Research On Deep Segmentation Method Of High-resolution Imagery By Integrating Terrain Features

Geomorphology is one of the basic elements of natural geographical environment.The research on extraction and classification of geomorphology has always been a popular topic in geography and geomorphology,meanwhile,it is one of the core topics in the field of digital terrain analysis.In recent years,the paradigm of object-based image analysis(OBIA)has been gradually adopted by many researchers to improve the accuracy and efficiency of research on classification,extraction and distribution of landform.As the prerequisite and basis of OBIA research paradigm,image segmentation will affect extraction accuracy directly.However,the existing segmentation methods mostly use the spectral information of the image for urban ground objects(buildings,roads,etc.)mapping.This kind of objects usually have more regular and clear boundaries or spectral information with high discrimination.Unlike these urban objects,natural geomorphological entities often have fuzzy and gradual boundary.Therefore,it is difficult to distinguish them directly through images.Especially for mapping micro-landform entities,such as loess sinkholes,loess emphemeral gullies,and etc.Due to the discrete distribution pattern and extremely small occupied area,precise mapping or recognizing natural geomorphological entities put forward higher requirements for image segmentation strategy.Feature extraction based on image segments is another important part of OBIA research paradigm.In the past,spectral features,geometric features and texture features were commonly used for landform mapping and classification.Unfortunately,these features are just based on the feature engineering,which can only represent the shallow basic information of imagery,and lead to the lack of robustness.Deep learning,represented by deep convolutional neural network,has a feature extraction mechanism from shallow to deep,and the extracted features are more robust.Some scholars have tried to extract deep features based on deep learning methods to assist in remote sensing image classification and extraction,and have achieved satisfactory results.If the deep learning method can be used to extract deep terrain features,it is expected to further improve the extraction accuracy.In this paper,we proposed an integrated framework for micro-landform extraction by combining the OBIA paradigm,digital terrain analysis and deep learning method based on high-resolution topographic datasets.The proposed framework included the multi-scale image segmentation,optimization and accuracy assessment,and deep learning-based terrain feature extraction.The main contents and conclusions of this paper are as follows:(1)The basic concept,classification system and basic characteristics of loess micro-landform were analyzed.In this paper,the shortcomings of the existing object-based image analysis paradigm in the extraction of natural geomorphological entities are sorted out.In addition,the commonly used segmentation methods are summarized.From the perspective of graph theory,we reconstructed the similarity operator of the Segmentation by Weighted Aggregation(SWA)based on terrain features.After that,the T-SWA method was then proposed,which can overcome the shortage that the MRS segmentation method requires a lot of parameters.Meanwhile,the evaluation method of segmentation accuracy based on the discrepancy theroy was extended,and a novel index,namely the area-number index(ANI),which combining the geometric descrepancy and quantity descrepancy,was proposed.The segmentation results were then effectively evaluated.The experimental results reflected that the T-SWA segmentation method was superior than MRS in segmentation accuracy.(2)The influencing factors of segmentation optimization and the conditions under which the segmentation object reaches the optimal were analyzed,and found out that the internal heterogeneity of the object is the most important factor for the segmentation optimization.At the same time,it was also found that the unsupervised optimization strategy is relatively more reasonable after summarizing the existing segmentation optimization methods.Furthermore,an unsupervised segmentation optimization method based on watershed extraction was proposed.This method can be used to quantify the inter-segment homogeneity and intra-segment heterogeneity.The experimental results show that the accuracy of the optimized segmentation was superior than that of the original segmentation results.(3)Deep terrain features extracted method was proposed based on deep convoluted neural network.Moreover,the fusion strategy of deep terrain features was also proposed based on the feature pyramid network.This paper also analyzed the influence of the combination of deep features and commonly used shallow features on the micro-landform extraction accuracy.It is found that the combination of spectral features and deep terrain features can obtain the highest accuracy.In addition,the parameters that affect the convolutional neural network were also discussed in this paper.The experimental results demonstrated that too many kernels are not useful to the improvement of extraction accuracy,and the 3-layer and 4-layer pyramids are more suitable for the deep terrain features fusion.(4)In order to verify the effectiveness of the proposed micro-landform extraction framwork,taking loess sinkholes,loess bank gullies and loess emphemeral gullies as examples.Three watersheds of the Loess Plateau were selected as study areas.The results show that the extraction accuracy of loess sinkholes is about 90%,and the extraction accuracy of loess bank gullies and loess emphemeral gullies is about 80%.By integrating terrain features,this paper proposed a set of methods,including segmentation and feature extraction.These methods improves the existing OBIA framework.Besides,the proposed methods can achieve high-precise extraction of typical loess micro-landform.It is a typical case of comprehensive use of digital terrain analysis and object-based image analysis to solve the extraction of natural geomorphological entities.The results can provide technical support for the planning and management of soil and water conservation in the study area.