Font Size: a A A

Deep Learning Landslide Reconition Method Combining High-resolution Optical Image And Low-resolution Terrain Raster Data

Posted on:2022-06-02Degree:MasterType:Thesis
Country:ChinaCandidate:Q C XieFull Text:PDF
GTID:2480306740455244Subject:Surveying the science and technology
Abstract/Summary:PDF Full Text Request
Frequent occurrence of landslides cause huge losses to the peoples' s property safety.Therefore,how quickly identify and detect landslides is of great significance for post-disaster reconstruction and disaster assesment.With the increasing enrichment of multi-source remote sensing image data,the use of computer to automatically process high-resolution remote sensing images can achieve faster landslide identification and detection,which greatly compensates for the shortcomings of manual field surveys.At the same time,the use of deep learning methods for automatic landslide indentication also solves the limitations of manual feature selection.The current use of convolutional networks for landslide recognition is usually based on high-resolution image data or high-resolution terrain data.Because different types of data have different characteristics,using a single data can not fully identify the characteristics of the landslide and extract the landslide area.Therefore,some scholars try to combine high-resolution remote sensing image data with high-resolution topographic data in order to use the advantages of different types of data to identify and detect landslides.Although this method can effectively identify and extract landslide features,it requires relatively high data accuracy and quality.The terrain data used is often high-resolution DEM data or its derived data.Due to the high cost of obtaining such high-resolution terrain data,there are fewer research on fusion of high-resolution optical image data and high-resolution terrain raster data for landslide identification and detection.Low-resolution terrain raster data has the characteristics of free and open,and is a data source can be well applied to the recognition of special features such as landslides.The landslide recognition of using the deep learning by fusing high-resolution optical image data and low-resolution terrain raster data can effectively combine the optical feature information and terrain geometric information,which is very suitable for landslide recognition and detection.However,there are relatively few studies on the fusion of the types of data,mainly due to the following two reasons: 1)Beacause of the heterogeneity between terrain raster data and optical image data is large,it is difficult to perform effective fusion through simple methods.And the resolution difference between low-resolution terrain raster data and high-resolution optical image is often large,which increases the difficulty of integration;2)For the terrain data provided by the open source platform and the optical image data,It is difficult to ensure that the time points of satellite shooting correspond to each other.In response to the above-mentioned problems,this thesis proposes a deep learning model based on convolutional neural networks,which fusions the high-resolution optical image data and the low-resolution terrain raster data to identify landslides.Specifically,the research content and results of this thesis are summarized as follows:(1)Through comprehensive analysis of multiple open source platforms and data,a reasonable selection of terrain raster data and high-resolution optical image data with the same shooting time point.Then,discussing the characteristics of terrain raster data such as slope,aspect,curvature and relative elevation,it is determined that these four types of terrain factor will be used as the basis for standardization,and the the standardized terrain factor will be superimposed by the channel,so that the model will be able to take into account the characteristics of various terrain factor,and automatically weigh the degree of influence of different factor.(2)This thesis constructs a convolutional neural network model fusioning high-resolution optical image data and low-resolution terrain raster data.In this model,the part of feature extraction contains two branches,which are used to extract terrain geometric features and optical image features.Then,by designing softmax fusion modules of different data source categories,the two types of data are weighted and fused,so that the model can automatically determine the contribution value of different types of data during the training process.At the same time,in order to solve the problem of many small and different-shape landslides in the experimental area.This thesis introduces the Spatial Aware Pool(SAP)module into the model,which can introduce the key information in shallow feature layer into the deep feature layer,so that the model can fully take into account the feature information of different scales during the training process.Through the analysis of experimental results,it can be seen that compared with the classic semantic segmentation network model that only using optical images as input data,the method proposed in this thesis has significantly improved various indicators of landslide detection.Its F1-score and kappa values were 84.15% and 83.62%,respectively.Compared with the DeepLab-V3+,which performed well in the classic model,its precesion increased by 3.2%,recall increased 1.3%,F1-score increased by 2.2%,kappa value increased by 2.3%.Therefore,the method proposed in this thesis can be better adapted to landslide indentification and detection,and has certain application value.
Keywords/Search Tags:CNN, Attention mechanism module, Fusion, Topographic features, SAP
PDF Full Text Request
Related items