SWAT 2015 International Conference: Comparative Analysis of Spatial Resolution Effects on Standard and Grid-based SWAT Models Presented by: Garett Pignotti Co-authors: Dr. Hendrik Rathjens, Dr. Cibin Raj, Vamsi Vema, Dr. Indrajeet Chaubey, & Dr. Melba Crawford 1
2 Outline I. INTRODUCTION II. BASELINE MODEL CALIBRATION Resolution III. RESOLUTION EFFECTS ON: i. SIMULATIONS ii. CALIBRATION IV. CONCLUSIONS & FUTURE EFFORTS Simulations Calibration Discretization Routing
Landscape Representation in SWAT Hydrologic Response Unit = HRU 3
4 SWATgrid Gridded Input Modified Routing SWATgrid Raster-based Delineation = Grid-based Interface (Rathjens & Oppelt 2012) Topaz (Garbrecht & Martz 1997) Landscape routing with SWATgrid (Rathjens et al. 2014) Landscape routing (Arnold et al. 2010)
5 Watershed Routing Standard Landscape
6 Research Questions & Objectives 1. What is the effect of resolution on output of both models? Identify a resolution for SWATgrid that both maximizes prediction accuracy while minimizing computation time 2. How do simulations of SWATgrid compare to standard HRU implementation? Compare model output 3. How do calibration parameters change with respect to resolution? Discriminate parameters sensitive to resolution change
7 Methods Study Site Cedar Creek Watershed 700 km 2
8 Methods Study Site DATA SOURCES: ELEVATION (30 M NED) LAND COVER (30 M NASS CDL) SOILS (250 M STATSGO) RESAMPLING: NEAREST NEIGHBOR (LU & SOIL) CUBIC CONVL. (DEM)
BASELINE MODEL CALIBRATION 9
10 Methods Resolution Effects 30 m Input Data 30 m Calibrated HRU Model AMALGAM (Vrugt et al. 2007)
11 Methods SWAT Modeling Management Strategy Jan 15 - Apr 22 N App. 1 Apr 22 Atrazine 2 Corn Year Soybean Year May 6 Cultivator May 6 Planting Jun 6 N App. 3 Oct 14 Harvesting Oct 15 Killing May 24 Zero Till May 24 Planting Oct 7 Harvesting Oct 8 Killing Oct 15 P App. 4 Nov 1 Chisel 1: Anhydrous of 53 kg/ha (N of 43 kg/ha); 2: Atrazine of 2.2 kg/ha; 3: Urea of 284 kg/ha (N of 131 kg/ha); 4: DAP (P 2 O 5 ) of 123 kg/ha (P of 54 kg/ha). Dominant management practices Tile Drainage Model Setup 1990 2010 Warm up: 1990-1993 Calibration: 1994-2003 Validation: 2004-2010
12 Basin Level Parameters Parameter Definition Units Range Default Calibrated SFTMP snowfall temperature o C -5-5 1.00 0.18 SURLAG surf. runoff lag coefficient day 0.5-2 4.00 1.00 SMTMP snow melt base temperature o C -5-5 0.50 2.46 TIMP snow pack temp. lag factor - 0.01-1 1.00 1.00 SMFMX SMFMN maximum melt factor minimum melt factor mm H 2 0/ o C mm H 2 0/ o C 1-10 4.50 6.54 1-10 4.50 1.00
13 Calibration Results Nash-Sutcliffe Efficiency: USGS 04180000 - Cedar Creek near Cedarville, n IN Obs i Sim 2 i NSE = 1 i=1 n i=1 Obs i Sim 2 Calibration Validation Monthly NSE 0.80 0.80 Obs. Mean (cms) 7.48 8.67 Sim. Mean (cms) 7.13 8.66 800 600 400 200 Monthly Flow (cms)1000 Observed Simulated 0
SIMULATION EFFECTS 14
Methods Simulation Effects 30 m Input Data 30 m Calibrated HRU Model Parameter Transfer Same Version (rev. 574) HRU model AMALGAM (Vrugt et al. 2007) GRID model Landscape & Standard Resolutions tested: 30, 60, 90, 150, 250, 500, & 1,000 m 15
Watershed Area Tends to Decrease with Increasing Resolution Watershed Area (km 2 ) 16 30 m 60 m 90 m 150 m 800 250m 700 500 m 600 1000 m 500 400 HRU GRID 300 30 60 90 150 250 500 1000 Resolution (m)
Watershed Area Land Use & Soil Distributions by % Area Remain Relatively Constant 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% Corn Soybean IN004 General Agr. IN005 Barren IN007 Range/Brush IN016 Pasture IN019 Forest IN025 Urban IN028 Wetland Water 0% 30 60 60 90 90 150 150 250 250 500 500 1000 1000 Input Input Data Data Resolution (m) (m) 17
21 Year Total Simulation Time (mins) Simulation Time Exponentially Increases with Finer Resolution Models 18 10000 1000 Intel Core i5-2400 CPU @ 3.10GHz 4 GB RAM 30 & 60 m GRID Models Not Possible 86,635 cells 27,253 cells 100 11,916 cells 10 2,451 cells 385 cells 1 GRID1000 GRID500 GRID250 GRID150 GRID90
Monthly Flow (cms) Monthly Flow at the Outlet is Under Predicted by the GRID Models 19 1000 HRU150 GRID-LAND150 800 GRID-STD150 600 400 200 0
0.5 3.4 6.3 9.3 12.2 15.1 18.0 21.0 23.9 26.8 29.8 32.7 35.6 38.5 41.5 44.4 47.3 50.2 53.2 56.1 59.0 62.0 64.9 67.8 70.7 73.7 76.6 79.5 82.4 85.4 88.3 91.2 94.1 97.1 Monthly Flow (cms) GRID Models Under Predict High Flows 1500 1000 500 0 1500 1000 500 0 1500 1000 HRU GRID - LAND GRID - STD Observed HRU 30 HRU 60 HRU 90 HRU 150 HRU 250 HRU 500 HRU 1000 Observed GRID 90 GRID 150 GRID 250 GRID 500 GRID 1000 500 0 Exceedance Probability 20
Avg. Annual Flow (cms) % Relative Error Average Annual Flow Follows Similar Trend but Different in Magnitude 60 HRU 50 GRID-LAND 40 GRID-STD 30 20 10 0-10 -20 30 60 90 150 250 500 1000 Resolution (m) 21
Simulate Monthly Avg. Flow NSE (cms) Watershed Area Impacts Predictions 0.9 10 HRU 0.8 9 HRU GRID-LAND 8 0.7 GRID-STD 7 0.6 6 0.5 5 0.4 4 0.3 3 0.2 2 0.1 1 0 300 400 500 600 700 800 Watershed Area(km 2 ) 2 ) 22
Subsurface Flows are Divergent between Models (150 m Comparison) 23 Relative Average Diff. Average Annual Basin Annual Values Basin (mm) Values (mm) -2000-150 200-100 400 600-50 8000 1000 50 PRECIP SNOW FALL SNOW MELT SUBLIMATION SURFACE RUNOFF Q LATERAL SOIL Q TILE Q GROUNDWATER (SHAL AQ) Q GROUNDWATER (DEEP AQ) Q REVAP (SHAL AQ > SOIL/PLANTS) DEEP AQ RECHARGE TOTAL AQ RECHARGE TOTAL WATER YLD PERCOLATION OUT OF SOIL ET PET TRANSMISSION LOSSES GRID-LAND GRID-STD HRU GRID-LAND GRID-STD
GRID Models Capture Spatial Hydrology 24
CALIBRATION EFFECTS 25
26 Methods Calibration Effects Input Data Unique Parameters HRU model AMALGAM (Vrugt et al. 2007) HRU model HRU model
Converged Monthly Monthly NSE NSE Optimized Objective Function Value Decreases with Coarser Resolutions 0.80.9 0.75 0.70.8 0.65 0.60.7 0.55 0.6 0.5 0.45 0.5 0.4 0.350.4 0.3 0.250.3 0.2 0.150.2 0.1 0.1 0.05 1 1001 2001 3001 4001 0 30 60 90 Simulation 150 250 500 1000 Resolution (m) 30 m 90 m 60 m 150 m 250 m 500 m 1000 m 27
Parameter Distributions Similar Up to 90 m 28
29 Summary & Concluding Remarks I. Lower flow for GRID models relative to calibrated HRU: II. Discretization & routing effects Implicit restriction for simulations: Number of grids = f(input resolution, watershed area) III. Potentially possible to scale to 90 m in this study IV. Use of SWATgrid best for specific applications FUTURE EFFORTS I. More rigorous analysis of discrepancies between models II. Test sensitivity of flow separation index III. Test in other watersheds
Thank You! https://engineering.purdue.edu/ecohydrology/ 30