Research On The Identification Of Typical Landform Types In The Loess Plateau Region Of Northern Shaanxi Based On Deep Learning

Geomorphology is one of the basic elements of the natural geographic environment.The extraction,identification,and classification of natural geomorphology has always been a research hotspot in geography,and it is also one of the core topics in the field of digital terrain analysis.On the basis of inheriting the underlying ancient topography,the Loess Plateau in my country has undergone more than 2 million years of loess transportation,accumulation,and long-lasting and strong interaction of modern water erosion,forming a complex and diverse landform pattern with thousands of ravines,fragmented,and undulating.The study of these differences in geomorphological morphology is not only the basis and key to the exploration of the spatial pattern,genetic mechanism,evolution process,and other geomorphological origin issues in the study area,but also an important entry point for the regional planning research of the Loess Plateau and the formulation of scientific measures such as soil erosion prevention and control.Scholars in the field have proposed a variety of automatic identification methods for many years,but the research mostly stays at the shallow terrain information mining based on DEM data sources and its derivatives,and the robustness is insufficient.With the widespread application of artificial intelligence technology,deep learning represented by deep convolutional neural networks has a feature information mining mechanism from shallow to deep,and the results obtained are more robust.How to use deep learning methods to deeply excavate various topographic features at the regional scale and realize automatic recognition of geomorphic types is particularly important for improving the existing digital terrain and landform analysis technology.Based on ASTER GDEM 30m data,this study takes 8 typical landforms in the Loess Plateau of northern Shaanxi as research cases,selects 19 training areas and 14 testing areas,and integrates digital terrain and deep learning methods to construct an automated geomorphological recognition network,including geomorphological feature extracting and geomorphological feature fusion.The main contents and conclusions of this research are as follows:(1)Determine the identification indexes of geomorphic morphological types based on DEM,and conduct preliminary selection of topographic factors from the macro and micro levels respectively,and on this basis,conduct quantitative analysis such as correlation and differentiability.Quantitative analysis results determine the five indicators of elevation,slope,topographic undulation,surface cutting depth,and mountain shadow to participate in the subsequent deep learning automatic recognition experiment of landform morphology types.(2)A deep learning geomorphological recogniticon network(ResNet-UNet)is designed,which consists of three parts:topographic feature extraction,geomorphic feature fusion and geomorphic recognition to meet the requirements of automatic recognition of typical landform types on a regional scale.The results show that,compared with the existing mainstream deep learning networks,the ResNet-UNet network performs better on the test set,followed by U-net and Linknet,with accuracy of 0.87,0.86,and 0.62,respectively.The network structure not only avoids the problems of network degradation and gradient explosion caused by deepening the number of deep convolutional layers,but also can obtain accurate pixel-level segmentation results with good robustness.(3)Analyze the exploration process of multi-modal and multi-source geomorphic data fusion,and verify the performance of geomorphic recognition network constructed based on the deep learning method.The experimental results show that the geomorphic data fusion based on multi-modes has better recognition accuracy.The recognition accuracy of hillshade+elevation+slope fusion on the training set is 95%,and the accuracy on the test set is 87%,indicating that the expression of the internal physical attributes and external visual attributes of the terrain are complementary,which can better express the surface morphology.The recognition accuracy of geomorphic data with more connotation geomorphic development and evolution process is not significantly improved,but the data volume and the computational complexity of the model are both increased.The high resolution remote sensing image data is more suitable for microtopography classification and recognition research.