Research On Remote Sensing Detection Method Of Typical Debris Flow Tracks In Xiaojiang River Basin

The Xiaojiang River Basin in Yunnan Province is a famous frequent debris flow area in China.Because of its geomorphic Foundation(spread out in the seismic zone),meteorological and hydrological conditions(rainfall concentration),human activities(excessive logging),debris flow occurs frequently in this area.According to statistics,every rainy season,there are frequent debris flows in this area.Due to the frequent occurrence of debris flows,there are often many economic losses caused by events such as silting up river channels,washing away bridge sections,damaging construction,flooding farmland,devouring villages,etc.Therefore,it is of great significance to grasp the temporal and spatial distribution of debris flow in time.In this paper,the highresolution two image(gf-2)is used as the data source,and the big and small Baini River in Xiaojiang basin,a famous debris flow frequent area in China,is selected as the research area.The object-oriented classification method and convolution neural network method are used to extract debris flow trace information in the research area.The main research contents and conclusions of this paper are as follows:(1)This paper expounds the research background and significance of debris flow trace information extraction,and studies the current situation of image extraction at home and abroad as well as the overall situation of debris flow information extraction.Some representative methods of debris flow information extraction at home and abroad are introduced.(2)The experimental data were processed by radiometric correction,geometric correction and image fusion.The spectral value and variance of each object in the study area are calculated.After analysis,it is found that the spectral characteristics of debris flow trace and other four kinds of ground objects are not well distinguished in four bands.Therefore,vegetation index(NDVI),water index(NDWI)and soil brightness index(NDSI)were selected as the research focus,and five sets of multi-dimensional images were constructed.(3)Object oriented method for debris flow trace information extraction.In this paper,object-oriented image segmentation technology,optimal segmentation scale and classification features are introduced in detail.The chessboard segmentation method,quadtree segmentation method,multi-scale segmentation method,edge detection segmentation method and brightness segmentation method of envi platform are discussed and experimented.It is found that the edge detection segmentation method based on envi platform is the best.Finally,using envi platform to extract the debris flow trace from the original image and five sets of multi-dimensional images.The results show that the accuracy of debris flow trace extraction from the original image is the lowest,with kappa coefficient of 0.7024 and overall classification accuracy of 0.9750;the accuracy of debris flow trace extraction from NDVI + NDWI + NDSI image is the highest,with kappa coefficient of 0.7527 and overall classification accuracy of 0.9778.(4)Convolution neural network method for debris flow trace information extraction.In this paper,the basic structure and main characteristics of convolutional neural network and envinet5 model based on u-net architecture of envi platform are introduced and discussed in detail.The general convolution network model training is based on the original remote sensing image.This paper proposes the convolution network training based on multi-dimensional image,which integrates the exponential characteristics into the training samples,constructs the original samples and five sets of multi-dimensional image samples,and trains these six sets of samples respectively.The results show that the accuracy of debris flow trace extraction from the original image is the lowest,with kappa coefficient of 0.7799 and overall classification accuracy of 0.9792;the accuracy of debris flow trace extraction from the original image + NDVI + NDWI + NDSI image is the highest,with kappa coefficient of 0.8386 and overall classification accuracy of 0.9843.(5)Comparing the two methods,the overall classification accuracy range of envinet5 model is 0.9782-0.9809,kappa coefficient range is 0.7799-0.8386;the overall classification accuracy range of object-oriented method is 0.9718-0.9778,kappa coefficient range is 0.7024-0.7527.It is not difficult to see that envinet5 model is better than object-oriented classification method in extracting the overall classification accuracy and kappa coefficient of debris flow trace.