development of a soil carbon map for the united republic of tanzania
TRANSCRIPT
Soil Organic Carbon Mapping Seminar
Rome, 23 November 2016
Development of a soil carbon map for the United
Republic of Tanzania
Bas Kempen in collaboration with TFS (MNRT), SUA, NSS (MoA), ZAFORMA, AfSIS, FAO-TZ
Project overview
• Carbon concentration: 0-10 cm, 10-20 cm, 20-30 cm, 0-30 cm (Walkley-Black)
• Bulk density & Carbon stock: 0-30 cm
• Regression-kriging
• Raster with 250 m spatial resolution
• Uncertainty: 90% Prediction Interval
• GeoTIFF format with xml metadata
• Hands-on training digital soil mapping (Wageningen)
• Hands-on validation workshop (Dar es Salaam)
Final report
Literate programming: R / LaTeX
• Documentation and computer script (R code) weaved in 1 document (PDF report).
• Transparent and reproducible.
trend, explanatory part
𝑍 𝐬 = 𝑚 𝐬 + 𝜀(𝐬)
dependent, target variable
stochastic residual, unexplanatory
part, can be spatially correlated!
Unlike ordinary kriging, in regression kriging the trend is no longer constant but a function of ’explanatory’ variables, for example:
𝑠𝑜𝑖𝑙 𝑐𝑎𝑟𝑏𝑜𝑛 𝐬 = 𝛽0 + 𝛽1 ∙ 𝑒𝑙𝑒𝑣𝑎𝑡𝑖𝑜𝑛 𝐬 + 𝛽2 ∙ 𝑠𝑙𝑜𝑝𝑒 𝐬 + 𝛽3 ∙ 𝑁𝐷𝑉𝐼 𝐬 + 𝑟𝑒𝑠𝑖𝑑𝑢𝑎𝑙 𝐬
Regression kriging
Regression kriging algorithm
1. Select explanatory variables and fit regression model
2. Compute residuals (fitted trend – observed values) at observation locations
3. Compute semivariogram (quantify the spatial correlation)
4. Apply the regression model to all unobserved locations (usually a grid)
5. Krige the residuals to all unobserved locations
6. Add up the results of steps 4 and 5
Soil point data
• NAFORMA: 1409
• NSS: 306
• AfSIS: 975
• AfSP v1.1: 525 (>1992)
• 3215 observations
Four depth layers
• 0-10 cm: 3,215; 13.2
• 10-20 cm: 3,187; 11.0
• 20-30 cm: 2,839; 7.8
• 0-30 cm: 2,926; 10.7
Environmental covariate layers
• 37 layers from:
- AfSIS Grids – MODIS: 24
- AfSIS Grids – SRTM: 3
- WorldGrids.org: 4
- SOTER map (SOTERSAF): 2
- e-SOTER: 3
- NAFORMA: 1
• Raster: 31x (250 m, 500 m, 1000 m)
• Polygon: 6x
Land cover
• NAFORMA LULC map (incl. Unguja , Pemba)
• ESA GlobCover v2.2 2005, 1000 m (WorldGrids)
- % mosaic vegetation
- % shrubland
MODIS Imagery: Vegetation (AfSIS grids)
1000 m
• FPAR
• LAI
• NPP (’00/’10)
• GPP (’00/’10)
250 m
• EVI
• NDVI
MODIS Imagery: Reflactance (AfSIS grids)
250 m
• Band1-R
• Band2-NIR
• Band3-B
• Band7-MIR
Yearly average
September
Model selection: selected model
• Explained variation (R2)
- CARBON:
• 0-10 cm: 0.36
• 10-20 cm: 0.39
• 20-30 cm: 0.36
• 0-30 cm: 0.45
- BULK DENSITY: 0.42
- SOC STOCK: 0.40
Spatial prediction
• Prepare a prediction grid of 250 m resolution (based on a MODIS image): >14 Million prediction points
• Sample covariate layers
• Computational intensive: prediction for two tiles of 7M points each requires >20 GB RAM memory
• Apply the regression-kriging algorithm
• Compute the uncertainty (90% prediction interval)
• Monte Carlo analysis to quantify uncertainty SOC stock
SOC stock
• Average: 4.1 kg/m2
• Total: 3.60 Pg
Uncertainty (95% CI):
4.0 - 4.15 kg/m2
.54 - 3.65 Pg
SOC stock uncertainty
• Average: 4.1 kg/m2
• Total: 3.60 Pg
• Uncertainty (95% CI):
- 4.0 - 4.15 kg/m2
- 3.54 - 3.65 Pg
Validation I
• 10-fold cross-validation
• Mean-error (bias)
• Mean absolute error / Root mean square error (accuracy)
Layer ME MAE RMSE RMedSE R2-RK R2-R
SOC concentration (g/kg)
0-10 cm -0.95 5.4 8.9 3.5 0.47 0.36
10-20 cm -0.66 3.8 6.2 2.3 0.49 0.39
20-30 cm -0.62 3.1 5.4 1.9 0.44 0.36
0-30 cm -0.44 3.4 5.4 2.1 0.59 0.45
Bulk density (g/cm3)
0-30 cm 0.00 0.1 0.14 0.08 0.56 0.42
SOC stock (kg/m2)
0-30 cm 0.00 1.23 1.8 0.9 0.53 0.40