Intelligent Landslide Detection Based On Multi-source Data Fusion

Landslide is a kind of geological disaster occurring frequently all over the world,which brings severe harm to the natural environment and human life and property.It is of great significance to study efficient and reliable landslide detection technology for disaster prevention.Traditional methods,such as field survey and visual interpretation of remote sensing images,have the disadvantages of low efficiency and high labor cost.With the explosive of information technology,massive remote sensing data are available for geographic information system,which promotes the wide application of intelligent recognition technology in landslide detection.Deep learning can realize end-to-end learning,reduce intermediate process errors,and the trained model can perform fast model inference,which is suitable for a wide range of landslide detection scenarios.In recent years,deep learning algorithms based on remote sensing technology have become a research hotspot in the field of landslide detection.Semantic segmentation models can generate pixel level landslide segmentation graph,and in this thesis,semantic segmentation models are studied for landslide detection.Huge challenge exists for landslide detection from remote sensing data,because sediments,vegetation and human activities over long periods of time,as well as diverse landscape make it hard to detect old landslides accurately and reliably from high-resolution remote sensing images(HRSIs).Moreover,terrain features like slopes,aspect and height variations cannot be sufficiently extracted from two-dimensional HRSIs but from digital elevation model(DEM).Then,a semantic feature extraction and fusion model on multi-source data is developed innovatively,which extracts and fuses sufficient semantic features from both HRSIs and DEM respectively such that landslides are detected more accurately.In the proposed model,a Siamese network without weight sharing is designed to independently extract color and texture features from HRSIs and topographic features like height variation from DEM,and then a higher semantic feature extraction and fusion network performs high-level semantic feature extraction before the pixel classification to segment landslide from background.To solve the problem of large difference in landslides’ size,a multi-scale channel attention module is designed to achieve the balance between accurate pixel classification and retention of detail information.The landslide detection model based on feature fusion of multi-source data achieves accurate segmentation of landslides on the landslide dataset with obvious features in Bijie city and the old landslide dataset with hidden features in Gansu Province.In order to further improve the Gansu Province’s landslide detection,a generative adversarial network(GAN)-based landslide detection model is designed,which takes the previously designed landslide segmentation model as the generator and a specially designed CNN as the discriminator network.The discriminator and generator are alternately trained,and the prediction result of generator approaches the real data distribution via confrontation between the generator and discriminator,which greatly improves the accuracy of landslide recognition.In order to realize intelligent automation of landslide detection and accurate identification of landslide with hidden features,a landslide detection model based on feature fusion of multi-source data and a model based on generative adversarial network are proposed are proposed.A series of comparative experiments are conducted on two datasets of Bijie’s landslides and Gansu’s landslides to test the performance of the proposed landslide detection models.Sufficient experiments are carried out on both landslides datasets,and the comparative experiment results verify the advantage of the proposed models.Compared with scheme only using HRSIs or the one fusing HRSIs and DEM via channel fusion,the innovative feature fusion scheme,which extracts distinct semantic features firom multi-source data and fuses them in a high-level semantic space,achieves better fusion effect with more accurate landslide segmentation,especially for old landslide detection.Moreover,GAN-based model can further improve the performance of the old landslide detection and provide a new method for the research of old landslide detection.