Sheng-Guo Wang 1* , Morgan Weatherford 2† , LeiLani Paugh 2† , Neil Mastin 2# , John Kirby 2# 1 University of North Carolina – Charlotte, 2 North Carolina Department of Transportation *PI, [email protected], † Str. Com. Chair/Expert, # Manager – 2016 TRB Annual Meeting 01-11-2016 Improvements to NCDOT’s Wetland Prediction Model 2015 AASHTO RAC Awarded “Sweet 16” High Value Research Project NCDOT RP 2013-13 Background Study Area Model Comparison for FW, RCP & SFLT Projects Wetland is important and worth protecting (NEPA 1970) – Need to improve reliability and performance of LiDAR based wetland model ● Part of NEPA Streamlining effort ● Model utilizes digital elevation model and terrain derivatives from bare- earth LiDAR data ● Layers created and analyzed in ArcGIS – Needs for more accurate and efficient method to identify wetlands – Avoid and minimize wetland impacts during project development Variables Table 13 different wetland delineation projects in North Carolina, USA, conducted by NCDOT within three ecoregion groups. Group A: 5 projects in Flatwoods (FW) Group B: 4 projects in Rolling Coastal Plain (RCP) Group C: 4 projects in Southeastern Floodplains and Low Terraces (SFLT) Result Logistic Regression Random Forest (RF) Model Comparison Conclusions 4. Full Automation process* dramatically facilitates the whole wetland prediction process, including automation of (i) Generating wetland variables (ii) Modeling (iii) Prediction (iv) Map display (v) Analysis 5. Beyond cost efficiencies, automated tools provide early awareness of potential wetland impact areas during the project planning. * US Pending Patent “Wetland Modeling and Prediction”, US 14/724,787 (05-28-2015) DISCLAMIER The contents of this poster reflect the views of the authors and do not necessarily reflect the official views or policies of the North Carolina Department of Transportation or the Federal Highway Administration at the time of publication. Acknowledgment Especially thanks to Sandy Smith and Scott Davis at Axiom for their expert field work, and UNCC WAM Research Team for their team work with the PI. Overall Scheme for Automation Random Forest Model and Automation Random Forest (RF) method based on decision trees is applied to identify wetland, which is built by a set of rules with random and optimization. Introduction Methodology true - predict: 1-1 green (wetland), 0-0 grey, 1-0 error red, 0-1 error yellow Field verification Full Automation Process – Wetland Prediction Modules – generate variables, modeling, prediction, post-treatment, display, analysis Generate DEM Derivatives – FHWA: 2011 Environmental Excellence Awards (EEA) to NCDOT and NCDENR for Excellence in Environmental Research: “GIS-based Wetland and Stream Predictive Models” Fig. Comparison of Error rate on all data: red – RF, blue – Logit 1. Raw modeling data 2. Generating Wetland Variables & Table 3. Training Data Set 4. Modeling 5. Models Modeling Prediction 6. Predict area data 7. Prediction Models 8. Wetland Prediction Post- Treatment 9. Post-treatment data 10. Post-treatment 11. Post-treated Wetland Prediction Analysis & Verification 13. Verification 12. Verification areas 14. Analysis Results Data Y N web photo our photo 1. Leads to more informed decisions and high model confidence • Improve protection of state’s natural resources • Additional research project now in place to examine automated defining of wetland types • Significant cost and time saving • Potential saving of $350,000 per project (depending on project size) 2. RF method is proposed for wetland modeling and prediction, and it improves prediction accuracy and outperforms Logit regression method 3. New technology is applied to the wetland identification SC Region 2 Logit Method SC Region 2 RF Method Training areas Prediction Training areas Prediction 80% data for model training, 20% data for test checking, 100% check 1 — 0 error = missing, 0 —1 error = over-estimate U3826 (SFLT) Logit U3826 (SFLT) RF Rowan County, NC U3826 (SFLT) Logit RF update RF Elv Download LiDAR data from NCFM http://www.ncfloodma ps.com/ Mosaic Filter Breach All Slope Curvature Slp Cv Prcv Plcv Asp Curv5 depan Download land cover data From GAP (USGS) Download soil data from NRCS (USDA) Aspect Extract Multi Values to Points Wetness Elevation Index Flow Direction Raw DEM WeiRe Wei Stochastic depression analysis rawda Other Image process flowdr batwi Maximum downslope elevation change DEM2 Tools using ArcGIS Developed Tools Other Processes Intermediate Variable Final Variable Soils GAP Mdec Reclassify Slope Area Ratio Flow Accumulation CA (contributing area) Histogram Projection Interpolate Feature to Raster Reclassify Final Data Table for Modeling Clip Clip Points for both riparian and non- riparian area Raster to Other Format Block Statistics Random Forest Logistic Regression Comparison Type Project Data Total Records 0 – 1 Error Rate 1 – 0 Error Rate Total Error Rate Total Error Rate Total Improve- ment FW B4168 80% 812 0.00% 0.00% 0.00% 3.69% 100.00% 20% 203 3.67% 2.13% 2.96% 2.46% -20.00% 100% 1015 0.77% 0.40% 0.59% 3.45% 82.86% R2514 80% 32827 0.00% 0.00% 0.00% 32.36% 100.00% 20% 8207 22.00% 8.47% 14.09% 32.75% 56.99% 100% 41034 4.43% 1.69% 2.82% 32.44% 91.32% Group from 5 projects 184,262 1.06% 4.75% 2.17% 20.67% 89.51% RCP R2554 80% 31469 0.00% 0.00% 0.00% 12.29% 100.00% 20% 7868 2.33% 6.41% 3.69% 11.64% 68.34% 100% 39337 0.47% 1.29% 0.74% 12.16% 93.91% B3654 80% 626 0.00% 0.00% 0.00% 7.19% 100.00% 20% 157 17.39% 1.49% 3.82% 3.82% 0.00% 100% 783 2.56% 0.32% 0.77% 6.51% 88.24% Group from 4 projects 104,731 0.72% 0.68% 0.70% 10.34% 93.21% SFLT B4135 80% 1769 0.00% 0.00% 0.00% 1.87% 100.00% 20% 443 1.89% 2.11% 2.03% 2.26% 10.00% 100% 2212 0.37% 0.43% 0.41% 1.94% 79.07% U3826 80% 20172 0.00% 0.00% 0.00% 11.45% 100.00% 20% 5044 3.51% 5.69% 4.48% 11.30% 60.35% 100% 25216 0.71% 1.13% 0.90% 11.42% 92.12% Group from 4 projects 30,588 0.74% 0.97% 0.85% 10.17% 91.61% Variable Full Name Formula and Illustrations elev Elevation Elevation of each cell: z(x, y) asp Aspect asp = 57.29578 * atan2 ([dz/dy], -[dz/dx]) slp Slope slp(x, y) = 57.29578 × atan( / 2 + / 2 ) cv Curvature Slope of the slope: cv = 57.29578 × atan( / 2 + / 2 ) prcv Profile Curvature Curvature on vertical (y) direction plcv Plan Curvature Curvature on horizontal (x) direction batwi Ratio of Slope by Drainage Area batwi = slp / drainage contributing area (calculated with breached DEM) wei Wetness Elevation Index Series of increasingly larger neighborhoods used to determine the relative landscape position of each cell. weiRe Reclassification of wei Wei value of each cell will be reclassified as 0 if original value is bigger than a predefined threshold, else is reclassified as 1. mdec Maximum Downslope Elevation Change Maximum difference of z(x,y) between the cell and its neighbor cells. rawda Stochastic Depression Analysis Stochastic depression analysis based on raw DEM. curv5 Smooth Curvature Each cell gets mean value of curvature from its 5*5 neighbors. 5 = =1 25 ()/25 depan Stochastic Depression Analysis Stochastic depression analysis based on breach-all DEM. gap Land Cover Data Categorized land use types. soil Soil Data Reclassified as 1 or 0 to indicate hydric or non-hydric soil type.