Topic Model Based High Resolution Remote Sensing Image Change Detection

With the development of science and technology as well as the continuous improvement of the level of people's livelihood,satellite remote sensing technology is gradually playing an irreplaceable role in many aspects,as one of the main research direction of it,remote sensing change detection technology has even more extensive and excellent applications. From military wars,disaster warning and relief to city construction planning and crop growth monitoring, it can appear in all these fields. Nowadays, remote sensing technology is constantly trending towards Three-High and Three-? Multi , which means high spatial resolution,high time resolution,high spectral resolution and multi-platform,multi-angle,multi-sensor. This trend makes the characteristics of remote sensing itself more and more obvious, namely huge amount of data,many objects and complex information of surface features. This also brings a great challenge to the traditional remote sensing change detection methods. Therefore, it is indeed urgent to propose new and more improved change detection methods.Topic model is a probabilistic generative model, and it's first orginated from the natural language text processing field. It is the extension of Bag of Words(BOW) model,aiming at abstracting the semantic-related latent topic information of the document by analysising the co-occurrence information of the words on the document level. Topic model has being widely used in the tasks such as dimensionality reduction of data,ambiguous searching and object categorization and so on. With the proposing of concept of visual-words and the advancing in research, topic model is now gradually applied in the field of digital image processing and has achieved well results.A topic model based change detection method is proposed for high resolution remote sensing images in this paper. It takes the pixel pairs of bi-temporal remote sensing images as the basic unit, and extracts their low-level features, such as relevacy, mean value, standard deviation, slope and intercept. Then visual words on the basis of these features are generated. After that, the classical topic model of latent dirichlet allocation is utilized to find the latent topic information, that is, changed topic or unchanged topic, thereby the goal of change detection will be achieved. Experiments show that this method could effectively detect changes in high resolution remote sensing images.