*4.1. Active Learning Versus Self-Learning*

AL requires analyst intervention for labeling of the most informative pixels to be used for further classification. Similarly, the self-learning approach also selects pixels with high uncertainty as the most informative ones, but the class labels of the most informative pixels are defined from unique sequence rules of time-series past land-cover maps, without any analyst intervention. Thus, manual labeling load can be reduced, which is the main advantage of the self-learning approach. When classification is conducted for large areas (e.g., state or country units) or inaccessible areas, this advantage can be greatly highlighted. However, the self-learning approach does not always aim at obtaining better classification accuracy than AL because in some cases, manual labeling by analyst might be more accurate than automatic labeling in the self-learning approach.

To investigate how classification performance of self-learning is compatible with AL, an additional comparative experiment was conducted. To mimic analyst intervention, manual labeling by analyst was replaced by defining the class label of the most informative pixels to that of corresponding pixels in the 2015 CDL. The same rule for the selection of informative pixels in the self-learning approach was also applied to AL for a fair comparison. The overall accuracy and Kappa coefficient of the AL classification result were 84.99% and 0.757, respectively. When we compared these accuracy statistics of AL with self-learning with the three-year CDLs that showed the best accuracy, the difference in overall accuracy was only 0.57 percentage points (84.99% versus 84.42%). The Kappa coefficient of AL was also very similar to that of self-learning (0.757 versus 0.759). Therefore, the classification accuracy of self-learning, which is compatible to that of AL, confirms the effectiveness of the presented approach in this study.
