Improvement of land cover classification accuracy by training sample clustering

Abstract

The subject of this article is land cover classification based on geospatial data. The supervised classification methods are appropriate for most of the thematic tasks of remote sensing because they provide the opportunity to set the characteristics of the initial classes in the form of a training sample set, in contrast to unsupervised methods. There are many approaches to processing such a set; however, their common disadvantage is that they do not consider the factor of training sample separability. This characteristic indicates the extent to which signatures representing different classes do not overlap. A low degree of separability is inherent in high-level training sample mixing. Thus, separability affects classification accuracy. One possible ways to increase separability is training sample clustering. Considering the above, the goal of this study is to develop a training sample clustering technique to improve land cover classification accuracy by increasing the separability of training samples. The tasks of this work are as follows: 1) develop a method for training sample separability assessment; 2) develop a training sample clustering technique based on training sample separability; 3) test the effectiveness of the developed technique by applying it to experimental land cover classification. In the experiments, two land cover classifications were obtained for each of the two selected study areas (i.e., one before and another after training sample clustering. Six land cover classes were defined for each experiment. The training samples were selected for each class. Conclusions. After the application of the developed technique, an increase in the separability of the training samples was evidenced by the developed separability index. In turn, this approach led to an improvement in land cover classification. For the first experiment, this was evidenced by an increase in the overall accuracy and kappa coefficient by 20% (from 63 to 83%) and 21% (from 60% to 81%), respectively. In the second experiment, the increase was 4% (from 77% to 81%) and 5% (from 66% to 71%), respectively.

Keywords

• classification
• supervised classification
• unsupervised classification
• clustering
• remote sensing
• training sample
• training sample separability