National Cooperative Soil Survey Conference 2007 Madison, WI Soil Spectroscopy for Rapid and Cost-Effective Soil Mapping Across Larger Landscapes Sabine Grunwald NRCS-CESU (Carolyn G. Olson) “Linking Experimental and Soil Spectral Sensing for Prediction of Soil Carbon Pools and Carbon Sequestration at Landscape Scale” Investigators: Grunwald S., J.O. Sickman and N.B. Comerford Graduate student: Gustavo M. Vasques
NASA Soils store ~ 3 times more C than biosphere (vegetation) ~ 2 times more C than atmosphere ~ 1.5 time more C than surface ocean
NASA Temp.
+ Environmental landscape factors Indicators of locally operating ecosystem processes C, N and P mineralizable pools Soil Carbon Sequestration 1 Carbon (C) pools: Total Recalcitrant Labile Anthropogenic and natural forcing functions Transfer of total, recalcitrant and labile C imprints into landscapes elucidates on mechanisms that induce C storage/ change Assess the actual pools and potential C sequestration potential
Soil Mapping 1 Accurate Rapid Cost-effective Large soil regions Relevant soil properties
StatisticsGIScience Soil & Environmental Sciences Cartography Quantitative methods Geo- statistics Digital Soil Mapping 1 GIS - multi-scale data integration Complex geospatial methods Environmental datasets: Field and analytical data Soil sensing Remote sensing Soil-landscape models: Functional (stochastic; deterministic) Mechanistic (simulation) Grunwald S. (ed) Environmental Soil-Landscape Modeling–Geographic Information Technologies & Pedometrics. CRC Press, New York.
Visible/near-infrared spectroscopy (VNIRS) is a fast, cheap and accurate alternative for the investigation of soil properties, and is now recognized as a powerful analytical tool in soil science Soil Sensing 2 Each soil has specific reflectance signature Modern soil surveying
Spectroradiometer 2 QualitySpec® Pro (Analytical Spectral Devices Inc., Boulder, CO)
TC: 7,132 mg kg -1 (avg.) Pine plantation Typic Aquods TC: 268,995 mg kg -1 Wetland Typic Argiaquolls cm cm cm Visible/Near-infrared Diffuse Reflectance Spectroscopy 2 Total carbon (TC): 169 mg kg -1 Upland forest Typic Quartzipsamments
Soil Study – Santa Fe River Watershed, Florida 3 Objectives: Investigate the usefulness of VNIRS for rapid and accurate assessment of soil carbon Understand the linkages between labile, recalcitrant (stable) and total organic carbon Assess the usefulness of VNIRS to map larger soil- landscapes
Data sources maps: DEM: National Elevation Model (US Geological Service) Land use: Florida Fish and Wildlife Conservation Commission (2003) Geology: FL Dept. of Environmental Protection Soil Orders: Soil Survey Geographic Database (SSURGO) Natural Resources Conservation Service Santa Fe River Watershed, Florida 3
Methodology 3 Laboratory soil data VNIR spectral data Pre-treatment: Log-normalization using base-10 logarithm Testing of 30 different preprocessing transformations Identify relationships Complete dataset Methods: Stepwise Multiple Linear Regression (SMLR) Principal Components Regression (PCR) Partial Least-Squares Regression (PLSR) Regression Tree (RT) Committee Trees (CT) (bagging) ~70% of data Model dataset ~ 30% of data used to test accuracy of model predictions Validation dataset predictions R 2, RMSE
Total Carbon (TC) 3 Sampling across land use- soil order trajectories
Layer 1 (0-30 cm) Layer 2 (30-60 cm) Layer 3 ( cm) Layer 4 ( cm) Observations (n) Mean14,8728,1053,9291,659 Std. Error of Mean1,8282,1271, Median10,5293,7051,8081,087 Mode2, Std. Deviation21,86725,43413,1982,117 Skewness Kurtosis Range199,318268,062112,72518,749 Minimum2, Maximum201,988268,995113,10918,917 Total Carbon (TC) [mg/kg] 3
Spectral scans of 554 soil samples collected in the SFRW at 4 different soil depths (0-30, 30-60, and cm) VNIR Scanning 3
Results: Prediction Performance - logTC [mg/kg] 4 Calibration Validation R 2 RMSER 2 RMSE SMLR PCR PLSR RT CT [30 pre-processing methods were tested]
PCR Validation Results: Prediction Performance - logTC [mg/kg] 4 SMLR [pre-processing: Savitzky-Golay 1st-derivative using a 1st-order polynomial with search window 9 (SGF-1-9) [pre-processing: standard normal variate transformation (SNV)] Laboratory measurements VNIR Predictions
PLSRRT Validation Results: Prediction Performance - logTC [mg/kg] 4 [pre-processing: Savitzky-Golay 1st-derivative using a 3rd-order polynomial with search window of 9 (SGF-3-9)] [pre-processing: Norris gap derivative with a search window of 5 (NGD-5)]
CT Validation Results: Prediction Performance - logTC [mg/kg]4 [pre-processing: Norris gap derivative with a search window of 7 (NGD-7)]
(mg/kg) TOCHCRCDOC07DOC02 Observations141 Minimum 2,670371, Maximum 201,98829,399181,7389,0008,995 Median 10,5292,8927, Mean 14,8283,70711, Std. Deviation 21,9933,29219, Total Organic Carbon and Carbon Fractions (0-30 cm) 5 TOC: Total organic carbon HC: Hydrolysable carbon (after digestion with 6N HCl) - Thermo Electron FlashEA Elemental Analyzer RC: Recalcitrant carbon was calculated as the difference between TOC and HC DOC: Dissolved organic carbon Shimadzu TOC Analyzer after hot water extraction, then filtered into 2 classes: <0.7 µm (DOC07) and <0.2 µm (DOC02).
SOC and fractions Best model CalibrationValidation Rc2Rc2 RMSE C Rv2Rv2 RMSE V TOCLOG-PLSR HCSAV-PLSR RCSAV-PLSR DOC07SAV-PLSR DOC02SNV-PLSR Statistics – Total Organic Carbon and Carbon Fractions 5 LOG: Log (1/Reflectance) SAV: Savitzky-Golay smoothing, and averaging SNV: Standard normal variate transformation
VNIRS vs. Conventional Lab Analysis 6 CharacteristicsVNIRSConventional Ease of sample preparation ++++ Ease of analysis++ Speed++++ Labor+++++ Equipments cost++ Use of supplies++++ Cost per sample++++ Accuracy+++
VNIRS & Landscape Scale Modeling 7 Organic matter (OM) lab measurements 0-30 cm OM predictions using VNIR spectral data method: committee trees (boosting) 0-30 cm Calibr.Validation R RMSE
VNIRS & Landscape Scale Modeling7 Semivariograms of OM measurements and VNIRS predictions show very similar spatial autocorrelation structure OM lab measurements OM derived from VNIRS
VNIRS & Landscape Scale Modeling7 OM map derived using lab measurements OM map derived from VNIR spectra
ErrorSandSiltClay Mean Min Max RMSE Lamsal S., PhD thesis Soil Texture Geospatial Modeling - SFRW 7 Soil 0-30 cm Method: Spatial stochastic simulation
National and Global VNIRS Applications8 Locations / soil-landscape settings: Africa (Shepherd and Walsh, 2002) Australia (Dalal and Henry, 1980) Australia (Viscarra Rossel et al., 2006) Brazil (Masserschmidt et al., 1999) Netherlands (Koistra et al., 2003) USA, Maryland (Reeves III et al., 2001) Global (USA & Africa) (Brown et al., 2006) …. many more Soil properties: Carbon; organic matter Texture Nutrients (N, P, Mg, Ca, …) Metals (Fe, Al,…..) CEC BD …. many more VNIRS
Conclusions8 NASA Soil mapping & expertise Soil and remote sensors incl. VNIRS Soils – environmental factors (GIS; soil-landscape analysis) Soil science - global context