P1252109393IaWCJ - PowerPoint PPT Presentation

Title: P1252109393IaWCJ


1
Using Environmental Process Models to Guide
Sensor Network Sampling Design

• Environmental processes
• Distributed parameter models
• Sensor network design what level of detail do
  you need?
• Example river hydrodynamics
• Other considerations with respect to sensor
  networks

Tom Harmon UC Merced School of Engineering and
the Center for Embedded Networked Sensing (CENS)
2
Modeling Environmental Processes

• Generally interested in fluids (air, water)
• and transported chemicals, organisms (e.g.,
  oxygen bacteria)
• Environmental matrices convey the flow
• Usually complex geometry (e.g., coastal margin
  with inlets)
• Often composed of nonhomogeneous materials (e.g.,
  soil)

3
Environmental process models

• By far the most common approach in environmental
  science
• Compartment (box) models
• Materials and energy in each compartment are
  homogeneously distributed
• Transfers between compartments are simple terms
  (yet can be tricky to estimate)
• Useful to some extent and easily applied (systems
  of ODEs)
• Can be designed, parameterized, tested using
  relatively sparse sensor networks
• Hydrodynamic models
• Fluid flow, mass and energy transport models
• Well-studied but more difficult to use
• Complex geometries, distributed parameters
• High granularity sensor networks will enhance the
  accuracy and usefulness of these modelsand vice
  versa

4
Scales and scaleability

• The scientific questions we ask influence the
  spatial and temporal scales over which we
  observe
• and the length and timescales of our modeling
  approaches
• which (along with the budget!) are used to drive
  our sensor network design

5
Questions and scales of observation
Quick example soil-water-plant system
How are nutrients cycled between the soil and
roots?
How are water and nutrients optimally applied to
a crop?
How sure are we that we are avoiding
over-irrigating/fertilizing (which can pollute
groundwater)?
6
Longer example Consider a river
Ready?...then climb aboard!
7
Whats around the bend?
River sensor network design exercise

• Textbook rivers and network design
• Real river issues
• Example San Joaquin-Merced Rivers confluence

8
Environmental issues and river mixing

• These are mainly centered around pollutants (like
  salinity, nutrients) and oxygen deficits (low
  dissolved oxygen)
• Pollutant mixing is governed by the turbulent and
  shear mixing processes described
• Dissolved oxygen (DO) would also be impacted by
  these processes, but also by
• Biochemical oxygen demand in the river and
  sediments (oxygen consumed)
• Reaeration processes (atmosphere-river surface)
  gas transfer

9
Textbook stuff Water flows downhill
z

• along rough, irregular boundaries
• The turbulence causes mixing of chemical species
• Local velocity differences serve to distribute
  mass carried by the fluid
• This is true for the individual profiles
  (turbulent mixing)
• Also true for the mean profile (shear mixing)

x
y
A
A
d
river bed
horizontal velocity distribution
10
Sampling implications
surface 1.0 bottom
snapshot velocity profile at a given time
z/d
time-averaged velocity profile with depth
river flow direction

• If you need to know average behavior, then dwell
• No problem with stationary sensors
• An issue with mobile sensors in rivers (see Singh
  et al. ICRA 2007)

11
Consider a source flowing into a river
Jason Fisher drawing

• Stationary flow from the pipe and in the river
• Temperature and chemical concentrations are not
  substantial (homogeneous flow)
• River geometry well-defined (banks, bottom
  bathymetry, etc)
• Sensors are available for fluid velocity,
  temperature, salinity

12
Mathematical description of chemical transport
(concentration, C)
Transverse turbulent (or eddy) diffusion
coefficients
longitudinal turbulent (or eddy) diffusion
coefficient BUT, this is overshadowed by shear
dispersion caused by the cross-channel velocity
profile (as opposed to turbulent velocity
fluctuations.so
dispersion coefficient
13
Other extreme Large scale river models

• This model can provide predictions of distributed
  water quality conditions given a well-defined
  source
• D typically overshadows ey in far-field
  observations (major stretches of a river)
• Thus, in far-field applications 1D models
  (advection-dispersion models) are the norm

For recent example (Hudson R.) Ho et al.,
Environ. Sci. Technol. 35(15), pp 3234-3241
(2002)
14
Many science (and regulatory) drivers point to
the need for more detail

• This model may be needed to resolve distributions
  in the mid-field
• 2D may be necessary in many applications, 3D in
  some, 1D for very coarse, large scale issues.

15
Parameter estimates for turbulent diffusion

• Based on shear velocity
• Vertical diffusion (some theoretical basis)
• Remember, this is generally not needed because d
  ltlt W, so vertical mixing has ample time to occur
• Could be necessary to incorporate in near-field
  studies (near the pipe)

16
Traditional estimates of turbulent diffusion
parameters
(ideal channels)

• Correlations developed for lab flumes
• Generally greater diffusion observed in a limited
  number of field studies
• River bank and bottom irregularities, river bends
• Classic experiments
• tracer releases with exhaustive measurements
• days to weeks of boat time, sampling, analysis

17
Lets look at a confluence setting
(salt for example)
18
On-going test site for multi-scale embedded
networked sensing
Californias San Joaquin Valley
19
(No Transcript)
20
San Joaquin-Merced River Confluence
about 100 m
about 7 m
Looking upstream looking
downstream (closeup)
21
Confluence geometry via bathymetry
22
San Joaquin-Merced River Confluence
23
Low Res vs High Res Velocity Cross-Sections
Harmon et al. Environ. Eng. Science, in press,
24(2), 2007 (March issue)
10x vertical exaggeration in plots
24
Salinity distributions
Harmon et al. Environ. Eng. Science, in press,
24(2), 2007 (March issue)
25
Experimental design Measuring mixing at
different scalesmid-field confluence zone
Influent concentrations 1 and 2 (well-mixed
upstream of confluence?)
use historical mixing coefficients to place
transects (e.g., 25 and 50 of complete mixing)
Mixing will follow (roughly) Taylor dispersion
principles - spreading -
variance - moments
26
Measuring mixing at different scales(2) near-
to mid-field observations
Continuous injection of salt or rhodamine dye
solution, oxygen-devoid water
NIMS AQ transect
27
Idealized models will only work for so muchriver
realities
Flows can be highly variable with day, week, and
(below) season
August 31, 2005
April 9, 2006 Merced R-San Joaquin R confluence
28
Water quality perturbations
agro-industrial Discharges (surface and
subsurface)
livestock
29
Opportunities, needs, challenges

• Generic tools observation network design
• Ideal solutions to simple cases ranging to
  numerical
• Fast parameter optimization, model inversion
  strategies
• Model structure identification
• Statistical algorithms connected to network
  design
• Data assimilation strategies
• Models and data from sensors
• Incorporating sensor error, model error, fault
  detection, etc. to yield network integrity
  estimates
• Data fusion strategies
• Remote sensing data ?? embedded sensing data
• Legacy data ?? embedded sensing data

30
Concluding remarks

• Sensor installation and network layout need to
  consider the environmental media and the science
  questions at hand (as well as networking
  technology issues)
• We (environmental scientists) have a good idea
  about how to model many of our systems
• Parameterization should be eased by networked
  sensing
• We need to begin to worry more about model
  structure (before not enough data to bother so
  much)
• Our first-cut models for systems can be used to
  design sensor network modalities and layoutbut
  we must become more nimble in doing this
• Laptop toolkits like design interfaces set in
  GIS-like context for better field transferability