| Landslides pose a serious threat to human life and natural resources.With the development of remote sensing technology and significant progress of deep learning,a technology which combines remote sensing image and deep learning has become an important means of landslide detection.This method provides a new way for pre-disaster warning and post-disaster rescue.Among remote sensing images,optical images show advantages in the field of landslide detection due to their high resolution,rich spectral information and wide coverage.Although researchers have achieved some remarkable achievements in landslide detection by using optical images,there are still some shortcomings in the current research:(1)landslide presents various sizes and shapes,so how to better extract and preserve the multi-scale shape information of landslide has become an urgent problem to be solved;(2)Optical remote sensing images are easy to be affected by weather,illumination and other factors,and the visual characteristics of landslides are not obvious.Especially in complex topographic areas,it is difficult to achieve ideal recognition effect only by using optical images.Transformer has a strong global feature extraction capability,which provides a new idea for more accurate extraction of landslide zhe.Therefore,to solve the above problems,this paper proposed two landslide detection models based on Transformer,with the main research contents as follows:(1)A shape-enhanced Vision Transformer model(ShapeFormer)is proposed.With Pyramid Vision Transformer(PVT)as the baseline network,the model can extract landslide features of different scales.At the same time,the shape feature extraction branch is added to strengthen the recognition effect of landslide contour.After testing on two public datasets,the model has reached the optimal accuracy of current research in the industry,and the experimental results demonstrate the potential of the proposed model.(2)The optical remote sensing image has a high resolution,but the visual features are not obvious enough for the loess landslide area.The Digital Elevation Model(DEM)enables detailed representations of complex terrain.The combination of optical image and DEM can extract more comprehensive three-dimensional landslide features,which is helpful to improve the ability of landslide detection model.Taking Pengyang County and Xiji County of Guyuan City,Ningxia as the research area,we executed a landslide dataset which combined high-resolution optical image and DEM data is produced in this paper.(3)In order to merge the features of optical image and DEM data more effectively,a two-branch model DEPNet was proposed in this paper.The model EfficientNet-B5 and PoolFormer were used as extractors of optical features and DEM data respectively,and the two features were combined in multilevel parallel at the decoding stage.This structure not only has a powerful feature extraction function,but also can fully combine the two kinds of data.Experiments on self-built dataset and public datasets prove that the model can produce ideal semantic segmentation effect.(4)This paper designs and implements a landslide detection system based on remote sensing images,which can display the landslide detection results of two public datasets and Guyuan dataset. |
