Research On The Extraction Of Water-eroded Valleys In The Loess Plateau Based On Deep Convolutional Neural Networks

The Loess Plateau is known as one of the unique geographical areas with the most geological research value in the world.During the development of the Loess Plateau for more than 2 million years,the complex topographic features,erodible loess,and concentrated rainstorms have made the Loess Plateau region the most severely eroded area in my country.Thousands of gullies,undulating hills,ridges,and severe gully erosion are important features of the loess landform.Loess gullies,dominated by the erosion of the loess landform,have always been an important research object for the development of loess landform erosion.They can be divided into two basic types:waterworn gullies and inherited gullies.Due to the activeness of various gully erosions in modern erosion of the Loess Plateau,the complexity of internal and external forces,the complexity of dynamic loess development and evolution,and the variability of influencing factors,there is a fusion of inherited gullies and waterworn gullies.Waterwom gullies stripping off inherited gullies are the core and key to the study of the essential characteristics of loess landforms.This study takes waterworn gullies in the Loess Plateau as the basic analysis object,in accordance with the evolution sequence of the loess landform,and under the constraints of the loess gullies,selects typical small watersheds in different types of areas of the Loess Plateau in northern Shaanxi that are dominated by water erosion as the experimental unit.Taking into account the characteristics of topography and geomorphology,using high-precision DEM and high-resolution remote sensing image data as basic information sources,the introduction of deep learning methods,combined with digital terrain analysis methods in GIS,aims to achieve water erosion for typical geomorphic areas automatic extraction and identification of gullies.The network is trained based on multi-source data so that it can learn the typical geomorphic features of loess waterworn gullies from remote sensing images,DEM and topographic feature factors.This article chooses the Loess Plateau as the research area,and collects various data in this area to make a training set.In addition,in the Loess Plateau of Ningxia and central Gansu,where there are a large number of inherited gullies and waterworn gullies,2 typical experimental plots were selected,and the U-Net neural network was used to distinguish between the 2 types of gullies to achieve effective extraction of waterworn gullies in the mixing loess gully areas.The main conclusions are as followed:(1)Based on high-resolution remote sensing image data,a convolutional neural network method for extracting waterworn gullies under the constraints of the gullies was designed,and the waterworn gullies were extracted from the study area in the middle and lower reaches of the Luohe River.The results of U-Net network model training show that the overall classification accuracy of verification reaches 92.82%,and the extraction accuracy of water erosion gullies is 89.09%.Using the best trained model to extract waterworn gullies in the overall key study sample area of Yijun,the recognition degree is high relatively.(2)Based on high-precision terrain data and high-resolution image data,a multi-band input U-Net fully convolutional neural network model is constructed.In typical watersheds,the extraction of waterworn gullies has been achieved,and the accuracy of identifying waterworn gullies has reached 94.1%.The combination of image features and topographical factors can make the extraction results of waterworn gullies more accurate.(3)Aiming at areas where there are a large number of integrated gullies where inherited gullies and waterworn gullies are mixed developing,based on high precision remote sensing image data,starting from the morphology of the loess,and assuming that the gullies controlled by the shoulder-line have the role of identifying waterworn gullies,the use of deep neural networks is designed.The network method classifies the 2 gullies.The extraction accuracy of water erosion gullies is higher,reaching 93.05%;the extraction accuracy of inherited gullies is 76.86%.(4)U-Net convolutional neural network has certain advantages in automatic extraction of water erosion gully due to its feature automatic learning characteristics,and the overall classification effect is better.When training a deep convolutional neural network model,the training model can be restricted by transforming terrain knowledge into control conditions acceptable to the machine.It takes the waterworn gullies of the Loess Plateau as the research starting point in the paper.The proposed method of extracting and classifying gullies based on deep convolutional neural network can realize rapid extraction of gully areas.The degree of erosion and development of modern loess landform has a certain indicating effect.