Digital soil mapping from legacy data and hyperspectral imagery in CapBon (Tunisia) First results and perspectives. Philippe Lagacherie 1, Cécile Gomez 2, Sinan Bacha 4, Rossano Ciampalini 2, Hedi Hamrouni 5, Pascal Monestiez 3 1. INRA LISAH Montpellier 2. IRD LISAH Montpellier 3. INRA BiosP Avignon 4. CNCT Tunis 5. Ministry of agriculture (DG ACTA) Tunis
Two objectives Test a DSM approach that map soil properties from sparse sets of measured soil profiles and globally avalaible soil covariates Applicable in 2012-2015 in many regions of the world (GlobalSoilMap.net) Develop a new DSM approach using a Vis-NIR hyperspectral image Applicable from 2015 in regions with bare or partially vegetated surfaces DIGISOL-HYMED project (2009-2012) Funding : http://www.umr-lisah.fr/digisolhymed GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 2
A Proof-of-concept area : The Cap- Bon Region (2,841 km²) GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 3
Legacy measured soil profiles and control data Soil profiles Control sampling 89 profiles (344 hrz) from a DG/ACTA survey (1973-1979) 91 profiles (345 hrz) from a IAO (Italy) survey (2000) 262 topsoil samples with certified (ISO) soil analysis 2009 2010 2011 1 profile/ 16 km² GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 4
Globally available soil covariates SRTM 90m / ASTER 30m Elevation Flow accumulation Landsat7 band5 Slope Wetnex index Landsat7 band7 Total curvature MRVBF Landsat7 NDVI Profile curvature MRRTF Landsat7 bands 5, 2 GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 5
Aisa-Dual hyperspectral image Image characteristics 338 km², 5 m resolution 450 2500 nm (280 bands) November 2 nd 2010, 10h00-12h30 43.5% of bare soils Spatial resolution degraded to 30 m to mimic near future satellite product (ENMAP, ) GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 6
Some first results DSM approach using sparse sets of soil profiles Detecting, correcting and interpreting the biases of measured soil profile data: A case study in the Cap Bon Region (Tunisia) (Ciampalini et al Geoderma, in revision) Documenting GlobalSoilMap.net grid cells from legacy measured soil profiles and global available covariates (Ciampalini et al accepted in DSM12 proceedings) DSM approach using a Vis-NIR hyperspectral image Using Vis-NIR hyperspectral data to map topsoil properties over bare soils in the Cap Bon region, Tunisia (Gomez et al accepted in DSM12 proceedings) GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 7
Bias detection and correction methods Detecting Biases 1) Creating virtual pairs of samples by simulating a soil property value at legacy soil profiles locations ( ) conditionned by the control sampling ( ) 2) Testing the significance of bias using a paired test (Wilcoxon signed rank-test) 3) Repeat 1) and 2) for n sets of simulations Correcting Biases 1) Compute the interpolation error of the control sampling using as validation data the soil property values of the measured soil profiles 2) Minimize interpolation error by tuning a proportional factor (y = ax) applied to the validation data Variance 280 260 240 220 200 180 160 140 120 100 Correction coefficient (a) VS Variance CLAY 80 0.5 0.7 0.9 1.1 1.3 1.5 1.7 Correction coeff. GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 8
Results (IAO data) Clay Silt Sand CEC OC ph_h2o Frequency of H O rejection 99.0 ** 85.0** 54.0* 66.0* 1.0 96.0** Optimised Correction factor 1.45 0.79 0.81 1.14-1.04 GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 9
DSM approach using sparse sets of soil profiles Detecting, correcting and interpreting the biases of measured soil profile data: A case study in the Cap Bon Region (Tunisia) (Ciampalini et al Geoderma, in revision) Documenting GlobalSoilMap.net grid cells from legacy measured soil profiles and global available covariates (Ciampalini et al accepted in DSM12 proceedings) DSM approach using a Vis-NIR hyperspectral image Using Vis-NIR hyperspectral data to map topsoil properties over bare soils in the Cap Bon region, Tunisia (Gomez et al accepted in DSM12 proceedings) Co-kriging of soil properties with Vis-NIR hyperspectral covariates in the Cap Bon region (Tunisia) (Ciampalini et al submitted in DSM12 proceedings) GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 10
Methods Prediction of soil properties following GSM specifications Clay%, Silt%, Sand%, OC, ph, CEC Depths : 0-5 cm, 5-15 cm, 15-30cm, 30-60 cm, 60-100 cm, 100-200cm Input data 30 m ASTER DEM derived variables : Elevation, Slope, Total Curvature, Profile Curvature, MRVBF, MRRTF, Flow Accumulation, Wetnex Index Landsat 7 TM+, nov 2011 derived variables:, b1 to b7, NDVI, b3/b2, b3/b7, b5/b7, (b5- b2)/(b5+b2) 89 profiles with 344 horizons (DG-ACTA survey), depths harmonisation with equal area spline A spatial soil inference system driven by an exploratory analysis Outputs Is there a spatial structure? Is there any correlated landscape covariate? YES NO YES Regression Kriging Ordinary Kriging NO Regression Means 95% confidence intervals (CI 95% ) of soil property values Proportion of true values in CI 95% obtained by cross validation GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 11
Results Exploratory analysis 10 soil properties were neither correlated with a landscape variable nor spatially structured (mainly ph and OC) no way to predict! Regression-Kriging, Ordinary kriging and Regression were selected by the spatial soil inference system for 16, 4 and 6 soil properties respectively Performances of DSM functions Only a minor part of the soil variability was mapped : between 0 and 38% decrease of CI 95% width Error are slightly underestimated : prop of true values in CI 95% between 85 and 96% Example of map : silt 5-15 cm GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 12
Some first results DSM approach using sparse sets of soil profiles Detecting, correcting and interpreting the biases of measured soil profile data: A case study in the Cap Bon Region (Tunisia) (Ciampalini et al Geoderma, in revision) Documenting GlobalSoilMap.net grid cells from legacy measured soil profiles and global available covariates (Ciampalini et al accepted in DSM12 proceedings) DSM approach using a Vis-NIR hyperspectral image Using Vis-NIR hyperspectral data to map topsoil properties over bare soils in the Cap Bon region, Tunisia (Gomez et al accepted in DSM12 proceedings) Co-kriging of soil properties with Vis-NIR hyperspectral covariates in the Cap Bon region (Tunisia) (Ciampalini et al accepted in DSM12 proceedings) GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 13
Method : Partial Least Square Regression Inputs : AISA spectra bands PLSR Ouputs : 8 soil properties: Clay%, Silt%, Sand%, CEC, ph, CaCO 3, i iron 129 topsoil samples located on bare soils 8 regression models built from cross validation GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 14
Performances of prediction models GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 15
Mapping of topsoil clay content GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 16
Using Vis-NIR hyperspectral images in DSM : ongoing researchs The pedometric way Co-kriging of soil properties with Vis-NIR hyperspectral covariates ( Lagacherie et al, 2012 EJSS, Ciampalini et al, DSM 2012 Sydney) Need to be refined! The signal processing way Using spectral unmixing techniques to predict soil properties over partly vegetated surface ( Ouerghemmi et al, 2011, geoderma) Merging Vis-NIR hyperspectral and legacy data Hyp + Measured legacy soil profiles subsurface soil property prediction Hyp+ Soil maps Extrapolate from bare soil surfaces R²cv = 0.51 Toward a DSM approach for mediterranean areas using hyperspectral satellite imagery GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 17
Lessons for Global soil mapping programs Conversion of legacy data into DSM inputs is not a straightforward step. Pedometric techniques may help to overcome some problems ( filtering errors, data harmonisation) In some regions of the world, processing only legacy soil data and globally available soil covariates may produce uncertain estimations of soil properties Sparse datasets Short range soil variations This however provide a strong rationale to planify new investments in soil data to fulfill user s requirements New covariates like hyperspectral imagery should be considered in the near future at least in the most difficult regions GSP Workshop «toward Global Soil Information» 20-23 March, FAO headquarter Rome (Italy) 18