Dynamic control of sensor networks with inferential ecosystem models

1
Dynamic control of sensor networks with
inferential ecosystem models

• Jim Clark, Environm, Biol, Stat
• Pankaj Agarwal, Comp Sci
• David Bell, Environment
• Carla Ellis, Comp Sci
• Paul Flikkema, EE, NAU
• Alan Gelfand, Stat

• Gabriel Katul, Environment
• Kamesh Munagala, Comp Sci
• Gavino Puggioni, Comp Sci
• Adam Silberstein, Comp Sci
• Jun Yang, Comp Sci

2
Motivation

• Understanding forest response to global change
  (climate, CO2)
• Forces at many scales
• Complex interactions
• lagged responses
• Uneven data needs occasionally dense, at
  different scales
• Wireless networks can provide dense data, across
  landscapes

3
Ecosystem models that could use wireless data

• Physiology
• PSN, respiration responses to weather, climate
• C/H2O/energy
• Atmosphere/biosphere exchange (pool sizes,
  fluxes)
• Biodiversity
• Differential demographic responses to
  weather/climate, CO2, H2O

4
Physiological responses to weather
Precip
light, CO2
H2O, CO2
Resp
PSN
Temp
Allocation
Sap flux
Fast, fine scales
H2O, N, P
5
Sensors for ecosystem variables
Demography
Biodiversity
Physiology
Precip Pt
Evap Ej,t
Transpir Trj,t
Light Ij,t
C/H2O/energy
Soil moisture Wj,t
Temp Tj,t
VPD Vj,t
Drainage Dt
6
WisardNet a wireless network

• Multihop, self-organizing
• Sensors for light, soil air T, soil moisture,
  sap flux
• Tower weather station
• Minimal in-network processing

sensor
sensor
gateway
node
Self-organizing wireless connections
7
Mapped stands All life history stages Seed
rain Seed banks Seedlings Saplings Mature
trees Interventions Canopy gaps Nutrient
additions Herbivore exclosures Fire Environmental
monitoring Canopy photos Soil moisture Temperature
Wireless sensor networks Remote sensing
8
Blackwood Division, Duke Forest
sensor
sensor
node
gateway
Fluid topology
9
The goods and the bads

• The good
• Potential to collect dense data
• Adapts to changing communication potential
• The bad
• Most data uninformative, redundant, or both
• Battery life of weeks to months, depending on
  transmission rate
• Checking and replacing batteries is the primary
  maintenance cost of network

10
the ugly

• Unreliable!

Battery life at 13 nodes
Network partially down
Failures
Junk
11
A dynamic control problem

• What is an observation worth?
• (How to quantify learning?)
• The answer recognizes
• Transmission cost of an observation
• Need to assess value in (near) real time
• Based on model(s)
• Minimal in-network computation capacity
• Use (mostly) local information
• Potential for periodic out-of-network input

12
A framework for data collection

• Predict or collect
• Transmit an observation if it could not have been
  predicted by a model.
• Fast decisions (real-time)
• Must rely on (mostly) local information
• Minimize transmission

13
Predictability of ecosystem data
Slow variables
Predictable variables

• Where could a model stand in for data?

Events
Less predictable
14
Which observations are informative?
Light
Shared vs unique data features (within nodes,
among nodes)
Precipitation
Exploit relationships among variables/nodes?
Soil moisture
Slow, predictable relationships?
15
Model-dependent learning

• Exploit relations in space, time, and with other
  variables
• Learning from previous data collection
• If predictable, data have reduced value

PAR at 3 nodes, 3 days
observations
16
Controlling measurement with models

• Inferential modeling concerns
• Some parameters local, some global
• Estimates of global parameters need transmission
• Data cant arrive faster than model converges
• Simple rules for local control of transmission
• Rely mostly on local variables
• Periodic updating from out of network
• Transmit if you cant predict

17
In network data suppression

• An acceptable error, ?
• A standard reactive model based on change,
• Alternative is the observation predictable,

zj local sensor data (no transmission) ?, z,
wt global data, periodically updated from full
model MF full, out-of-network model MI simplifie
d, in-network model
18
Out-of-network model is complex
Calibration data (sparse!)
Sensor data zj,t
Data
y,E,Tr,Dt
y,E,Tr,Dt-1
y,E,Tr,Dt1
Process
Parameters
Location effects
Process parameters
time effect ?t1
time effect ?t-1
Measurement errors
time effect ?t
Process error
heterogeneity
Hyperparameters
19
Soil moisture example

• Simulated process, parameters unknown
• Simulated data
• TDR calibration, error known (sparse),
• 5 sensors, error/drift unknown (dense, but
  unreliable),
• Out-of-network estimate process/parameters
• Use estimates for in-network prediction
• Transmit only when predictions exceed threshold

20
Model summary
Process
Sensor j Rand eff
TDR calibration
Inference
Back to node j
21
Simulated process data
truth y 95 CI 5 sensors z Calibration w
Colors
Red dots
Drift parameters ? Estimates and truth (dashed
lines)
22
Estimates from training phase
Process parameters ? Estimates and truth (dashed
lines)
23
Keepers based on acceptable error
Transmit obs only if
Increasing drift reduces predictive capacity
From plug-in values
Keepers (40), very few until drift accumulates
24
Better data, few observations
Known constraint on missing data
Reanalysis model for missing data
truth y 95 CI 5 sensors z Calibration w
Colors
Red dots
25
Additional variables
26
Collect or predict

• Inferential ecosystem models a currency for
  learning assessment
• In-network simplicity point predictions based on
  local info, periodic out-of-network inputs
• Out-of-network predictive distributions for all
  variables (reanalysis step)
• A role for inference in data collection, not just
  data analysis

27
Advantages over reactive data collection in
wireless networks

• Change in a variable is not directly linked to
  its information content
• Example all soil moisture sensors may change at
  similar rates, making them largely redundant
• Predictability emphasizes change that contains
  information
• The capacity to predict an observation summarizes
  its value and assures that it can be estimated to
  known precision in the reanalysis