Effects of LiDAR DEM Smoothing and Conditioning Techniques on a Topography‐Based Wetland Identification Model. Issue 5 (27th May 2019)

• Record Type: Journal Article
• Authors: O'Neil, Gina L.; Saby, Linnea; Band, Lawrence E.; Goodall, Jonathan L.
• Is Part Of: Water resources research. Volume 55:Issue 5(2019)
• Page Start: 4343
• Page End: 4363
• Publication Date: 2019-05-27
• Subjects: wetland -- Random Forests -- LiDAR -- hydroconditioning -- smoothing; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2019WR024784
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: Accurate and widely available wetland inventories are needed for wetland conservation and environmental planning. We propose an open source, automated wetland identification model that relies primarily on light detection and ranging (LiDAR) digital elevation models (DEMs). LiDAR DEMs are increasingly available and provide the resolution needed to map detailed topographic metrics and areas of likely soil saturation, but the choice of smoothing and conditioning techniques can significantly impact the accuracy of hydrologic parameter extraction. So far, the effect of these preprocessing steps on wetland delineation has not been thoroughly analyzed. We test the response of a Random Forest wetland classifier, using topographic wetness index, curvature, and cartographic depth‐to‐water index as input variables, to combinations of smoothing techniques (none, mean, median, Gaussian, and Perona‐Malik) and conditioning techniques (Fill, Impact Reduction Approach, and A* least‐cost path analysis) for four sites in Virginia, USA. The Random Forest model was configured to account for imbalanced data sets, and manually surveyed wetlands were used for verification. Applying Perona‐Malik smoothing and A* conditioning yielded the highest accuracy across all sites and considerably reduced model runtime. We found that models could be further improved by individualizing the smoothing method and scale to each input variable. Using only topographic information, the wetland identification model could accurately detect wetlands in all sites (81‐91% recall). Model overprediction varied across sites, represented by precision scores ranging from 22 to 69%. In its current form, the wetland model shows strong potential to support wetland field surveying by identifying likely wetland areas. Plain Language Summary: Accurate wetland inventories are needed for wetland protection and conservation. We propose an automated tool that locates wetlands using light detection and ranging (LiDAR) digital elevation models (DEMs). LiDAR DEMs are increasingly available and show elevation changes that likely affect soil saturation. However, the ability of LiDAR DEMs to describe saturated areas is affected by smoothing and conditioning. Smoothing blurs DEMs to remove elevation changes that are too small to indicate features of interest, and conditioning ensures accurate simulation of hydrologic flow paths. The effects of different smoothing and conditioning methods on wetland mapping have not been studied. We tested how our wetland tool is influenced by five smoothing techniques and three conditioning techniques for four sites in Virginia, USA. We found that Perona‐Malik smoothing and A* conditioning improved predictions and reduced tool runtime for all sites. Also, we found that predictions could be further improved by varying smoothing parameters specific to each input. Using only elevation information, the wetland tool predicted 81‐91% of true wetlands across our sites. The proportion of wetland predictions that were correct varied (ranging from 22 to 69% across sites). Overall, the results suggest strong potential for the model to support environmental groups to delineate wetlands. Key Points: For four sites, we tested the effects of terrain preprocessing on a Random Forest model that uses LiDAR to delineate wetlands Perona‐Malik smoothing and A* conditioning performed best in all sites, and models further improved by individualizing smoothing by input For all sites, the model detected most wetlands (81‐91%) but with varying precision (22‐69%), indicating its best use as a screening tool