{"title":"Modeling the Flow and Geomorphic Heterogeneity Induced by Salt Marsh Vegetation Patches Based on Convolutional Neural Network UNet-Flow","authors":"Zhipeng Chen, Feng Luo, Ruijie Li, Chi Zhang","doi":"10.1029/2023JF007336","DOIUrl":null,"url":null,"abstract":"<p>The two-way interactions between biological and physical processes, bio-geomorphic feedback, play a vital role in landscape formation and evolution in salt marshes. Patchy vegetation represents a typical form of scale-dependent feedback in salt marshes and is primarily responsible for the formation of efficient drainage networks. The intuitive manifestation of scale-dependent feedback is the heterogeneity of flow and landscape. Process-based modeling is an essential tool for exploring flow heterogeneity, but calculations for small spatial scales and over long time frames can be prohibitively costly. In this study, we proposed a deep learning model architecture, UNet-Flow, based on convolutional neural networks (CNNs), which is used to build a surrogate model to simulate a flow field induced by salt marsh patchy vegetation. After optimizing and evaluating the model, we discovered that UNet-Flow exhibits a speed improvement of over four orders of magnitude compared to single-process simulations using the free surface flow model TELEMAC-2D, with acceptable levels of error. Furthermore, we proposed an approach that combines the process-based model SISYPHE with the deep learning method to model geomorphic heterogeneity. After numerous simulations of flow heterogeneity modeling using UNet-Flow, we obtained a significant logarithmic relationship between scale-dependent feedback strength and vegetation stem density, and an ascending-descending trend in feedback strength was observed as the number or surface area of vegetation patches increased. Finally, we investigated the relationship between geomorphic heterogeneity and vegetation-related variables. This study represents a noteworthy effort to study bio-geomorphology using deep learning methods.</p>","PeriodicalId":15887,"journal":{"name":"Journal of Geophysical Research: Earth Surface","volume":null,"pages":null},"PeriodicalIF":3.5000,"publicationDate":"2024-02-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Geophysical Research: Earth Surface","FirstCategoryId":"89","ListUrlMain":"https://onlinelibrary.wiley.com/doi/10.1029/2023JF007336","RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"GEOSCIENCES, MULTIDISCIPLINARY","Score":null,"Total":0}
引用次数: 0
Abstract
The two-way interactions between biological and physical processes, bio-geomorphic feedback, play a vital role in landscape formation and evolution in salt marshes. Patchy vegetation represents a typical form of scale-dependent feedback in salt marshes and is primarily responsible for the formation of efficient drainage networks. The intuitive manifestation of scale-dependent feedback is the heterogeneity of flow and landscape. Process-based modeling is an essential tool for exploring flow heterogeneity, but calculations for small spatial scales and over long time frames can be prohibitively costly. In this study, we proposed a deep learning model architecture, UNet-Flow, based on convolutional neural networks (CNNs), which is used to build a surrogate model to simulate a flow field induced by salt marsh patchy vegetation. After optimizing and evaluating the model, we discovered that UNet-Flow exhibits a speed improvement of over four orders of magnitude compared to single-process simulations using the free surface flow model TELEMAC-2D, with acceptable levels of error. Furthermore, we proposed an approach that combines the process-based model SISYPHE with the deep learning method to model geomorphic heterogeneity. After numerous simulations of flow heterogeneity modeling using UNet-Flow, we obtained a significant logarithmic relationship between scale-dependent feedback strength and vegetation stem density, and an ascending-descending trend in feedback strength was observed as the number or surface area of vegetation patches increased. Finally, we investigated the relationship between geomorphic heterogeneity and vegetation-related variables. This study represents a noteworthy effort to study bio-geomorphology using deep learning methods.