**1. Introduction**

Production of thematic maps such as land use/land cover and crop type maps using remote sensing data has been regarded as one of most important applications of remote sensing, as it can provide useful information with periodicity and at a variety of scales [1–4]. Since thematic maps are usually used in land surface monitoring and environmental modeling, it is critical that they be reliable [5]. For example, crop type maps are usually fed into physical models for crop yield estimation or forecasting.

Many studies have been carried out to generate a reliable thematic map from remote sensing data. From the data availability aspect, multi-sensor/source data including optical, SAR, and GIS data have been used as inputs for classification [6–9]. To properly treat input data for classification, advanced classification methodologies such as machine learning approaches and object-based classification

have also been applied to either single-sensor data or multiple data sets [10–12]. Even though a proper classification methodology and appropriate data sets are applied to classification, supervised classification usually requires a large amount of high-quality training data. However, this is not always possible to obtain particularly when supervised classification is to be conducted for large or inaccessible areas. It is thus necessary to develop a new classification framework that can alleviate the difficulty of collecting a lot of training data.

To resolve this issue, several approaches have been proposed in the remote sensing community, such as semi-supervised learning (SSL) and active learning (AL) [13–22]. The idea central to these approaches is the use of unlabeled data to complement the training data [13,14]. AL and SSL are very similar in that they begin with an initial classification using a small amount of training data, followed by further classifications using the new training data derived from informative unlabeled pixels in the initial classification result [15–19]. The informative pixels are ones that provide useful information for properly modifying the decision boundary already determined from a small amount of training data, which ultimately lead to an improvement of classification accuracy. However, SSL and AL adopt different ways of extracting the new informative training data from unlabeled data. The SSL approach selects the most confident pixels from the initial classification result as new informative training data, where the most confident pixels mean ones that are likely to be classified unambiguously by a classifier and have the higher confidence [20–22]. Various SSL approaches, such as transductive support vector machine and graph-based methods, have been developed to extract new training data from the initial classification result [22]. New training data can be added directly to the training data without class assignment by an analyst because the classification algorithm itself already assigns the class labels to the most confident pixels. If the initial classification result includes many wrongly classified pixels, however, the new training data extracted from the SSL approach would be wrongly labelled, resulting in the poor classification performance [23]. In addition, the new training data with higher confidence tend to provide redundant information that is not useful for modifying the decision boundary. Thus, there might be no significant improvement to the classification accuracy, compared with AL [23]. Conversely, the most informative pixels in the AL approach are defined as ones that a classifier fails to properly classify, which correspond to pixels with higher uncertainty or lower confidence in the initial classification result. After these pixels are extracted, an analyst then manually assigns their class labels. Since the analyst designates class labels for uncertain or ambiguous pixels, these new training data can positively contribute to modifying the decision boundary. However, it is difficult to apply AL to areas where prior information on land-cover classes is not readily available to facilitate the analyst's interpretation. If the class label assigned by the analyst is incorrect, the accuracy of the classification may deteriorate.

Recently, several studies have proposed combining AL and SSL to take full advantage of both approaches [24,25]. Muñoz-Marí et al. [24] proposed a semiautomatic approach that integrated a hierarchical clustering tree with active queries to generate land-cover maps. Based on hierarchical clustering with a small amount of training data, the most coherent pixels were exploited and an active learning query was applied to extract the most informative pixels. Dópido et al. [25] also developed a SSL approach that adapted AL methods to integrate self-learning. Pixels adjacent to initial training data were selected as candidates for new training data. AL first extracted the most informative pixels from the adjacent pixels, and then these pixels were used as the new training data. In both approaches, the large number of training data could be selected from unlabeled data and a significant improvement in classification accuracy was obtained for hyperspectral image classification. However, despite utilizing spectral and spatial similarities to assign class labels to the most informative pixels within the self-learning framework, there was still uncertainty or difficulty with the class assignment.

Regarding the issue of class assignment, supplementary information from past land-cover maps [19] and predefined rules [26] in an area of interest could be incorporated into self-learning frameworks. For example, information on crop cultivation systems, such as crop rotations, could be effectively used as a kind of temporal contextual information. Although this information could facilitate the collection of additional high-quality training data, the new training data extracted from self-learning tends to be over-sampled for specific class labels that occupy the largest proportion of the study area. As a result, the biased training data might degrade the classification accuracy [19]. To the best of our knowledge, little emphasis has been placed on both class assignment and extraction of unbiased training sets in self-learning approaches for remote sensing data classification.

In this paper, a new self-learning approach is presented for crop classification that can collect a large number of labeled training data without analyst intervention. Rule information on class changes is first generated from past land-cover maps, and then the class labels for the new training data are assigned based on the rules. The impact of the rule information on classification accuracy is also investigated by changing the number of past land-cover maps used. The methodological developments and applicability of this self-learning approach are demonstrated by a crop classification experiment using time-series Moderate Resolution Imaging Spectroradiometer (MODIS) Normalized Difference Vegetation Index (NDVI) data sets and cropland data layers (CDLs) as classification inputs and supplementary data, respectively.
