Research On Landslide Disaster Detection Model Based On Deep Learning

Landslides are common and extremely harmful geological disasters that pose a serious threat to global human safety and result in damage to property economies and the surface environment.After a landslide,it is very important to quickly obtain the location information of the landslide by using satellite images for rescue and recovery work.It is also helpful to build a landslide database,draw landslide susceptibility maps,and coordinate the harmonious development of humans and nature.Therefore,it is very important to quickly and accurately detect the landslide area.Deep learning,with its high robustness and strong feature expression capability,has developed rapidly and been widely used in recent years.Based on the deep learning one-stage object detection model,the performance capability of the deep learning model on landslide features from satellite images is explored.A landslide disasters detection model with better parametric number,speed and accuracy is constructed.It is used for rapid detection of landslide disasters areas on satellite remote sensing images.The details are as follows.(1)Starting from deep learning data sources,multiple data enhancement processes are performed with landslide remote sensing images as the base data.In the optimization process of deep learning algorithm,different batch data will produce different global and local information.Increasing the variety of data can provide multiple local information and enhance the optimization results of the model.Augmentation methods include single-sample and multiple-sample data augmentation.Among them,the single-sample data enhancement methods include:geometric transformation,color transformation,and hybrid transformation.The multisample data enhancement methods include: blended image,erased image,label smoothing,etc.Taking Qiaojia County and Ludian County in Yunnan Province as the study area,three channels(RGB)optical remote sensing images of historical landslides were acquired from Google Earth,with a total of 1818 images.After enhancement,8256 landslide sample sets were obtained,and the image data were labeled and processed.The 8000 non-landslide sample sets are constructed.The experimental results show that the enhanced data type makes the model have better robustness and generalization ability.The miss detection rate is 1.56% in the detection of potential landslide areas.In particular,the Mosaic data augmentation approach has a better contribution to the landslide detection model.(2)To study the feature representation capability of the model for landslide disasters areas.Using the enhanced image data,the phantom residual module and fully connected network are used to construct the model framework,and the feature representation model is trained by combining with the optimization algorithm.In the experiments,the learning ability of the model is evaluated by its ability to distinguish landslides from non-landslides.The distribution of feature maps is used to analyze the model's ability to represent landslide areas.The model was trained on about 16,000 remote sensing image data and achieved a high accuracy rate,i.e.,the model has a high feature representation capability.It was further found that the bottom and top layers of the model had better learning ability for small and large landslides,respectively,after analyzing the feature maps.(3)A landslide disasters detection model with small number of parameters is proposed.The study proposes a landslide detection model YOLO-SA(You Only Look Once-Small Attention)based on the one-stage object detection model YOLOv4(You Only Look Once)and reconstructs the model framework by using ghost convolution module,group convolution,and attention mechanism.The model addresses the problems of slow detection speed of two-stage model and low detection accuracy of one-stage detection model,and greatly reduces the number of parameters of the model with high detection accuracy while maintaining the model with high detection speed.The model training is based on the enhanced dataset,and the training process migrates the landslide features to learn what the model has learned.YOLO-SA is compared with 11 advanced models,including Faster-RCNN,3 types of Efficient Det,2 types of Centernet,SSD-efficient,and 4 types of YOLOv4 models.The results show that the number of YOLO-SA parameters is reduced to1.472 mb compared to Efficient Det-B0;the accuracy is improved to 94.08%compared to Centernet-hourglass;and the speed is up to 42 f/s.In addition,the effectiveness of the YOLO-SA model for potential landslide detection is verified,with an F1 score of 90.65%.