A novel co-training approach for urban land cover mapping with unclear Landsat time series imagery

Landsat time-series (LTS) imagery shows potential for dynamic mapping in urban areas. However, unclear observations (clouds, cloud shadows, snow/ice, and SLC-off data) in the dataset inevitably restrict the efficacy of many of the state-of-the-art classifiers. In this work, we present a novel co-training classification approach consisting of two steps to cope with the unclear observations. Firstly, we develop a method called the MCCR classifier that deals with the unclear observations in an error-recoverable way. The clear observations are utilized to recover the unclear ones in the training samples by the use of the matrix completion (MC) algorithm, and the collaborative representation (CR) classifier is exploited to handle unclear observations in the unlabeled data. Secondly, considering that the random forest (RF) classifier is able to cope with contaminated data in an error-tolerant way, a co-training approach (CotrRM) based on the RF and MCCR classifiers is also proposed to further improve the classification efficacy. The CotrRM method is executed by iteratively constructing semi-labeled training sets based on the crisp and soft predictions of the two individual classifiers on the unlabeled data. To validate the effectiveness of the proposed MCCR classifier and CotrRM method, LTS imagery of the city of Wuhan (a metropolitan city of China) from four years (11 images from 2000, 16 images from 2005, 13 images from 2010, and 15 images from 2015) was adopted. The experiments showed that the MCCR classifier performs as well as the RF classifier for the mapping of urban land cover with contaminated LTS imagery. Moreover, the proposed CotrRM method has the ability to further improve the classification performance. The proposed approach can not only work effectively in the classification, but can also recover the unclear observations in the LTS imagery, courtesy of the MC algorithm. The overall accuracies of the land-cover changes between each two adjacent periods are all over 85%. Given the effectiveness and flexibility, the proposed method could also be applied in other unclear data classification.