Automated Machine Learning Method To Extract The Information Of Landslide Hazards In Miansi Of Wenchuan

Landslide is one of the main natural disasters. It has become an effective method to analyze the landslide disaster by using remote sensing technology. In the application of remote sensing landslide disaster, it often faces the huge amount of data of high resolution remote sensing image. And traditional based on visual interpretation and object-oriented image analysis are limited. According to the actual needs of the high resolution remote sensing image identification and classification of landslide, while taking into account sample texture, shape, etc. And consider the characteristic information and space relationship. This paper random forest algorithm and QBC Committee voting method, proposed a new sample query algorithm RF-QBC, and built for the landslide of high resolution remote sensing image automatic extraction of new technology and methods, through image segmentation, feature extraction, sample query steps accomplish landslide disaster information automation extraction. The idea of the overall design is object oriented analysis under the framework of a series of machine learning algorithm set, core content is in the image segmentation combined with the characteristics of the RF-RFE selection algorithm and RF-QBC algorithm to construct automatic sample selection strategy. First of all, the segmentation of remote sensing images, DEM and other auxiliary data are based on oriented object segmentation. And the segmentation results are as of the object of classification. The spectral and texture features of the segmentation results are selected by using the RF-RFE algorithm. Secondly, use the RF-QBC algorithm to construct the sample query strategy, and the iterative calculation of the sample's voting entropy. Finally, all the labeled samples are taken as the final classification samples. Finally, the random forest classification is performed by the group of samples to achieve the accurate extraction of landslides. The main work in this paper is as follows:(1)Construct a landslide information automation extraction process. Integrated RF-RFE feature selection method, RF-QBC voting Committee sampling method, and random forest algorithm, a method of automatic extraction of landslide hazard information based on machine learning method was constructed. Experimental results show that the proposed method in ensuring the landslide extraction results spatial recognition and the overall positioning has some advantages. The visual solutions classification results compare the extraction effect of the method, which can meet the actual landslide disaster application requirements, and has a higher objective.(2)Select the feature parameters after image segmentation. Parameters of the algorithm using RF-RFE screening image segmentation of spectrum, texture, shape and other features, choose the best subset of features and the Ute syndrome subset of feature importance to sort. The RF-RFE algorithm can effectively eliminate the characteristic subgroups and improve the classification performance of the classifier. Experimental results show that the increasing number of feature, which improve the effect of image object information quantity is not obvious. The final feature subset selected by RF-RFE can truly reflect the characteristics of the type of the experimentation area.(3)Establish the automatic sampling strategy of landslide. Using the RF-QBC active learning method, the automatic selection strategy is established for the non labeled segmentation object, and the maximum information entropy in the iterative process is selected as the final classification. RF-QBC algorithm of the landslide sampling results that in user accuracy, F-measure of remote sensing image sampling evaluation index and the evaluation indexes of classification model has better performance than the original Boosting-QBC algorithm.(4)By using the automatic sampling method to train the random forest algorithm classifier, the landslide is extracted accurately. RF-QBC method is superior to the Boosting-QBC algorithm in the high resolution remote sensing image landslide sampling, and the random forest classification accuracy of RF-QBC landslide sampling is about 11% higher than that of Boosting-QBC. In the whole experimental area, the RF-QBC landslide classification results improve the positioning accuracy, while ensuring the spatial consistency of the results of the landslide extraction, especially in the valley, river and landslide distribution. The predict landslide classification results of RF-QBC algorithm are basically same as the establishment interpretation signs of field verification points. The RF-QBC's extraction results can meet the actual landslide disaster application demand, and visual interpretation compared to the objectivity of the method is higher.