| In the southwest of China,especially in Sichuan Province,geological disasters occur frequently,which seriously threatens people's lives and property.In the event of a geological disaster,the disaster is generally understood by geological survey staff entering the disaster area for on-the-spot exploration or visual interpretation through remote sensing images.Although detailed disaster data can be obtained,it is time-consuming,laborious,threatening to the safety of geological survey staff and requires a high professional quality,which cannot meet the work requirements of disaster prevention and mitigation.With the development of image processing technology and computer software & hardware,the intelligent interpretation of remote sensing image has attracted wide attention because of its high efficiency and simple operation.Among,the convolutional neural network technology is the most prominent.Its application requires a large number of training samples as the data base,but so far there is no remote sensing image sample database of geological disaster.In order to make the convolutional neural network technology better applied to extracting the geological disaster information from remote sensing image,this paper aims to construct a scientific and practical remote sensing image sample database of geological disaster.Based on this,the specific research content and corresponding results of this paper are as follows:(1)A classification system for remote sensing image sample database of geological disaster was designed.Through the analysis of the interpretation marks of the main types of geological disasters(debris flow & landslide)on remote sensing images and the existing land classification system,combined with the characteristics of the study area and research purpose of this paper,a three-level classification system is designed.There are 2 first-class types of geological disaster area and non-geological disaster area,7 second-class types of landslide,debris flow,waters,vegetation,cultivated land,construction land,bare land and 14 three-class of river,paddy field,woodland,residential land,etc.The classification system takes into account both land use conditions and natural land cover,and is consistent with the regional characteristics of Sichuan Province.(2)Following the construction rules of regionality,richness and subjectivity established in this paper,the entity construction of the Southwest Jiaotong University-remote sensing image Sample Database of Geological Disaster(SWJTU-GDSD7)was completed using the construction process proposed in this paper.There are a total of 1,316 images,except for the 134 samples of landslide,each of which contains about 200 images.(3)Design experiment to verify the practicability of the remote sensing image sample database of geological disaster in the application of convolutional neural network technology to intelligently extracting the geological disaster information from remote sensing image.Experiments show that: 1)Samples of different scales have a good performance in the process of identifying the geological disasters by convolutional neural networks technology,but they show some differences for different convolutional neural network models;2)With the continuous increasing number of training samples,the recognition accuracy of each network model is also increasing.3)Through the test of the whole remote sensing image,the overall accuracy is 87.47% and the recall rate of geological disaster is 91.69%,which has achieved a good recognition result.(4)The disaster sample database construction system is developed.The system based on ArcEngine adopts C/S architecture and uses MATLAB and IDL to write core processing code,then calls the corresponding processing program with C#,and its interface is designed by DevExpress.There are two core modules of data preprocessing and sample processing.The system greatly improves the efficiency of the sample database construction. |
PDF Full Text Request