Transcript Predictive Modeling

Mining Climate and Ecosystem Data
l
l
l
l
l
l
Team
Motivation
Transformative Computer Science Research
–
Predictive Modeling
–
Complex Networks
–
Association Analysis
–
High Performance Computing
Evaluation Plan
Relationship to Physics Based Models
Management and Collaboration
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Team
Vipin Kumar, UM
Fred Semazzi, NCSU
Auroop Ganguly, UTK/ORNL Nagiza Samatova, NCSU Arindam Banerjee, UM
Joe Knight, UM
Shashi Shekhar, UM
Peter Snyder, UM
Jon Foley, UM
Alok Choudhary, NW Abdollah Homiafar, NCA&T Michael Steinbach, UM Singdhansu Chatterjee, UM
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Team Qualifications
Team members are well-recognized experts in
Analysis of ecosystem and climate data
•
Foley, Ganguly, Kumar, Semazzi, Snyder, Steinbach
Data mining, machine learning, nonlinear dynamics & signal
processing
•
Banerjee, Chatterjee, Choudhary, Ganguly, Homaifar, Kumar,
Samatova, Shekhar, Steinbach
High performance computing
•
Choudhary, Kumar, Samatova
Weather and climate models
•
Ganguly, Foley, Semazzi, Snyder
Remote sensing and land cover change
•
Foley, Knight, Snyder
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
•
The planet is warming
• Multiple lines of evidence
• Credible link to human GHG
(green house gas) emissions
•
Consequences can be dire
• Extreme weather events
• Regional climate and ecosystem
shifts
• Abrupt climate change
• Stress on key resources and
critical infrastructures
•
There is an urgency to act
• Adaptation: “Manage the
unavoidabale”
• Mitigation: “Avoid the
unmanageable”
•
The societal cost of both action
and inaction is large
Mining Climate and Ecosystem Data
Anomalies from 1880-1919 (K)
Climate Change: The defining issue of our era
Figure Courtesy: ORNL
Key outstanding science challenge: Actionable
predictive insights to credibly inform policy
NSF RSV
June 2010
‹#›
Physics-based Models are Essential but Not Adequate
l
l
Models make relatively reliable predictions at global Disagreement between IPCC models
scale for ancillary variables:
–
Sea Surface Temperature (SST)
–
Temperature/humidity profiles over land
–
Wind spread at different heights
They provide least reliable predictions for variables
that are crucial for impact assessment:
–
Regional precipitation and extremes
–
Hurricane intensity and frequency
–
Droughts and floods
“The sad truth of climate science is that the most crucial
information is the least reliable” (Nature, 2010)
Regional hydrology (“P–E” changes in
2030s) exhibits large variations
among major IPCC model projections
Hypothesis-driven “manual” conceptual models try to address this gap:
l
l
Hurricane models (Emanuel et al, BAMS, 2008)
Regional-scale precipitation extremes (O’Gorman & Schneider, PNAS, 2008; Sugiyama et al,
PNAS, 2010)
We need a systematic approach to semi-automatic data-driven model inference.
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Data-Driven Knowledge Discovery in Climate Science
l
l
From data-poor to data-rich transformation
–
Sensor Observations: Remote sensors like satellites and
weather radars as well as in-situ sensors and sensor
networks like weather station and radiation measurements
–
Model Simulations: IPCC climate or earth system models
as well as regional models of climate and hydrology, along
with observed data based model reconstructions
Data-guided insights can complement physics-based models
–
Transform global ancillary information to regional critical climate changes and
extremes
–
Assess reliability of model projections and inform physics model parameterizations
–
Validate predictions with both held-out data and science understanding
–
Provide relatively hypothesis-free discovery processes to supplement hypothesisguided data analysis
"The world of science has changed ... data-intensive science [is] so different that it is worth
distinguishing [it] … as a new, fourth paradigm for scientific exploration." - Jim Gray
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Challenges in Analyzing Eco-Climate Data
Global Sea Surface Temperature
Discovering teleconnections
El Nino Events
Correlation Between ANOM 1+2 and Land Temp (>0.2)
90
0.8
0.6
60
0.4
30
latitude
0.2
Nino 1+2
Index
0
0
-0.2
-30
-0.4
-60
-0.6
-0.8
-90
-180 -150
-120
-90
-60
-30
0
30
60
90
120
150
180
longitude
Challenges due to data characteristics
•
•
•
•
•
•
•
Spatiotemporal, non-stationary, non-i.i.d.
Nonlinear multiscale dependencies
Low frequency variability
Massive data sets
Long range spatial dependencies
Long memory temporal dependencies
…
Relationship between El Nino
and Fires in Indonesia
Changes in Global Forest Cover
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Transformative Computer Science Research
Enabling large-scale data-driven science for complex, multivariate,
spatio-temporal, non-linear, and dynamic systems:
The Expedition is an end-to-end
demonstration of this major
paradigm for future knowledge
discovery process.
Relationship Mining
Enable discovery of complex
dependence structures such as
non linear associations or long
range spatial dependencies
Nonlinear, spatio-temporal, multivariate, persistence, long memory
Complex Networks
Enable studying of collective
behavior of interacting ecoclimate systems
Nonlinear, space-time lag,
geographical, multi-scale
• Fusion plasma
• Combustion
• Astrophysics
• ….
Predictive Modeling
Enable predictive modeling
relationships
Community structure- of typical and extreme
behavior from multivariate
function-dynamics
spatio-temporal data
kernels , features, dependencies
Nonlinear, spatio-temporal,
multivariate
High Performance Computing
Enable efficient large-scale spatio-temporal analytics on
future generation exascale HPC platforms with complex memory hierarchies
Large scale, spatio-temporal, unstructured, dynamic
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Relationship Mining
Objective
•
Computer Science Innovations
•
•
•
•
Dynamic behavior of the
high and low pressure fields
corresponding to NOA
climate index (Portis et al,
2001)
To develop methods for discovery of complex
dependence structures (e.g., nonlinear
associations, long-range spatial dependencies)
Define a notion of “transactions” for
association analysis in spatio-temporal data
Adapt association analysis to continuous
data
Reduce the number of redundant patterns
and assess statistical significance of the
identified patterns
Define a new approach to association
analysis that trades off completeness for a
smaller, simpler set of patterns and far
smaller computational requirements
Mining Climate and Ecosystem Data
NSF RSV
Impact
•
Facilitate discovery of complex
dependence structures in dynamic
systems
•
Understand relationships between fire
size and frequency to precipitation,
land cover, ocean temperature, etc
•
Characterize impacts of sea surface
temperature on hurricane frequency
and intensity
•
Understand feedback effects, the key
uncertainty in climate science
June 2010
‹#›
Complex Networks
Objectives
•
To develop algorithms for characterization of
network structure, function, and dynamics
•
To enable large-scale comparative analysis of
climate networks to increase prediction confidence
Node degree in a correlation-based
climate network
Computer Science Innovations
Impact
•
Construct multivariate networks to capture
nonlinear spatio-temporal interactions
•
•
Characterize network dynamics at multiple
spatial scales over different time periods
•
•
Scale graph mining algorithms to the
required size and number for comparative
analysis of multiple real-world networks
Understand intricate interplay between
topology and dynamics of the climate
system over many spatial scales
•
Explain the 20th century great climate
shifts from the collective behavior of
interacting subsystems
Mining Climate and Ecosystem Data
NSF RSV
Apply network structure-function-dynamics
methods to complex systems
June 2010
‹#›
Alok Choudhary, NWU
Nagiza Samatova, NCSU
Vipin Kumar, UMN
Predictive Modeling
Objectives
•
To advance nonparametric multivariate spatiotemporal probabilistic regression models to
incorporate nonlinear dependencies
=
GP1
To develop latent variable models to characterize
spatio-temporal climate states
•
To advance multivariate extreme value theory
through geometric and probabilistic
generalizations of quantiles
•
•
•
•
Take into account nonlinear spatial,
temporal, and multivariate dependencies
Minimize assumptions on temporal
evolution, e.g., no`Markov’ assumptions
Deal with curse of dimensionality in
nonlinear extreme value analysis (recency,
frequency, duration) and quantiles
Scale algorithms yet ensure uncertainty
quantification
Mining Climate and Ecosystem Data
GP2
Correlated Gaussian Processes for multivariate spatiotemporal
regression w/ nonlinear dependencies
•
Computer Science Innovations
+
Multivariate quantiles for extreme value analysis
•
Impact
•
•
•
NSF RSV
Uncertainty quantified predictions of climate
variables, e.g., precipitation, over space-time
Abrupt climate change detection, typical
climate characterization
Modeling regional climate change and
extreme climate events, e.g., hurricanes
Applications beyond climate data, e.g.,
finance, healthcare, bioinformatics, networks
June 2010
‹#›
High Performance Computing
Objectives
•
To investigate a “co-design” approach to
scalable analysis of complex dependence
structures
•
To develop parallel and scalable statistical and
data mining functions and kernels
•
To develop sophisticated parallel I/O
techniques that fully exploit complex memory
hierarchy (disks, SSDs, DRAMs, accelerators)
Computer Science Innovations
•
•
The first“co-design” approach that will take into
account exploration of nonlinear associations,
long-range spatial dependencies, I/O
requirements and optimizations, presence of
accelerators and multiple suitable programming
models for HPC systems
Novel optimizations to deal with data-andcompute intensive computations on exascale
platforms with complex memory hierarchies
Mining Climate and Ecosystem Data
NSF RSV
An architecture with accelerators such as GPUs.
Some/all nodes have accelerators and many cores.
Impact
•
•
Analytical kernels that will scale many
data mining algorithms
Analytics algorithms designed for
effective exploration of complex spatiotemporal dependence structures
June 2010
‹#›
Synergy with Physical Modeling Community
CS and Climate Science Synergies
•
•
Physics based models inform data mining
•
More credible variables (e.g., SST)  more
crucial variables (e.g.,
precipitations/hurricanes)
•
Better modeled processes (e.g., atmospheric
physics)  more crucial processes (e.g.,
land surface hydrology)
•
Global and century scale  regional and
decadal scale
•
•
•
Foley, Ganguly, Semazzi, and Snyder
Collaborators: Potter (NASA), Erickson (ORNL)
Active involvement in the IPCC assessment reports
Strong Ties w/ Climate Community
Data mining informs physics based models
•
•
Prominent climate science expertise on the team
•
Better understanding of climate processes
Partnership with ORNL, NCAR, LLNL, LBL climate
modeling and simulated data archival groups:
•
Improved insights for parametrization
schemes
•
Partnership with NASA, ORNL, NOAA observed
and reanalysis data archival and analysis groups:
•
Mining Climate and Ecosystem Data
NSF RSV
Community Climate System Model (NCAR, ORNL,
LBL), Earth Systems Grid, IPCC data archives
(PCMDI at LLNL and HPSS at ORNL), Climate
data analysis and extremes (ORNL, LBL)
Remote sensed satellite and other data (NASA), in
situ sensors like DOE's ARM (ORNL, PNNL),
Reanalysis and observed data assimilation (NOAA)
June 2010
‹#›
Measures of Success in CS Research
Does the proposed research enable significantly better data analysis
than before or enable new kinds of analysis?
l
l
Predictive Modeling
– Improved capabilities for multivariate
spatio-temporal regression that can
simultaneously capture the dependencies
between spatial-temporal objects (e.g.,
temperature, pressure) and help find
novel climate patterns not found by
standard approaches
– New capabilities for detecting extreme
events using quantiles and quantile
regression for multivariate data
Complex Networks
– Detection of known climate patterns and
tracking of the evolution of climate
patterns in time using complex networks
that capture non-linear relationships in
multivariate and multiscale data
Mining Climate and Ecosystem Data
l
l
NSF RSV
Relationship Mining
– Creation of entirely new capabilities in
association mining from approaches that extend
/ modify traditional approaches to handle
spatio-temporal data
– Creation of a completely novel approach for
association analysis that reduces the number of
patterns and the time required to find them
– Improvement in the performance of complex
networks and predictive models by using
nonlinear relationships that more faithfully
represent reality
High Performance Computing
– Scalable analytics code for spatio-temporal data
– Enabling of large scale data driven science that
serves as a demonstration of the value of the
data driven paradigm
June 2010
‹#›
Climate-based Measures of Success
l
Ultimate success is to have data-driven analysis included as a standard
part of climate projections and impact assessment (e.g., for IPCC).
l
Achieving this will require measuring and demonstrating success in two
key areas
–
Filling critical gaps in climate science (Nature, 2010)




–
Improve climate change impact assessments and inform policy


l
Reduction of uncertainty in climate predictions at regional and decadal scales
Improved predictive insights for key precipitation processes
Providing credible assessments of extreme hydro-meteorological events
New knowledge about climate processes (e.g., teleconnection patterns)
Distinguishing between natural and anthropogenic causes
Improved assessments of climate change risks across multiple sectors
Specific use cases will provide the context for these evaluations
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Science Impact View: Evaluation Methodology
• Can we improve the projection (e.g., hurricane frequency, intensity)?
• Can the data-driven model identify:
•
• which regional climate variables are informative/causal (e.g., SST, wind speed)?
• what is the relationship (+/- feedback) between these variables?
To what extent does the data-driven model inform climate scientists?
Forecast &
Compare w/
Train Model Observation Data
Hindcast Prediction
1870
1950
• Reanalysis data
• Observed data
Forecast & Multimodel Comparison
2010 (now)
2100
• Model projection data
• Observed data
Illustrative Example for
Hurricane Frequency Use Case
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
Conceptual View: Driving Use-Cases
CS Technologies
Characterize feedbacks contributing
to abrupt climate regime changes
Yr5+
Yr5
Identify drought and pluvial event
triggers to improve prediction
Understand and quantify the
uncertainty of feedback
effects in climate
Yr4
Project climate extremes at
regional and decadal scales
Yr3
Understand relationships between fire
size and frequency to precipitation, land
cover, ocean temperature, etc.
Yr2
Yr1
Characterize impacts of sea
surface temperature on hurricane
frequency and intensity
Data Challenges
Non-stationary Long-memory, long- Heterogeneous
range, multiscale
non-i.i.d
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
CS Technology View: Deliverables Roadmap
Year 1-2
Multivariate spatio-temporal regression,
Climate state estimation
Extreme value theory based Uncertainty quantification in
on geom./prob. quantiles single regime-shift detection
Multivariate network Handling nonlinear spatio- Dynamic community
construction
temporal interactions
detection & tracking
A notion of “transactions”
in spatio-temporal data
Year 5-5+
Year 3-4
Adapt association analysis
to continuous data
Impact of SST on hurricane
intensity & frequency
Scalable parallel
analysis kernels
Statistical significance
assessment of relationships
Parallel I/O techniques to
exploit memory hierarchy
Identify drought and pluvial event
triggers to improve prediction
Mining Climate and Ecosystem Data
Graph perturbation theory
for “what-if” exploration
NSF RSV
Attribution in multiregime-shifts
Comparative multiple
network analysis
Detection of nonlinear
associations/relationships
Co-design scalable complex
dependence structures
Data driven analysis as
a standard (e.g. in IPCC)
Community software,
benchmarks outreach
June 2010
‹#›
Use Case: Hurricane Frequency
Steps for Discovery of Multivariate Non-linear Interactions
Yr1: UTK/ORNL, UMN, NWU
Yr1.5: NCSU, UMN, NWU
IPCC AR4 Models: CMIP3 datasets
Monthly mean sea surface temperature
Monthly mean atmospheric temperature
Daily horizontal wind at 250/850 hPa
1. Pre-process ancillary climate
model outputs
2. Construct multivariate
nonlinear climate network
3. Detect & track
communities
Steps for Predictive Modeling of Hurricanes
Yr1.5: UMN, UTK/ORNL
Yr2: UMN, UTK/ORNL
Yr2.5: UMN, UTK/ORNL
SST impact on
hurricane
frequency &
intensity
3. Find non-linear
relationships
4. Validate w/
hindcasts
Mining Climate and Ecosystem Data
NSF RSV
5. Determine non-stationary
non-i.i.d. climates states &
Build hurricane models
June 2010
‹#›
Use Case: Regional Precipitation
Step1: Conceptual physics models
(O’Gorman and Schneider 2009) and
relationship mining identify
variables in 3D (space, time, vertical)
neighborhoods with information
relevant for predicting precipitation
Step 2: Precipitation mean and
extremes are projected with ancillary
variables in 3D neighborhoods with
predictive modeling
Step 3: Complex networks are
constructed over oceans using
relationship mining
Step 4: Complex networks develop
proxies for global and regional scale
ocean dynamics leading to set of
potential predictors
Step 6: The (Step 2) 3D
Step 5: Teleconnections are developed to
neighborhood-based predictions
predict regional precipitation change and their
are combined with teleconnctionextremes based on both relationship mining
based predictions with fusion of
and predictive modeling
predictive modeling
Mining Climate and Ecosystem Data
NSF RSV
Step 7: Regional precipitation
prediction gains are run
through cross-validation and
interpreted with climate science
June 2010
‹#›
Use Case: 1930s Dust Bowl
1930s Dust Bowl Drought
Affected almost two-thirds of the
U.S. Centered over the
agriculturally productive Great
Plains Drought initiated by
anomalous tropical SSTs
(Teleconnections)
NASA NSIPP Model: 14 C20C datasets
Ensemble of 14 100-year simulations of the 20th Century
Forced with observed monthly SSTs
Allows for assessment of how much the SSTs control
Great Plains climate variations
Information content in C20C model
simulations along with ancillary
However, the abrupt change
observations can be used to identify
in the precipitation regime
and decade-long persistence and predict drought “triggers” and the
contribution of complex feedbacks to
of drought thought to be
abrupt precipitation regime changes
related to strong landand drought persistence
atmosphere feedbacks (e.g.,
precipitation recycling)
Global SST anomalies averaged over
1932-1938. Boxes represent subregions used to identify the relative
importance of SST anomalies in
initiating the the Dust Bowl drought.
Time series of precipitation
anomalies averaged over the Great
Plains region for 14 ensemble
members of C20C run. Anomalous
Tropical SSTs explain most of the
20th Century drought events.
Mining Climate and Ecosystem Data
NSF RSV
June 2010
‹#›
From Schubert et al. (2004)
Use Case: Regime Shift in Sahara & Sahel
Regime Shift in Sahara
Sudden shift from `Green
Sahara’ to `Desert Sahara’
Sahara consisted of extensive
vegetation, lakes, wetlands
according to Geologic data.
Complex system with more
than one stable state.
Vegetation cover and precipitation patterns
Sudden transition from
vegetated to desert conditions
around 5500 years ago
Sahel zone
-- What were underlying conditions?
-- What triggered the regime shift?
Data analysis for hypothesis generation, testing:
Sahel drought led to widespread famine,
ecosystem degradation, and dispersion of its
inhabitants
Reduction in Precipitation
Such sudden, large change in environmental–or “regime shift”—often
occurs without advance warning. Underlying conditions may predispose
a system to a regime shift. A fairly small event, e.g., storm, drought,
etc., may trigger the shift to a new stable state.
Mining Climate and Ecosystem Data
Northern Africa is dominated by the
Sahara. The transition between the
Sahara and the savanna/tropical forest to
the south occurs in the Sahel.
Key Questions stay unanswered:
Regime Shift in Sahel
Sahel is the transition zone between the
Sahara to the north and the tropical savanna
and evergreen forests to the south
Onset of major 30-year drought over the Sahel
region in 1969
Precipitation patterns are tightly
correlated with vegetation cover patterns.
NSF RSV
+
Reduction in Vegetation
Possible hypothesis:
Land-atmosphere feedback process
June 2010
‹#›