This document discusses using Bayesian networks to model dryland salinity in Western Australia. It provides the following key points:
- Bayesian networks can be used to estimate unknown quantities and model probabilistic relationships by building models from available data on topics like medical diagnosis, risk assessment, and physical state estimation.
- An example application is modeling dryland salinity in Western Australia, where the goal is to map current and predict future areas affected by rising saline groundwater levels resulting from past vegetation clearing.
- The challenge is that direct observation of salinity levels across the large area is not possible, so the network would integrate landscape data from remote sensing and other sources to estimate salinity probabilities.
- The document outlines
1. BN, BBN, CPN, ‘Bayesian Networks’
Why use BN
• To estimate quantities which are unobservable
q
• Modelling – building models, ‘elicitation’
• Mixed data, probabilistic relationships
– medical diagnosis
– inference, risk, decision support
– estimation of physical state
p y
– ‘data integration’
6. Sample ‘truth’
Observational
data – spatial
Y/N (date?)
Knowledge about PROCESS
Knowledge ??
Landscape position important
- more likely in valleys
Salinity affects vegetation
- Visible effects ? – images
Groundwater levels S il t
G d t l l Soil type,
Vegetation type, etc ???
7. Network diagram – dryland salinity – FIRST VERSION
Is each location (likely to be) saline or not ? - Not observable directly
Meaning of network, then
- A. How do we observe (get data) on ‘Landform Position’ ? everywhere
- B. How do we observe ‘vegetation condition’ ?
(A from processing DEM; B [surrogate] classification from Landsat)
8. Network – dryland salinity – FIRST VERSION - getting the data
processing classification task
task
DEM IMAGE
‘Raw data’ ?arrows? ‘Raw data’
?what happens?
9. Hydrologists
H d l i t concept di
t diagram - NOT a BN
Ground water
depth and rate Salinity
of rise
?
Hydrological model
- deterministic X Data
Model Parameters
10. Water Poverty ‘Network’
limitations for agriculture ?
volume, critical supply gap,
uncertainty supply
??
Opportunity
cost labour Poverty
measure
WP or surrogate
g
Water-related
health costs
Education/Invest
ment constraints
?
11. Land Monitor – Information Gap
• The three highest priority environmental issues
- Land salinisation,
- Salinisation of inland waters, and
- Maintaining biodiversity
(Western Australian State of the Environment Report, 1998)
• About 1.8 million ha in WA are already salt-affected,
and this area could double in the next 15 to 25 years
years.
• Effects on Vegetation
• No Accurate map, No spatially explicit information
on change, or prediction
12. Salinity Problem & Impact
Resource Problem affects people
Economic & Social Problem
Prediction 25% - 35% land lost
$$ - 40% Australia’s grain
Farming is not subsidised in Aust
g
Business, Land value, Banks $
Built infrastucture : road network
Maintenance;
Town Buildings ‘Rescue Towns’
13. Land Monitor - Components
I.
I Institutional support (agencies).
(agencies)
2. Demonstrated Technical Capacity (CMIS)
Define necessary data (Landsat TM 1988-2000 DEM)
1988-2000,
and methods
3.
3 Funding Support (National Govt)
------------------------
4. Public Interest
LANDSAT TM – Complete Australian Archive since 1988
14. CMIS Methods and technical developments
• Rectification & Registration, Calibration (robust regression)
• Discriminant Analysis (C
(CVA etc)
)
• Enhanced ML classification (PP – uncertainty)
• DEM (pre)Processing – derived variables
• Data Integration - CPN, Decision Trees
• Trend summary and representation (vegetation condition)
(others e.g. NN, LD, D i i T
( th NN LD Decision Trees …)
)