| Remote sensing can obtain data in agriculture,forestry,geology,ocean,military,urban and other aspects.It is one of the research hotspots in the field of remote sensing to detect and extract targets through remote sensing images,so as to obtain disaster information,plan disaster prevention and relief and post disaster reconstruction.Collapse and landslide are two common and harmful geological disasters.Accurate,fast and efficient acquisition of landslides in remote sensing images is extremely important for disaster information acquisition and assessment,disaster relief and post-disaster reconstruction.Traditional collapse-landslip body extraction methods often require manual design of extraction rules,and the universality of artificially designed rules is usually low.Existing deep learning collapse-landslip body extraction methods ignore the characteristics of collapses and landslides in remote sensing images and the requirements for quick access to disaster information,and generally have the disadvantage of high model complexity.Based on the method of deep learning,combined with the characteristics of collapselandslip body,and considering the timeliness of rescue and disaster relief,this paper constructs a Branch Net network structure that takes accuracy and complexity into account and adds multi-scale design and attention mechanism for remote sensing image collapse-landslip body extraction,and carries out experiments on self-made data sets to verify and evaluate the network performance.The main contents of this paper are as follows:1.In this paper,146 large-scale collapse-landslip body samples are made,which have large amount of data,wide data sources,long time-consuming,large number and rich types of collapse-landslip body,which can effectively improve the accuracy and universality of collapse-landslip body extraction model.2.In this paper,a multi-scale and relatively lightweight network Branch Net is built,and an attention mechanism is added,which can artificially control the parameters and calculation amount by adjusting the hyperparameters.The excellent performance of Branch Net is verified by qualitative and quantitative analysis on three different collapse-landslip body datasets.3.This paper proposes a method of superimposing the binarized visible light vegetation index band on the three-band image and sending it to the network for training.The effectiveness of the method is verified by the comparison experiment of whether to use the vegetation index band.4.This paper proposes a post-processing method of collapse-landslip body extraction based on polygon extraction and filling,and the post-processing method in this paper will generate the analysis report of each collapse-landslip body extraction result image,provide the number,length and area of collapse-landslip body,and serve for disaster analysis.The experimental results show that: the Branch Net proposed in this paper not only has less parameters and calculation amount,but also has higher accuracy,which can obtain information of landslide disasters more quickly and efficiently;Adding vegetation index to participate in the training of local method is helpful to improve the accuracy of landslide extraction;The post-processing method based on polygon extraction and filling successfully solves a large number of problems that the extraction results of landslide are incomplete due to the internal vegetation of landslide,which is convenient for the statistical analysis of landslide disasters. |
