Automatically samples selection in disaster emergency oriented land-cover classification

The automation level of classification for remote sensing image need to be improved to satisfy the timeliness and high-precision requirements in disaster emergency monitoring and assessment. But, the artificial selection of typical samples restricts the automatic interpretation of disaster information, a problem particularly acute for the development of business operation systems. This paper implements a totally automatic object-oriented land cover classification system based on automatic sample selection. First, the candidate object samples are acquired by fuzzy clustering. Second, image features and land type features are extracted from imagery and prior knowledge, respectively. Afterward, samples can be selected by applying preset thresholds on these features. Distance metric learning is then used not only for further sample selection, but also for more accurate supervised classification. Zhouqu Debris flow disaster images are computed by this method. Results show that the classification outcomes with samples selected automatically are very close to those samples selected by hand. Our results are more stable and objective than those produced manually. Moreover, it is more convenient to batch process images automatically.