1. Bayesian network models in
MARS:
Case study Lake Vansjø
Task 7.3: Combining abiotic and biotic models for
river basin management planning
Jannicke Moe, Raoul Couture, Anne Lyche Solheim
(NIVA)
MARS WP7 meeting
18.10.2016, Den Helder (Netherlands)
18.10.2016J Moe, RM Couture, AL Solheim 1
2. Progress since the Oslo meeting
18.10.2016J Moe, RM Couture, AL Solheim 2
More details: http://www.slideshare.net/JannickeMoe/mars-wp7-bnvansjojmo20151113
3. Lake Vansjø –
basic info
• Vansjø-basin Vanemfjorden
• Catchment dominated by forest
and agriculture
• Long history of eutrophication
• Extreme rain events
• Moderate ecological status due
to eutrophication
• Phytoplankton (dominated by
Cyanobacteria), macrophytes, total P
18.10.2016J Moe, RM Couture, AL Solheim 3
Haande, Lyche Solheim,
Moe & Brænden 2011.
NIVA report
6. Mapping the BN for Vansjø to
the MARS conceptual model (DPSIR)
18.10.2016J Moe, RM Couture, AL Solheim 6
DRIVER
DRIVER
PRESSURE
(nutrient
loads etc.)
STATE: ABIOTIC
INDICATORS
STATE: ABIOTIC
INDICATORS
STATE: BIOTIC
INDICATORS
STATE
:
BIOTIC
IND.
RESPONSE
STATE:
WFD STATUS
• What about IMPACT - functions and services?
7. Other BNs for Vansjø include IMPACT
18.10.2016J Moe, RM Couture, AL Solheim 7
Barton et al. 2016. Eutropia – integrated valuation of
lake eutrophication abatement decisions using a
Bayesian belief network. In: Z.Neal (ed.). Handbook of
Applied Systems Science. Routledge.
• IMPACT nodes can be linked to STATE nodes
• Suitability for fishing
• Suitability for bathing
IMPACT
IMPACT
STATES
8. A BN for multiple stressors in lake Vansjø
Moe, Haande & Couture. Ecological Modelling (2016)
18.10.2016J Moe, RM Couture, AL Solheim 8
• Aim: predict effects of scenarios on ecological status
• 4 modules: different sources of information
9. Module 1: Scenarios (from REFRESH)
•Climate scenarios:
• Reference
• «Hadley»: higher temperature, more precipitation
•Management scenarios:
• Reference
• Best: less TP (~Consensus world)
• Worst: more TP (~Techno or Fragmented world)
•Will re-do using MARS scenarios for climate and
land-use
18.10.2016J Moe, RM Couture, AL Solheim 9
10. Module 2: Output from process-based
models
18.10.2016J Moe, RM Couture, AL Solheim 10
• Process-based models:
• Persist (hydrology)
• INCA-P (catchment)
• MyLake (lake) input to BN
• 60 realisations of the model
(parameter combinations)
give rise to probability
distributions in the BN model
11. 18.10.2016J Moe, RM Couture, AL Solheim 11
Module 3: Monitoring data -
cyanobacteria
• Multiple regressions:
Identify significant predictor
variables
structure of nodes and
arrows in BN model
• Regression tree analysis:
Identify breakpoints in
predictor variables
discretisation (setting
intervals) of nodes in BN
Empirical relationships between abiotic and biotic variables
quantified by data analysis (WP4)
12. What are inside the arrows?
- conditional probability tables (CPT)
18.10.2016J Moe, RM Couture, AL Solheim 12
CPT for Cyano
•Based on 90 observations
CPT for Status Phytoplankton
•Based on knowledge (combination rules)
States
Probabilities
13. 18.10.2016J Moe, RM Couture, AL Solheim 13
Module 4: Ecological status
• Status for different quality elements are combined in
CPTs according to the national classification system
• E.g. status of phytoplankton:
• If status of cyanobacteria < chl-a,
the combined status is averaged
• If status of cyanobacteria > chl-a,
cyanobacteria are not considered
14. Results of model for Scenario: reference
18.10.2016J Moe, RM Couture, AL Solheim 14
Probability of Poor-Bad status equal for
Cyanobacteria and Chl-a (~45%)
15. Results of model for Scenario:
best management, future climate
18.10.2016J Moe, RM Couture, AL Solheim 15
Probability of Poor-Bad status higher for
Cyano (40%) than for Chl-a (36%)
16. Results - all scenarios
18.10.2016J Moe, RM Couture, AL Solheim 16
0
20
40
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Probab
0
20
40
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
0
20
40
Re
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Chla
Probability(%)
(d)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Cyanobacteria
(e)
0
20
40
60
80
100
Re
(f)
Poor-Bad
Moderate
High-Good
0
20
40
60
80
100
Re
Probability(%)
(g)
0
20
Ref Had Ref
Climate sc
Worst Ref
Management
Pr
Poor-Bad
Moderate
High-Good
-• Chl-a: Climate change impact is negative, but small
compared to land use impact
• Cyanobacteria: responses to scenarios are similar to chl-a,
but...
• Including cyanobacteria reduces the probability of good
ecological status for phytoplankton
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Secchi depth
Probability(%)
(a)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Total P
(b)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Phys.-chem.
(c)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Chla
Probability(%)
(d)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Cyanobacteria
(e)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Phytoplankton
(f)
Poor-Bad
Moderate
High-Good
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Lake
Probability(%)
(g)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Secchi depth
Probability(%)
(a)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Total P
(b)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Phys.-chem.
(c)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Chla
Probability(%)
(d)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Cyanobacteria
(e)
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Phytoplankton
(f)
Poor-Bad
Moderate
High-Good
0
20
40
60
80
100
Ref Had Ref Had Ref Had
Climate scenario
Worst Ref Best
Management scenario
Lake
Probability(%)
(g)
17. Problems encountered
1) How to link predicted and observed values
(Total P (pred.) is now the "cause" of Total P (obs.))
- No better solution found
2) How to handle the poor match between predicted and
observed values (especially Total P)
- Improvement needed in the process-based model
18. Problems encountered
3) How to deal with missing or few values for CPTs
(columns with all zeros)
- Will try combination with expert judgement
4) How to make better use of additional information
(data on cyanobacteria from 400 other Norwegian
lakes)
- Will try built-in method for updating CPT with new data
Vansjø + 400 lakes
19. How our BN can be of use for water
management in locally and elsewhere
• as a bridge between the coarse MARS conceptual
model and the detailed process-based models
• aggregating input and output of process-based models
• linking abiotic and biotic components
• including biotic components where data are sparse but
knowledge is available
• for quickly re-running scenarios
• a kind of model emulator
• forwards and backwards
• for incorporating and visualising uncertainty
• for communication with stakeholders: model
structure, scenarios, results and uncertainties
18.10.2016J Moe, RM Couture, AL Solheim 19
20. Next steps for Lake Vansjø BN
February - April 2017
• Apply MARS future scenarios - aggregate the
outcome of WP4
• Improve the CPT for cyanobacteria
• Expert judgement; update with large-scale dataset
• Add colour (organic C) as abiotic state variable,
with potential negative impact on cyanobacteria
• from empirical analysis in WP4
• Try PTI (Phytoplankton Trophic Index) as
additional biotic state variable
18.10.2016J Moe, RM Couture, AL Solheim 20