A Study On The Automatic Interpretting Of Deris Flow Fan Under Topic Modeling Framwork

Debris flow is a common type of sudden geological hazards, which seriously threaten the safety of mountain masses. As product of debris flow sedimentation, the debris flow fan is not only an important symbol of studying the growing degree of debris flow, but also the metrics to assess its hazard range. Therefore, to accurately identify debris flow fan is of great significance to disaster state evaluation.Take debris flow and other geological hazards as an example, the priority is generally visual interpretation method, which features high precision though, but quite inefficient. Especially for the unexpected geological hazards, it barely meets real-time requirements of rapid interpretation. On one hand, the improvement of the image resolution lead to the image exhibiting more complex spectral character, and thus pixel-based interpretation no longer apply to high resolution remote sensing image. On the other hand, object-oriented remote sensing analysis is restricted by multi-scale segmentation technology. In short, it is particularly urgent to carry on an automatic or semi-automatic rapid technique for debris flow extraction.Probabilistic topic model is a kind of hierarchical clustering model for text modeling, and it can be a good method to deal the complex spectral character by establishing the mapping relationship between the objects of the text model and remote sensing image analysis objects. Geological hazards are often expressed as a complex scene composed of multi-objects, meanwhile, the topic model has the ability to build a symbiotic relationship between different objects, which accounts for the fact that the inherent characteristics of the topic model was in line with the demand for geological hazards scene classification. Therefore, under the framework of topic modeling, this article, aiming at the current application needs to extract debris flow hazards, will focus on studying high-resolution remote sensing research methods applicable to automatic interpretation of debris flow fan, as well as verify its effectiveness in practical applications by analyzing the interpretation accuracy.Adopting Google Earth remote sensing images for visual interpretation, this paper selects debris flow-prone areas Jinsha River and its tributaries as a study area to obtain several samples of the debris flow fans in the study area, and to build a scene library of debris flow fans. In order to achieve automatic interpretation of debris flow fans, firstly, this paper completes the representation of scene image library by constructing the Bag of Visual Word model. Then, this paper also introduces Latent Dirichlet Allocation model to complete the topic modelling framework structure. Afterwards, it changes to adopt Support Vector Machine to introduce the class information to complete automatic interpretation of debris flow fans scene, and results in following conclusions by analyzing the experimental:(1) The debris flow fans can get higher accurate recognition which can be up to 90% when based on automatic interpretation of debris flow fan under the LDA topic modeling framework, and the following selected scene image matters, including debris flow fans scene, woodland scene, residential area scene, agricultural land scene, water scene, with the topic number range from 15 to 35, and vocabulary size range from 500 to 700.(2) In this paper, two methods for automatic interpretation for debris flow fans scene, on one hand, one of which is to directly input histogram images of the scene image into SVM classifier; on the other hand, the histogram images of the scene images will be first input with LAD model to dig up potential implicit topics of the scene, then the SVM classifier will be included. When the vocabulary size is 700, the recognition accuracy of the latter is improved by 18.5% compared to the former, which proves that adopting the LDA model to remodel the topic model, has significant validity for automatic interpretation of debris flow fans.(3) Bag of Visual Words model is a very effective method of image representation, and as far as the remote sensing image analysis is concerned, LDA model is also an efficient unsupervised learning models. Experimental results show that the method adopted in this paper are simple, strongly robust, and can achieve high recognition effect from automated interpretation of debris flow fans scene.(4) The remote sensing images can be regarded as a scene image composed of a variety of objects, the scene classification interpretation not only can achieve the same purpose as the visual interpretation, but also partly overcome the defects of visual interpretation. Therefore, this paper proposes an automatic interpretation model for geological hazards under topic model framework, which opens up a new idea for future automatic interpretation of adverse geological hazards.