What Are The Drone Anti-jamming Systems Technology?
Global Soil Mapping Proposal for Multiscale Participatory Approach
1. Global Soil Mapping
A proposal for a participatory multiscale
approach to GSM
Tomislav Hengl
ISRIC World Soil Information, Wageningen University
GlobalSoilMap.net presentation, 11 Feb 2011
2. Outline
Introduction
This talk
My backgrounds
GlobalSoilMap.net
Misconceptions about DSM/GSM
Mapping eciency
Soil geodata usability
Soil prediction methods
A proposal for GSM
Global Soil Mapping is not trivial
Nested regression modeling
The participatory approach
Malawi show case
Input data
Results
Summary points
GlobalSoilMap.net presentation, 11 Feb 2011
3. Topics
My backgrounds;
Some misconceptions about DSM/GSM;
A proposal for GSM:
A Global Multiscale Prediction Model
The crowd-sourcing approach to soil data collection (Open
Soil Proles, Soil covariates)
Global task-oriented Land (Soil) Information System
Report on the results ( Malawi show case).
Get some feedback.
GlobalSoilMap.net presentation, 11 Feb 2011
4. Previous projects
My expertise: spatio-temporal data analysis in FOSS (R),
digital soil mapping, geomorphometry, geostatistics. . .
I have worked with various type of data
(climatic/meteorological, species occurrence records,
geochemicals. . .);
Recently published a repository of cca 100 global layers at
resolution of 0.05 arcdegrees (5.6 km).
Author of A Practical Guide to Geostatistical Mapping .
Main organizer of the GEOSTAT summer school for PhD
students (R+OSGeo).
GlobalSoilMap.net presentation, 11 Feb 2011
5. My dream is to build an Open multipurpose GLIS
Soil properties (soil information system)
- physical and chemical soil properties, nutrient
capacity, water storage, acidity/salinity…
Model library Live weather channel (meteorological forecasting)
- anticipated temperature (min, max), rainfall, frost
hazard, drought hazard, flood hazard…
Fertilization
Irrigation Plant monitoring channel (MODIS/ENVISAT)
Pest treatment - current biomass production, biomass anomalies
Best crop calendar (pest and diseases), plant health…
Yield estimates
Environmental risks Socio-economic data (site-specific)
GLOBAL - administrative units, new laws and regulations,
LAND INFORMATION market activity, closest offices, agro-dealers…
SYSTEM
Suggest the best
land use practice Query site
attributes
Information Update with
incorrect? ground truth data
Spatial location (site)
GlobalSoilMap.net presentation, 11 Feb 2011
6. GlobalSoilMap.net
An international initiative to make soil property maps (7+3) at
six depths at 3 arcsecs (100 m).
the lightmotive is to assemble, collate, and rescue as much of
the worlds existing soil data ;
Some 30 people directly involved (ISRIC is the main project
coordinator).
International compilation of soil data.
The soil-equivalent of the OneGeology.org, GBIF, GlobCover
and similar projects.
See full specications at
http://globalsoilmap.org/specifications
GlobalSoilMap.net presentation, 11 Feb 2011
7. World soils in numbers
The total productive soil areas: about 104 million square
km.
GlobalSoilMap.net presentation, 11 Feb 2011
8. World soils in numbers
The total productive soil areas: about 104 million square
km.
k
To map the world at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
GlobalSoilMap.net presentation, 11 Feb 2011
9. World soils in numbers
The total productive soil areas: about 104 million square
km.
k
To map the world at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
We would require some 65M proles according to the strict
rules of Avery (1987).
GlobalSoilMap.net presentation, 11 Feb 2011
10. World soils in numbers
The total productive soil areas: about 104 million square
km.
k
To map the world at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
We would require some 65M proles according to the strict
rules of Avery (1987).
World map at 0.008333333 arcdegrees (ca.1 km) resolution is
an image of size 43,200 Ö21,600 pixels.
GlobalSoilMap.net presentation, 11 Feb 2011
11. World soils in numbers
The total productive soil areas: about 104 million square
km.
k
To map the world at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
We would require some 65M proles according to the strict
rules of Avery (1987).
World map at 0.008333333 arcdegrees (ca.1 km) resolution is
an image of size 43,200 Ö21,600 pixels.
27 billion pixels needed to represent the whole world in 100 m
(productive soil areas).
GlobalSoilMap.net presentation, 11 Feb 2011
12. GSM in comparison with other similar projects
4.0
GLWD
EcoRegions HWSDv1
5.6 km MOD12C1
MOD13C2 CHLO/SST
3.5
FRA
Resolution (m) in log-scale
WorldClim
GPWv3
3.0
DMSP-OLSv4
GlobCov2 OneGeology?
2.5
SRTM GADM GlobalSoilMap?
2.0
1990 1995 2000 2005 2010 2015 2020
Year
GlobalSoilMap.net presentation, 11 Feb 2011
13. Misconceptions #1
Mapping eciency can be expressed as cost in $ per
area.
To map world soils at 100 m using per unit costs of
$2/km2 would cost ca.$300 million1 .
1
Pedro Sanchez; the NY GlobalSoiMap.net meeting (17th Feb 2009).
GlobalSoilMap.net presentation, 11 Feb 2011
14. Survey costs and mapping scale
q
Minimum survey costs in EUR / ha (log−scale)
3
q
2
q
1
q
0
−1
q
9.5 10.0 10.5 11.0 11.5 12.0 12.5
Scale number (log−scale)
GlobalSoilMap.net presentation, 11 Feb 2011
15. Mapping accuracy and survey costs
The cost of a soil survey is a function of mapping scale, roughly:
log(X) = b0 + b1 · log(SN) (1)
We can t a linear model to the empirical table data from
e.g.Legros (2006; p.75), and hence we get:
X = exp (19.0825 − 1.6232 · log(SN)) (2)
where X is the minimum cost/ha in Euros (based on estimates in
2002). To map 1 ha of soil at 1:100,000 scale, for example, one
needs (at least) 1.5 Euros.
GlobalSoilMap.net presentation, 11 Feb 2011
16. The GSM calculus
The total productive soil areas: about 104 million square
km.
GlobalSoilMap.net presentation, 11 Feb 2011
17. The GSM calculus
The total productive soil areas: about 104 million square
km.
k
To map the world soils at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
According to Pedro Sanchez, soils could be mapped for
$0.20 USD per ha ( $300 million USD).
GlobalSoilMap.net presentation, 11 Feb 2011
18. The GSM calculus
The total productive soil areas: about 104 million square
km.
k
To map the world soils at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
According to Pedro Sanchez, soils could be mapped for
$0.20 USD per ha ( $300 million USD).
We would require some 65M proles according to the strict
rules of Avery (1987).
GlobalSoilMap.net presentation, 11 Feb 2011
19. The GSM calculus
The total productive soil areas: about 104 million square
km.
k
To map the world soils at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
According to Pedro Sanchez, soils could be mapped for
$0.20 USD per ha ( $300 million USD).
We would require some 65M proles according to the strict
rules of Avery (1987).
World map at 0.008333333 arcdegrees (ca.1 km) resolution is
an image of size 43,200 Ö21,600 pixels.
GlobalSoilMap.net presentation, 11 Feb 2011
20. The GSM calculus
The total productive soil areas: about 104 million square
km.
k
To map the world soils at 100 m (1:200 ), would cost about
5 billion EUR (0.5 EUR per ha) using traditional methods.
According to Pedro Sanchez, soils could be mapped for
$0.20 USD per ha ( $300 million USD).
We would require some 65M proles according to the strict
rules of Avery (1987).
World map at 0.008333333 arcdegrees (ca.1 km) resolution is
an image of size 43,200 Ö21,600 pixels.
We would need immense storage capacities one image of
the world at a 100 m resolution contains 27 billion pixels
(productive soil areas only!).
GlobalSoilMap.net presentation, 11 Feb 2011
21. Mapping eciency
The costs-per-area measure is not really informative (it is easy to
spend money).
We propose instead a measure called mapping eciency, dened
as the amount of money needed to map an area of standard size
and explain each one percent of variation in the target variable:
X
θ= [EUR · km−2 · %−1 ] (3)
A · RMSE r
where X is the total costs of a survey, A is the size of area in
km
−2 , and RMSE r is the amount of variation explained by the
spatial prediction model.
GlobalSoilMap.net presentation, 11 Feb 2011
23. Information production eciency
information
An additional measure of mapping eciency is the
production eciency, i.e.the amount of money spent to produce
a given quantity of soil information:
X
Υ= [EUR · B−1 ] (4)
gzip
where gzip is the amount of data (in Bytes) left after compression:
gzip = fc · (fE · M ) · cZ [B] (5)
where fc is the loss-less data compression factor, fE is the
extrapolation adjustment factor, cZ is the variable coding size, and
M is the total number of pixels.
GlobalSoilMap.net presentation, 11 Feb 2011
24. Map information content
Variable coding can be set by deriving the (global) eective
precision of a soil property map:
RMSE
∆z = ; Z = {Z(s), ∀s ∈ A} (6)
2
Following the Nyquist frequency concept from signal processing,
there is no justication in saving the predictions with better
precision than half the average accuracy.
GlobalSoilMap.net presentation, 11 Feb 2011
25. Map information content
Eective information content (bytes remaining after compression)
in a soil map for a given map extent is basically a function of three
factors:
Support size (point or block).
Size of a map in terms of number of pixels, determined, in
fact, by the eective pixel size (which is in fact determined
by sampling intensity).
Eective precision (Eq.6) estimated using validation points.
GlobalSoilMap.net presentation, 11 Feb 2011
26. Conclusions
Mapping eciency (cost / area / percent of variance
explained) is an objective criteria to compare spatial prediction
methods. $ / area is incomplete (anyone can spend money to
produce maps the question is how good are the maps?).
GlobalSoilMap.net presentation, 11 Feb 2011
27. Conclusions
Mapping eciency (cost / area / percent of variance
explained) is an objective criteria to compare spatial prediction
methods. $ / area is incomplete (anyone can spend money to
produce maps the question is how good are the maps?).
Maps are not what they seem always assess and visualize
the accuracy of your maps.
GlobalSoilMap.net presentation, 11 Feb 2011
28. Conclusions
Mapping eciency (cost / area / percent of variance
explained) is an objective criteria to compare spatial prediction
methods. $ / area is incomplete (anyone can spend money to
produce maps the question is how good are the maps?).
Maps are not what they seem always assess and visualize
the accuracy of your maps.
Soil mapping is an iterative process, in each iteration we
explain a bit more of variability.
GlobalSoilMap.net presentation, 11 Feb 2011
29. Conclusions
Mapping eciency (cost / area / percent of variance
explained) is an objective criteria to compare spatial prediction
methods. $ / area is incomplete (anyone can spend money to
produce maps the question is how good are the maps?).
Maps are not what they seem always assess and visualize
the accuracy of your maps.
Soil mapping is an iterative process, in each iteration we
explain a bit more of variability.
We might not ever be able to explain 100% variability in the
target soil variable.
GlobalSoilMap.net presentation, 11 Feb 2011
30. Misconceptions #2
Each node will produce soil property maps for their
area of interest, which can then be stitched together2
These maps will become the most used soil
information in the World.
2
This is not species on GlobalSoilMap.net, but there is a general agreement.
GlobalSoilMap.net presentation, 11 Feb 2011
31. A hierarchical approach to GSM
Country nodes continental nodes (major players) Global
coverage.
Each country node is responsible for producing maps for their
territory. The nodes havea complete freedom to select
applicable spatial prediction methods (delivery tempo,
data sharing policy etc.).
As long as the technical specications are satised (10
properties, 6 depths, upper lower condence limits, 100 m),
the maps will be put on GlobalSoilMap.net.
Inputs and methods to be used for GSM are secondary.
GlobalSoilMap.net presentation, 11 Feb 2011
32. Lessons from geodata usability
Geodata usability is a function of: (1) adequacy, (2)
consistency, (3) completeness, (4) accuracy of the
metadata, (5) data interoperability, (6) accessibility and
data sharing capacity, (7) attribute and thematic
accuracy.
GlobalSoilMap.net presentation, 11 Feb 2011
33. Lessons from geodata usability
Geodata usability is a function of: (1) adequacy, (2)
consistency, (3) completeness, (4) accuracy of the
metadata, (5) data interoperability, (6) accessibility and
data sharing capacity, (7) attribute and thematic
accuracy.
Each of these aspects can be optimized.
GlobalSoilMap.net presentation, 11 Feb 2011
34. Lessons from geodata usability
Geodata usability is a function of: (1) adequacy, (2)
consistency, (3) completeness, (4) accuracy of the
metadata, (5) data interoperability, (6) accessibility and
data sharing capacity, (7) attribute and thematic
accuracy.
Each of these aspects can be optimized.
In reality, we can only increase each of the listed factors up to
a certain level, then due to objective reasons, we reach the
best possible performance given the available funds and
methods. Any other improvement would require additional
funds (or radical improvement of the data/operation models).
GlobalSoilMap.net presentation, 11 Feb 2011
35. Soil proles from various projects (65k points)
GlobalSoilMap.net presentation, 11 Feb 2011
36. Conclusions
A hierarchical (isolation) approach to global soil mapping
(stitching of country maps) would probably lead to products
that are inconsistent, incomplete and irreproducible.
GlobalSoilMap.net presentation, 11 Feb 2011
37. Conclusions
A hierarchical (isolation) approach to global soil mapping
(stitching of country maps) would probably lead to products
that are inconsistent, incomplete and irreproducible.
Considering the current state of legacy data, any GSM will
need to be largely based on extrapolation and downscaling.
GlobalSoilMap.net presentation, 11 Feb 2011
38. Conclusions
A hierarchical (isolation) approach to global soil mapping
(stitching of country maps) would probably lead to products
that are inconsistent, incomplete and irreproducible.
Considering the current state of legacy data, any GSM will
need to be largely based on extrapolation and downscaling.
The Global Soil Mapping initiative should be about building
live repositories (Open Soil Proles, Soil Covariates) and tools
(Global Soil Information Facility).
GlobalSoilMap.net presentation, 11 Feb 2011
39. Conclusions
A hierarchical (isolation) approach to global soil mapping
(stitching of country maps) would probably lead to products
that are inconsistent, incomplete and irreproducible.
Considering the current state of legacy data, any GSM will
need to be largely based on extrapolation and downscaling.
The Global Soil Mapping initiative should be about building
live repositories (Open Soil Proles, Soil Covariates) and tools
(Global Soil Information Facility).
k $300
To map the world soils at 100 m (1:200 ), would cost ca.
million USD. To update such map would cost (again!) $300
million USD.
GlobalSoilMap.net presentation, 11 Feb 2011
40. Conclusions
A hierarchical (isolation) approach to global soil mapping
(stitching of country maps) would probably lead to products
that are inconsistent, incomplete and irreproducible.
Considering the current state of legacy data, any GSM will
need to be largely based on extrapolation and downscaling.
The Global Soil Mapping initiative should be about building
live repositories (Open Soil Proles, Soil Covariates) and tools
(Global Soil Information Facility).
k
To map the world soils at 100 m (1:200 ), would cost ca. $300
million USD. To update such map would cost (again!) $300
million USD.
The future of digital soil mapping lays in task-oriented Soil
Information Systems (idea by Gerard Heuvelink).
GlobalSoilMap.net presentation, 11 Feb 2011
41. Misconceptions #3
There are many possible DSM techniques that are
equally suitable for GSM.
Each node should use which ever technique they nd
applicable.
GlobalSoilMap.net presentation, 11 Feb 2011
42. GSM techniques
Data rich areas Data poor areas
Know extrapolation
Profile data and polygon maps ledge
trans
fer
Profile data only
Polygon maps only
No soil data available
Purely Knowledge-
Hybrid Extrapolation
geostatistical driven
methods methods
methods methods
Figure: Groups of techniques suitable for global soil mapping; after
Minasny and McBratney (2010).
GlobalSoilMap.net presentation, 11 Feb 2011
43. Conclusions
Most of the DSM techniques are in fact somehow connected
(weighted averaging per polygon is in fact type of regression,
SOLIM is type of multiple linear regression), hence, there are
not as many techniques.
GlobalSoilMap.net presentation, 11 Feb 2011
44. Conclusions
Most of the DSM techniques are in fact somehow connected
(weighted averaging per polygon is in fact type of regression,
SOLIM is type of multiple linear regression), hence, there are
not as many techniques.
For the consistency and completeness of nal outputs it is
probably better to build one global model for each soil
property (or even one multivariate model).
GlobalSoilMap.net presentation, 11 Feb 2011
45. Conclusions
Most of the DSM techniques are in fact somehow connected
(weighted averaging per polygon is in fact type of regression,
SOLIM is type of multiple linear regression), hence, there are
not as many techniques.
For the consistency and completeness of nal outputs it is
probably better to build one global model for each soil
property (or even one multivariate model).
Selection of covariates and prediction techniques needs
to be clearly driven by objective accuracy assessment.
GlobalSoilMap.net presentation, 11 Feb 2011
46. Other global mapping projects
SRTM (DEM) 100 m near-to-global coverage.
MODIS products a variety of RS-based products
(vegetation indices, LAI, land cover maps etc) at resolutions
250 m, 500 m, 1 km and 5.6 km.
GlobCov ESA's ENVISAT global consistent land cover
map (300 m).
WorldClim maps of bioclimatic variables interpolated using
dense point data (1 km).
... there are many more examples (see also: publicly available
data sets).
All these are based on using unied methodology.
GlobalSoilMap.net presentation, 11 Feb 2011
47. Diculties
There is probably not enough point data in the world to make
soil property maps at so ne resolution (maps will be largely
based on extrapolation and downscaling).
GlobalSoilMap.net presentation, 11 Feb 2011
48. Diculties
There is probably not enough point data in the world to make
soil property maps at so ne resolution (maps will be largely
based on extrapolation and downscaling).
The most serious problem of GSM is the discrepancy between
the countries considering the amount of (eld) data.
GlobalSoilMap.net presentation, 11 Feb 2011
49. Diculties
There is probably not enough point data in the world to make
soil property maps at so ne resolution (maps will be largely
based on extrapolation and downscaling).
The most serious problem of GSM is the discrepancy between
the countries considering the amount of (eld) data.
Soils are NOT vegetation it is much more dicult to
map distribution of soils accurately (RS is helpful, but only up
to a certain degree).
GlobalSoilMap.net presentation, 11 Feb 2011
50. Diculties
There is probably not enough point data in the world to make
soil property maps at so ne resolution (maps will be largely
based on extrapolation and downscaling).
The most serious problem of GSM is the discrepancy between
the countries considering the amount of (eld) data.
Soils are NOT vegetation it is much more dicult to
map distribution of soils accurately (RS is helpful, but only up
to a certain degree).
The nal global soil property maps might be of poor accuracy
in 50% of the world.
GlobalSoilMap.net presentation, 11 Feb 2011
51. Question:
Can we do GSM @ 100 m with such limited data?
GlobalSoilMap.net presentation, 11 Feb 2011
52. Opportunities
getting the legacy data
There is an enormous potential of
together (there must be thousands and thousands of soil
proles unused).
GlobalSoilMap.net presentation, 11 Feb 2011
53. Opportunities
getting the legacy data
There is an enormous potential of
together (there must be thousands and thousands of soil
proles unused).
There is an impressive enthusiasm about this project (many
national soil survey agencies see this as an opportunity to get
funding).
GlobalSoilMap.net presentation, 11 Feb 2011
54. Opportunities
getting the legacy data
There is an enormous potential of
together (there must be thousands and thousands of soil
proles unused).
There is an impressive enthusiasm about this project (many
national soil survey agencies see this as an opportunity to get
funding).
World (scientists, policy makers, crediting organizations,
private sector, ... farmers) need soil information!
GlobalSoilMap.net presentation, 11 Feb 2011
55. The proposal
We propose that, for the purpose of achieving the
highest geodata usability, the project should promote
use of a single (participatory) global multiscale
nested regression-kriging model (5 km, 1 km, 250 m
and 100 m resolution)
and then engage local DSM teams to contribute soil
ground truth data, polygon maps and predictions
that can be integrated into one information system.
GlobalSoilMap.net presentation, 11 Feb 2011
56. Global Multiscale Nested RK
Predictions are based on a nested RK model:
z(sB ) = m0 (sB−k ) + e1 (sB−k |sB−[k+1] ) + . . . + ek (sB−2 |sB−1 ) + ε(sB ) (7)
where z(sB ) is the value of the target variable estimated at ground
scale (B), B−1 , . . . ,B−k are the higher order components,
ek (sB−k |sB−(k+1) ) is the residual variation from scale sB−(k+1) to a
higher resolution scale sB−k , and ε is spatially auto-correlated
residual soil variation (dealt with ordinary kriging).
GlobalSoilMap.net presentation, 11 Feb 2011
57. Some drawbacks
GM-NRK makes all other DSM eorts in the World
redundant(!);
GM-NRK ignores all other sub-100 m resolution data and
mapping eorts;
It could also delay delivery of soil property maps because the
mapping activities would be more dicult to organize
internationally;
GlobalSoilMap.net presentation, 11 Feb 2011
58. The best combined spatial predictor
participatory
To avoid these diculties, we propose using a
approach to GSM a combination of GM-NRK and local
prediction models. Assuming that at local and global scales
independent inputs/models are used to generate predictions, the
best combined predictor can be obtained by using:
1 1
zGM−NRK (s0 ) ·
ˆ RMSE r (GM−NRK) + zLM (s0 )
ˆ · RMSE r (LM)
zBCSP (s0 ) =
ˆ (8)
2
1
RMSE r (Mj)
j=1
where RMSE r is the prediction error estimated using
cross-validation (Eq.3).
GlobalSoilMap.net presentation, 11 Feb 2011
59. The proposed system
Multiscale prediction
Spatial aggregation model
5.6 km
ISRIC
1 km
downscaling
GlobalSoilMap.net
250 m continental nodes
automated validation Regional mapping
100 m
new submission organization
1x1
degree tiles FTP service (clearing house)
(7 properties,
6 depths) PostGIS Raster DB
GeoTiff (3 arcsec)
soil property maps
Data portal
WMS KML GeoTIff
(visualization: web browser) (visualization: Google Earth) (analysis: GIS)
GlobalSoilMap.net presentation, 11 Feb 2011
60. GM-NRK in action: Malawi showcase
2740 soil observations, from which some 8001000 contain
complete analytical and descriptive data.
GlobalSoilMap.net presentation, 11 Feb 2011
61. GM-NRK in action: Malawi showcase
2740 soil observations, from which some 8001000 contain
complete analytical and descriptive data.
1:800k polygon soil map.
GlobalSoilMap.net presentation, 11 Feb 2011
62. GM-NRK in action: Malawi showcase
2740 soil observations, from which some 8001000 contain
complete analytical and descriptive data.
1:800k polygon soil map.
Some 30-40 gridded layers at various resolutions
(covariates).
GlobalSoilMap.net presentation, 11 Feb 2011
63. Data sets available for Malawi
(a) (b) (c)
48.8
32.7
16.6
0.5 10°
11°
12°
13°
14°
15°
16°
38000
32667
27333
22000
17°
33° 34° 35°
GlobalSoilMap.net presentation, 11 Feb 2011
64. Gridded maps for Malawi
Parent General land Erosion Land
Climate Biomes
material use deposition management
Rainfall map of the world
5.6 km
MODIS-based long term Land Surface
Temperature (day/night)
Elevation
1 km Geologic Provinces of Africa
Soil polygon map (FAO classes)
ENVISAT Land Cover map (GlobCov)
MODIS (MCD12Q1) land cover dynamics
250 m
MODIS (MCD13Q1) Enhanced Vegetation
Index (EVI) and medium infrared band (MIR)
TWI, TRI, Slope,
Surface roughness,
100 m Insolation
Landsat ETM
thermal band
GlobalSoilMap.net presentation, 11 Feb 2011
65. Loading the data
# library(GSIF)
# This library is still not available, hence just load the functions:
source(http://globalsoilmap.org/data/functions.R)
# load the input data:
source(http://globalsoilmap.org/data/malawi.RData)
ls()
# mw_soil.utm --- soil polygon map at 1:800k scale;
# malawi.utm --- ca 2000 soil profiles for the whole Malawi;
# malawi.poly.utm --- country borders (lines);
This will load all point, polygon data and and R functions required
to run this exercise. The input gridded data can be obtained from:
download.file(http://globalsoilmap.org/data/malawi_grids.zip,
+ destfile=paste(getwd(), malawi_grids.zip, sep=/))
# 313 MB
GlobalSoilMap.net presentation, 11 Feb 2011
66. geology for CLYPPT. At 250 m resolution, the models are again more significant: the predictors explain 18.7%
RegressionsoilanalysisPHIHO5,elevation, EVI maps and soil types for PHIHO5, andMODISelevation,
of variability for ORCDRC, 21.1% for
Infrared band and type map for ORCDRC,
and 26.8% for CLYPPT. The best predictors are:
again
Medium
EVI and soil maps for CLYPPT. At finest resolution, we use a smallest subset of predictors (DEM derivatives and
Landsat thermal infrared band). Consequently, the R-squares are somewhat lower: 5.5% for ORCDRC; 12.1% for
PHIHO5 and 9.3% for CLYPPT. The overall best predictors are elevations, landsat TIR and Topographic Wetness
Index (Table 12.2).
Table 12.2 Summary results of regression analysis for three selected soil variables at various scales (case study Malawi).
Best predictors Best predictors Best predictors Best predictors
Variable name OSP code N and R-square and R-square and R-square and R-square
(5 km) (1 km) (250 m) (100 m)
rainfall,
MODIS MIR, soil elevation, landsat
Soil organic temperature of elevation
ORCDRC 785 types TIR, TRI
carbon warmest month (R2 =0.213)
(R2 =0.187) (R2 =0.055)
(R2 =0.315)
precipitation, LAI, MODIS EVI, soil elevation, TWI,
TWI
pH PHIH5O 793 daily LST types TRI
(R2 =0.213)
(R2 =0.464) (R2 =0.211) (R2 =0.121)
soil mapping units, elevation, MODIS elevation, TWI,
geological units
Clay content CLYPPT 756 daily LST EVI devmean
(R2 =0.127)
(R2 =0.148) (R2 =0.268) (R2 =0.093)
It is clear from the results shown in Fig. 12.5 that at each scale different predictors play different role. These
results also confirm that some soil properties, such as clay content, can be better explained using fine-scale
predictors (SRTM DEM derivatives), others such as organic carbon are controlled by global (coarse) predictors
GlobalSoilMap.net presentation, 11 Feb 2011
67. Organic carbon (values in log-scale)
5 km 1 km 250 m
3.200
2.533
1.867
1.200
0 100 km
GlobalSoilMap.net presentation, 11 Feb 2011
68. pH visualized in GE (1 degree block)
GlobalSoilMap.net presentation, 11 Feb 2011
69. Conclusions
GSM at 100 m is doable (even without 6M proles!).
GlobalSoilMap.net presentation, 11 Feb 2011
70. Conclusions
GSM at 100 m is doable (even without 6M proles!).
The multiscale approach allows us to extrapolate in
large area (even to areas where we have no soil data!).
GlobalSoilMap.net presentation, 11 Feb 2011
71. Conclusions
GSM at 100 m is doable (even without 6M proles!).
The multiscale approach allows us to extrapolate in
large area (even to areas where we have no soil data!).
Selection of covariates and prediction techniques needs
to be clearly driven by objective accuracy assessment.
GlobalSoilMap.net presentation, 11 Feb 2011
72. Conclusions
GSM at 100 m is doable (even without 6M proles!).
The multiscale approach allows us to extrapolate in
large area (even to areas where we have no soil data!).
Selection of covariates and prediction techniques needs
to be clearly driven by objective accuracy assessment.
The point data is the key to GSM we need to motivate
governmental agencies and private persons to contribute to
OSP.
GlobalSoilMap.net presentation, 11 Feb 2011
73. Conclusions
GSM at 100 m is doable (even without 6M proles!).
The multiscale approach allows us to extrapolate in
large area (even to areas where we have no soil data!).
Selection of covariates and prediction techniques needs
to be clearly driven by objective accuracy assessment.
The point data is the key to GSM we need to motivate
governmental agencies and private persons to contribute to
OSP.
We need to start developing and testing tools if you
have the inputs and the tools to generate outputs, they can be
re-generated as many times as you wish.
GlobalSoilMap.net presentation, 11 Feb 2011
74. GSM products (revisited)
SoilGrids.org covariates at 5 km, 1 km (250 m).
SoilProles.org Open Soil Proles (once we reach 1M
points we should be able to produce soil property maps with
reasonable accuracy).
R/Python package automated analysis of point and
gridded data.
GSIF Global Information Facilities for soil data.
GlobalSoilMap.net presentation, 11 Feb 2011
75. Next steps
Re-implement the method using a `clean' data set (USA
data).
GlobalSoilMap.net presentation, 11 Feb 2011
76. Next steps
Re-implement the method using a `clean' data set (USA
data).
Finalize the blue-paper (technical specs and methods for
GSM).
GlobalSoilMap.net presentation, 11 Feb 2011
77. Next steps
Re-implement the method using a `clean' data set (USA
data).
Finalize the blue-paper (technical specs and methods for
GSM).
Package a showcase that anyone can use.
GlobalSoilMap.net presentation, 11 Feb 2011
78. Next steps
Re-implement the method using a `clean' data set (USA
data).
Finalize the blue-paper (technical specs and methods for
GSM).
Package a showcase that anyone can use.
Set-up web-services (ISRIC servers) and start publishing
the data (launch OSP, worldmaps).
GlobalSoilMap.net presentation, 11 Feb 2011