). From a practical viewpoint, using three or four CDLs was the best choice for this study area. Based on these experiment results, the presented approach could be applied effectively to areas with insufficient training data but access to past land-cover maps. However, further consideration should be given to select the optimal number of past land-cover maps and reduce the impact of errors of rule-based labels.

No abstract