Adjusting Emergent Herbaceous Wetland Elevation with Object-Based Image Analysis, Random Forest and the 2016 NLCD

Abstract

Emergent herbaceous wetlands are characterized by complex salt marsh ecosystems that play a key role in diverse coastal processes including carbon storage, nutrient cycling, flood attenuation and shoreline protection. Surface elevation characterization and spatiotemporal distribution of these ecosystems are commonly obtained from LiDAR measurements as this low-cost airborne technique has a wide range of applicability and usefulness in coastal environments. LiDAR techniques, despite significant advantages, show poor performance in generation of digital elevation models (DEMs) in tidal salt marshes due to large vertical errors. In this study, we present a methodology to (i) update emergent herbaceous wetlands (i.e., the ones delineated in the 2016 National Land Cover Database) to present-day conditions; and (ii) automate salt marsh elevation correction in estuarine systems. We integrate object-based image analysis and random forest technique with surface reflectance Landsat imagery to map three emergent U.S. wetlands in Weeks Bay, Alabama, Savannah Estuary, Georgia and Fire Island, New York. Conducting a hyperparameter tuning of random forest and following a hierarchical approach with three nomenclature levels for land cover classification, we are able to better map wetlands and improve overall accuracies in Weeks Bay (0.91), Savannah Estuary (0.97) and Fire Island (0.95). We then develop a tool in ArcGIS to automate salt marsh elevation correction. We use this ‘DEM-correction’ tool to modify an existing DEM (model input) with the calculated elevation correction over salt marsh regions. Our method and tool are validated with real-time kinematic elevation data and helps correct overestimated salt marsh elevation up to 0.50 m in the studied estuaries. The proposed tool can be easily adapted to different vegetation species in wetlands, and thus help provide accurate DEMs for flood inundation mapping in estuarine systems.

Compound Effects of Flood Drivers and Wetland Elevation Correction on Coastal Flood Hazard Assessment

“…The Savannah DEM is obtained from a former study in which a DEM‐correction tool was developed to adjust wetland surface elevation in the estuary and help correct a maximum overestimation of 0.25 m in the existing lidar‐derived DEM (Muñoz et al, 2019). The DEM‐correction tool is based on a simple yet efficient algorithm that identifies and updates “emergent herbaceous wetlands” to present‐day conditions with remotely sensed data and random forest technique.…”

Fusing Multisource Data to Estimate the Effects of Urbanization, Sea Level Rise, and Hurricane Impacts on Long-Term Wetland Change Dynamics

“…Several studies have reported that LiDAR-derived DEMs contain surface elevation errors (i.e., vertical bias) in coastal wetlands [41]- [43], especially in salt marsh species where overestimation of true elevation can be as high as 0.65 m [44]. Similarly, Muñoz et al, [24] reported an overestimation of salt marsh elevation (e.g., emergent herbaceous wetlands) up to 0.50 m in Weeks Bay of the NGOM DEM. The authors estimated vertical bias with a 'DEM-correction' tool that automates the elevation correction process based on the spatial distribution of emergent herbaceous wetlands from C-CAP, NLCD and updated wetland maps.…”

“…Significant improvements in LCLUC assessment have been reported when integrating ML with hierarchical classification strategies and object-based image analysis [23], [24]. However, deep convolutional neural networks (CNNs) outperform traditional (shallow) ML approaches (e.g., random forest, support vector machine and decision tree) in a variety of applications including object detection, segmentation and spatial structure pattern analyses [25], [26].…”

“…We combine multi-annual imagery and generic DEMs for wetland cover change analysis in the Mobile Bay watershed, AL. Specifically, we use readily available datasets consisting of (i) satellite-based Landsat imagery, (ii) airborne imagery of the National Agriculture Imagery Program (NAIP), and (iii) LiDAR-derived DEMs corrected for wetland elevation error [24]. The LCC model is trained with land cover maps of the National Land Cover Database (NLCD) [34] and validated with maps of both the NLCD and the Coastal-Change Analysis Program (C-CAP) [35].…”

Inland Lakes Mapping for Monitoring Water Quality Using a Detail/Smoothing-Balanced Conditional Random Field Based on Landsat-8/Levels Data

“…For the image classification problem, the RF is not the best-performing algorithm. However, due to its simplicity, ease of implementation, strong generalization ability, and good performance on many datasets, it has been widely used in academic research and industrial applications [39,40]. The RF is an algorithm that integrates multiple trees through the idea of ensemble learning.…”