Research On Landslide Identification Based On Remote Sensing Images Based On Super-resolution Reconstructio

With the continuous improvement of the road construction process in Yunnan Province,the maintenance of the road and its surrounding environment is receiving more and more attention from practitioners.In Yunnan Province,due to the influence of factors such as high mountains and valleys,climate variability and man-made disturbances generated by road construction,this has prompted landslide hazards to become the most prominent geological hazards in the province,and the generation of a large number of landslide hazards has seriously affected the stability of the road during its operation period,therefore It is particularly important to carry out preliminary identification of landslide targets in the highway and its surrounding areas,to achieve landslide location as well as preliminary analysis of landslide characteristics,and to reduce the risk during the operation period of the highway.In this paper,GF-2 remote sensing images and different datasets are used as the data base,and a deep learning-based dataset and model optimisation method is investigated for the landslide identification problem.The main work accomplished in this paper is as follows:（1）In order to make full use of the previous low-resolution landslide images to solve the problem of difficulty in obtaining high-resolution training sets for landslides,this paper uses ESRGAN,a super-resolution reconstruction model based on generative adversarial networks,to achieve super-resolution reconstruction of low-resolution landslide image sets.The model improves the feature extraction performance and stability of the generator by removing the batch normalization layer,adding a multi-stage residual network with residual scaling factors,and using The model improves the feature extraction performance and stability of the generator by removing the batch normalization layer,adding a multi-level residual network and residual scaling factor,and using a relative discriminator to improve the discrimination between the pseudo-image and the original image.The experimental results show that the method can better improve the details and visual perception in the lower spatial resolution landslide images,and the reconstructed higher spatial resolution landslide images are used for the Mask R-CNN（Res Net50）model where BBox AP₅₀ reaches 55.2%and Mask AP₅₀ reaches 54.9%,and the Mask R-CNN（Res Net101）model,BBox AP₅₀ reached 62.9%and Mask AP₅₀ reached62.8%.This dataset effectively improves the accuracy of landslide identification models.（2）To address the performance shortcomings of the current Mask R-CNN model for landslide detection,this paper improves and optimizes the Mask R-CNN model.This paper proposes a feature image extraction and fusion structure that combines the Res Ne St and RFP structures to achieve deep feature learning of the sample dataset,in addition to the modular form of the one-stage target recognition structure in this paper replacing the region suggestion network in the Mask R-CNN model to improve the overall model recognition accuracy by improving the accuracy of the first-stage landslide recognition suggestion frame.The experimental results show that the optimised Res Ne St and recursive pyramid structure combined with the one-stage target detector detector head as a replacement structure for the area proposal network can effectively improve the landslide identification accuracy,and the BBox AP₅₀ and Mask AP₅₀ are 91.4%and 63.5%respectively for the optimal identification accuracy.This paper is a study on the identification of landslide targets based on deep learning methods of remote sensing imagery,aiming to improve the identification accuracy of landslide targets in roads and their surrounding areas through the application of dataset and model optimisation methods,in order to provide timely and accur ate preliminary identification results of landslide data in the work of landslide disaster emergency rescue and disaster assessment.