2. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Tectonic setting Deformation
GEOLOGIC
COMPLEXITY
=
Distribution of faults
and lithological
Sedimentation boundaries as a
patterns function of scale
79
Igneous events Au
Hodkiewicz (2003); Ford and Blenkinsop (2008); Ford and McCuaig (2010)
3. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Tectonic setting Deformation
GEOLOGIC
COMPLEXITY
=
Uncertainty
=
Sedimentation Lack of knowledge
patterns
?
Igneous events
Hodkiewicz (2003); Ford and Blenkinsop (2008); Ford and McCuaig (2010)
4. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
147 – 478 m
5. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Small complexity
(volcanic rocks)
Large complexity
(volcanic and
sedimentary)
Small complexity
(sedimentary rocks)
6. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
- 339 sampling points (0 – 20 cm)
- soil and land use survey (purposive sampling)
- particle-size distribution
- transformed to additive log-ratios
- terrain attributes
- 10 m DEM from topographic maps
- multiple linear regression models
- stepwise + repeated 10-fold cross-validation
- R (packages stats, base and caret)
- GRASS, QuantumGIS and SAGA GIS
7. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Table 1. Confusion matrix for the whole study area
Reference
Prediction User accuracy
Sedimentary Volcanic
Sedimentary 132 16 0.8000
Volcanic 33 158 0.9080
Mapper accuracy 0.8919 0.8272
Accuracy 0.8555 (95% CI: 0.8134 – 0.8911)
NIR 0.5133 (P-Value [Acc > NIR]: < 2e-16)
Kappa 0.7099 (Mcnemar's Test P-Value: 0.02227)
8. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Table 2. Confusion matrix for the small complexity area
Reference
Prediction User accuracy
Sedimentary Volcanic
Sedimentary 95 0 1.0000
Volcanic 0 109 1.0000
Mapper accuracy 1.0000 1.0000
Accuracy 1.0000 (95% CI: 0.9821 – 1.0000)
NIR 0.5343 (P-Value [Acc > NIR]: < 2.2e-16)
Kappa 1.0000 (Mcnemar's Test P-Value: NA)
9. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Table 3. Confusion matrix for the large complexity area
Reference
Prediction User accuracy
Sedimentary Volcanic
Sedimentary 37 16 0.5286
Volcanic 33 49 0.7538
Mapper accuracy 0.6981 0.5976
Accuracy 0.6370 (95% CI: 0.5499 – 0.7180)
NIR 0.5185 (P-Value [Acc > NIR]: 0.003605)
Kappa 0.2798 (Mcnemar's Test P-Value: 0.022271)
11. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Small complexity
Large complexity
Small complexity
12. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Geology
- Maciel Filho (1990)
- Sub-horizontal (5º)
- Uncertainty?
Elevation
- Brazilian Army
(1970s)
- Aerial photography
- Uncertainty?
13. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Conclusions
● Yes... we can build predictive models using terrain
attributes (and other environmental co-variates)
● However... large geologic complexity = large
uncertainty = large prediction errors
● Therefore... improving predictive models depends on
improving geologic information available
14. Larger geologic complexity implies larger uncertainty
Alessandro Samuel-Rosa, Ricardo Simão Diniz Dalmolin & Pablo Miguel
Future work
● What is the “required” level of accuracy of the prediction
models?
● Do the users of soil information understand “uncertainty”?
Traditional soil maps “seem” to be very accurate!
● Give an especial attention to the uncertainty of the
environmental co-variates (know where we are going to
“fail”)