| In recent years,many geological disasters,such as earthquakes and landslides,have taken place in China,which have caused serious harm to our society.However,the capability of on-site disaster detection and rapid analysis is still insufficient,and the technical support is not enough.To this end,the state has proposed an “emergency surveying and mapping” layout,which puts forward a clear need for rapid analysis techniques of disaster scene.It also promotes the development of rapid detection techniques for remote sensing geological disaster targets.The "geological disaster target" here is a collective term for damaging objects such as landslides,debris flows and other geological disasters and collapsed houses caused by earthquakes.In recent years,the rapid development of deep learning has gradually deepened its application research in the field of remote sensing,and there have been many research results based on deep learning methods to quickly analyze remote sensing images.Target detection is a key step in remote sensing image analysis,but traditional methods are difficult to meet the urgent needs of emergency mapping in timeliness and universality.Therefore,we use the convolutional neural network method in deep learning to study the rapid detection of geological disaster targets.A series of experimental schemes are proposed to solve the problems of lack of public data sets and the poor effect of directly using open source network models.The purpose is to provide data for the research of target detection algorithms,and to carry out basic experimental research for rapid detection of disaster targets based on convolutional neural networks.Finally,it will provide services for the construction of emergency rescue and disaster prevention and mitigation systems.Specifically,the main research work and preliminary results of this article are as follows:(1)In view of the extremely lack of publicly available geological hazard target detection data sets,we label and produce a UAV imagery geological hazard target data set.In this paper,we collected about 7,000 high-resolution drone images of geological disaster areas,and selected the most representative 1062 images to label three types of typical geohazard target samples,that is,collapsed houses,landslides,and debris flows.This dataset contains a total of 16,535 sample labels,and its name is DED-1.At the same time,this paper introduces the format of DED-1,the particularity of label making,and the characteristics of the dataset itself.(2)Contrastive analysis of parameter tuning training and detection results based on mainstream deep convolutional neural networks.We selected five mainstream deep convolutional neural network models for training and testing on our geological disaster target dataset.Qualitative and quantitative analysis of network performance to verify the availability of this dataset and the applicability of different network models.Experiments show that the model trained by DED-1 can detect the geological disaster targets in the UAV images quite well,the detection time is short,and it has a good target classification and positioning effect.There are problems of missed detection and low detection precision.This article analyzes the causes from the particularity of the disaster target and the applicability of the model,and discusses the optimization direction of dataset production and network model.(3)The strategy optimization and improvement of the target detection model for the DED-1 dataset.Because there are still some missed detection problems in directly fine-tuning these detection networks,this article focuses on the specificity of the DED-1 dataset,and add some strategies to the model for tuning and improvement.The strategies used include data augmentation,improved Anchors initialization based on K-Means clustering algorithm,adding label smoothing,and using a more appropriate learning rate update strategy.We also proposed two new learning rate update methods that are better than the default method.Experiments have found that the above four strategies can improve the detection precision when they are added to the model separately.Among them,the data augmentation strategy has the largest contribution to the precision improvement.The thesis also combines these strategies,and the precision of model detection after the combination of multiple strategies is improved by about17 percentage points.The final precision can basically meet the needs of rapid identification and macro positioning of emergency detection.We also discussed the issue of precision evaluation and tried to find more suitable evaluation methods for DED-1.The innovative work of the paper are as follows:(1)A typical geological disaster target detection dataset is constructed to make up for the extreme lack of such dataset.It provides basic experimental data for the research of geological disaster target detection.(2)The detection results of geological hazard targets have been improved based on various strategy tuning methods.At present,there is very little research on rapid detection of multiple types of geological disaster targets.This paper combines the characteristics of geological disaster data,and then selects a variety of optimization strategies that have been proved to be effective by experiments,and designs a better learning rate update method.Finally,a multi-strategy joint approach is used to further improve the detection precision of geological disaster targets. |
