Information Technology and Systems Center

1
Data Mining in Earth Sciences
Rahul Ramachandran, Sara Graves and Ken
Keiser Mathematical Challenges in Scientific
Data Mining IPAM January 14-18, 2002

• Information Technology and Systems Center
• University of Alabama in Huntsville
• http//datamining.itsc.uah.edu

2
Outline

• Introduction
• ADaM System
• Data Mining Taxonomy for Earth Science
• Event/Relationship based
• Application Examples
• Dimensionality Reduction
• References

3
Reasons for Data Mining of Earth Science Data

• Greatly increased data volume due to improvements
  in data collection/access/availability/storage
  technology (instruments, computational resources,
  internet)
• Terra are about 1 terabyte per day - more than
  can be analyzed by conventional means
• High variability in data formats and content
• Need for high returns on expensive data
  investments
• Need for improved access/availability of data,
  information and knowledge
• Need for higher level products for the
  non-specialist and interdisciplinary/cross-domain
  researchers
• Questions/queries are getting more complex due,
  in part, to heterogeneous nature of the data

4
Characteristics of Earth Science Data

• High variability
• Type
• Geostationary
• Polar Orbiting
• Structure
• Raster
• Vector
• Resolution
• Fine AVHRR 1km
• Coarse SSM/I 20km
• Multi/Hyper Spectral
• Processing stage
• Level 0 Raw data instrument counts
• Level 1 Annotated with Geo-reference information
• Level 2 Transformed by algorithm into
  geophysical parameter
• Level 3 Spatial/Temporal resampling
• Level 4 Includes additional model data

5
Characteristics of Earth Science Data

• Need to know physical basis (domain knowledge)
  before applying statistical techniques
• Multiple time scales
• Wide variety of data formats
• Includes spatial/temporal information
• Typically needs domain-specific algorithms

6
ADaM History

• Algorithm Development and Mining (ADaM) System
• The system provides knowledge discovery, feature
  detection and content-based searching for data
  values, as well as for metadata.
• It contains over 120 different operations to be
  performed on the input data stream.
• Operations vary from specialized atmospheric
  science data-set specific algorithms to different
  digital image processing techniques, processing
  modules for automatic pattern recognition,
  machine perception, neural networks and genetic
  algorithms.
• Developed a Event/Relationship Search System for
  the environment

7
ADaM Engine Architecture
Preprocessed Data
Patterns/ Models
Results
Data
Translated Data
Processing

Preprocessing
Analysis
Selection and Sampling Subsetting
Subsampling Select by Value Coincidence
Search Grid Manipulation Grid Creation
Bin Aggregate Bin Select Grid Aggregate
Grid Select Find Holes Image Processing
Cropping Inversion Thresholding Others...
Clustering K Means Isodata
Maximum Pattern Recognition Bayes Classifier
Min. Dist. Classifier Image Analysis
Boundary Detection Cooccurrence Matrix
Dilation and Erosion Histogram Operations
Polygon Circumscript Spatial Filtering
Texture Operations Genetic Algorithms Neural
Networks Others...
8
ADaM Mining Environment
Data Mining Server
Mining Results
Event/ Relationship Search System
9
Data Mining Taxonomy
10
Event-based Mining

• Known events/Known algorithms
• Tropical Cyclones from AMSU-A data
• Known events/Learned algorithms
• Rainfall estimation from SSM/I data
• Lightning Detection from OLS data
• Unknown event/Unknown algorithm
• Target Independent Mining

11
Known Event/Known Algorithm
I know what phenomena to detect and I have
the algorithm to do so!
Results
Add algorithm to Mining Environment

• Relationship analysis
• Coincidence searches
• Input for other algorithms

Earth Science Data Sets
12
Tropical Cyclone DetectionEstimating Maximum
Wind Speed

• Scientist Dr. Roy Spencer (GHCC/MSFC NASA)
• Data used Advanced Microwave Sounding Unit-A
• Radiometer can detect temperatures at different
  levels of the atmosphere
• Surface winds in tropical cyclones are directly
  related to the warm middle- and upper-atmosphere
  temperatures which exist around the cyclone
  center
• AMSU-A measures this warmth at several
  frequencies near 55 gigahertz (GHz)
• Calibrated using aircraft reconnaissance
  measurements in tropical depressions, tropical
  storms, and hurricanes from the 1998 Atlantic
  hurricane season
• Tropical cyclone detection based on ice
  scattering, water vapor and wind speed

13
Tropical Cyclone DetectionEstimating Maximum
Wind Speed
Advanced Microwave Sounding Unit (AMSU-A)Data

• Water cover mask to eliminate land
• Laplacian filter to compute temperature
  gradients
• Science Algorithm to estimate wind speed
• Contiguous regions with wind speeds above a
  desired
• threshold identified
• Additional test to eliminate false positives
• Maximum wind speed and location produced

Calibration/ Limb Correction/ Converted to Tb
Hurricane Floyd
Data Archive
Mining Environment
Result
Results are placed on the web and made available
to National Hurricane Center Joint Typhoon
Warning Center
14
Known Event/Learned Algorithm
Data Mining System
I know what phenomena I want to detect but I
do not know the characteristics of the phenomena
Results
Refine your algorithm using iteration

• Relationship analysis
• Coincidence searches
• Input for other algorithms

Earth Science Data Sets
15
Rainfall Estimation and Identification Study
using SSM/I data

• Scientist Dr. Steve Goodman (GHCC/MSFC NASA)
• To determine whether generic pattern recognition
  techniques could be applied to SSM/I data to
  detect rain
• Minimum Distance Classifier, Back-propagation
  Neural Network and Discrete Bayes Classifier were
  compared against a Science Algorithm ( WetNet PIP
  Algorithm)
• US Composite rainfall product was used as ground
  truth

Subsetted SSM/I data
NEXRAD Composite data
16
Rainfall Estimation and Identification Study
using SSM/I data

• SSM/I and US rain data over southeastern United
  States for the period January and July 1995 were
  compared in the study
• SSM/I and Radar data were gridded and registered
  to establish spatial and temporal coincidence
• BPNN performance was comparable to that of the
  WetNet PIP SSM/I rain rate algorithm
• Performance of Bayes classifier was not as good
  as that of the WetNet PIP SSM/I rain rate
  algorithm. This is perhaps due to the small
  sample size used for estimating density functions
  of the two classes (rain and non-rain)

17
Lightning Detection in Operational Linescan
System (OLS) Images

• Scientist Dr. Steve Goodman (GHCC/MSFC NASA)
• To identify lightning streaks in night time
  portions of OLS images
• OLS is carried by DMSP satellites and produces a
  visible and thermal image
• Lightning shows up as bright horizontal streaks
  as do city lights and moonlight reflected off the
  clouds
• Approach based on morphological filtering and
  gradient detection was selected
• Both visible and thermal band used

18
Lightning Detection in Operational Linescan
System (OLS) Images

• Erosion and dilation was used to find areas
  in/near clouds, other areas were removed
• Gradient detection in the direction of satellite
  propagation was applied to the visible image to
  extract horizontal streaks
• Texture measures were used to identify areas of
  small patchy cloud cover which exhibited small
  bright streaks
• Genetic algorithm was used to tune parameters of
  the classification during training

Results ( Accuracy)
Correctly Detected
False Positives
False Negatives
Training Results
80
0.7
19.2
Test Results
78.2
4.3
17.3
19
Unknown Event/Unknown Algorithm
Data Mining System
I want to find anomalies in the data sets !
Results
Let the miner discover it

• Relationship analysis
• Coincidence searches
• Input for other algorithms

Earth Science Data Sets
Example Target Independent Mining
20
Target Independent Mining of SSM/I Data

• Mine for data in a target independent manner (no
  specific phenomena under consideration)
• Interested in transient phenomena that move
  through an area
• Transient phenomena characterized as deviation
  from normal
• Objective Data Reduction with minimum loss of
  information
• Size of remotely sensed data prevents it from
  being maintained on-line
• Data is archived in much slower tertiary storage
• Need to develop techniques to minimize the need
  for data access from the tertiary storage
• Procedure Overlay the earths surface with a
  constant grid consisting of cells
• For each cell a maximum and minimum trend line is
  computed
• Maximum trend line is computed by forming a set
  of maximum values for a day over some period
  (month)
• Median for a series of months is used to form the
  maximum trend line
• Same procedure used to calculate minimum trend
  line

21
Target Independent Mining of SSM/I Data
Trend Lines Represent What Is Normal
22
Target Independent Mining of SSM/I Data

• Extracted metadata not oriented toward any
  particular transient phenomena
• Laboratory tests show 98 data compression while
  preserving 92 of MCSs detectable in raw data
• MCS events represented only 6.7 of extracted
  metadata

23
Relationship-based Mining

• Coincident Association
• VARGA Algorithm for multispectral data
• Localized Spatial Association
• Cumulus Cloud Classification in GOES Imagery
• Temporal Association

24
Coincident Association Mining

• Use Market Basket analysis to mine for
  
  
  
  
  association rules in vector data
• Rule has form X ?Y
• Rule characterized by
• Support
• of vector instances that have X ? Y
• How likely the rule is applicable?
• Confidence
• What of vector instances that contain X also
  contain Y?
• Estimate of conditional probability

25
Coincident Association Applied to Multi-spectral
Data Mining

• Developed and implemented Vector Association Rule
  Generation Algorithm (VARGA) as a modification to
  market-basket association rule mining.
• Modified to minimize memory usage for large
  multi-spectral satellite data such as SSM/I (90
  megabytes per day uncompressed)
• Example SSM/I Rule
• 19V, 180.0 37H, 140.0 -gt 37V, 200.0
  0.117037 0.945986

26
Localized Spatial Association Mining

• Extract association rules to characterize texture
  (Dissertation of Dr. John Rushing)
• Each pixel on an nxn neighborhood is
  characterized by the triple (X,Y,I)
• The X and Y offsets from the pixel at the
  neighborhood center
• Its intensity I
• Association rules can then be characterized by
  relationships between the triples

27
Association Rule Example

• The rule specified in figure can be applied to
  this image in 9 of the 16 pixel locations due to
  the pixel offsets in the rule.
• Of these 9 locations, the antecedent matches at 5
  locations, and both the antecedent and consequent
  match at 3 locations.
• This yields a support of 3/9 33.33 and a
  confidence of 3/5 60.

 
28
Association Rule Example
29
GOES Cumulus Cloud Classification Why Texture
Features?

• Cumulus cloud fields have a very characteristic
  texture signature in the GOES visible imagery

30
GOES Cumulus Cloud Classification The Need

• Cloud systems are important modulators of earths
  radiation budget
• Large uncertainties are associated with cloud
  radiative forcing
• Radiative energy budget is impacted by change in
  distribution of clouds
• Cumulus clouds are a cloud field type that could
  respond strongly to climate change
• Knowledge of cloud geometry, size and spatial
  distribution is needed for the representation of
  cumulus clouds in radiative transfer models
• To derive models of cloud field characteristics,
  automated cumulus cloud detection schemes are
  required to analyze large amounts of data

31
GOES Cumulus Cloud Classification Purpose of
this study

• Compare different techniques for detecting
  Cumulus cloud fields in Geostationary Operation
  Environmental Satellite (GOES)
• Comparison based on
• Accuracy of detection
• Amount of time required to classify
• Feature measures used along with the Maximum
  Likelihood Classifier
• Texture features
• Gray Level Co-Occurrences Matrix
• Gray Level Run Length Features
• Association Rules
• Edge Detection Features
• Sobel Filter
• Laplacian Filter
• Combination of Sobel and Laplacian Filter

32
GOES Cumulus Cloud Classification Texture
Features (1)

• Gray Level Co-Occurrence Matrix
• First texture feature vector to be developed
• GLCM is used as a benchmark
• It is based on positional operator
• Positional operator defines relationship between
  pixels in terms of x,y offset or as a distance,
  angle offset
• Co-occurrence matrix is an NxN matrix where N is
  the number of gray levels and functions are
  computed on the matrix
• Gray Level Run Length Features
• Gray level statistical features based on
  homogeneous gray level runs
• Run is a series of consecutive pixels of the same
  intensity
• Run length are at orientations in increments of
  45 degrees starting at 0 degrees

33
GOES Cumulus Cloud Classification Texture
Features (2)

• Association Rules
• Often used in business applications to identify
  relationships in databases
• Adapted to discriminate textures in images
• Based on frequently occurring local image
  structures

Triples ( Pos X, Pos Y, Pixel Intensity) Rule
(0,0,2) (1,1,2) gt (1,0,0) Then calculate
Support and Confidence of this Rule
34
GOES Cumulus Cloud Classification Edge Detection
Features

• These techniques are used for detecting
  discontinuities in an image
• These techniques apply a local derivative
  operator on the image
• Sobel Filters
• It calculates the magnitude of rate of change of
  gray level and the direction of this change
  vector
• Magnitude Gx Gy
• Direction tan-1(Gx/Gy)
• Gx (z7 2z8 z9) (z1 2z2 z3)
• Gy (z3 2z6 z9) (z1 2z4 z7)
• Laplacian Filters
• It is a second order derivative
• F(z) 4z5 (z2 z4 z6 z8)

z1 z2 z3
z4 z5 z6
z7 z8 z9
35
GOES Cumulus Cloud Classification Experiment
Process

• Training
• Samples selected from 1000x1000 GOES scene
• Only two classes are used Cumulus and Others (
  includes background)
• For validation, samples were labeled by at least
  two experts and only pixels where experts agreed
  were used for training
• Maximum likelihood classifier was trained using
  GLCM, GLRL, Association Rules and Edge detection
  features
• Window size was varied 5x5 11x11
• Testing
• 12 different GOES images (512x512) where used for
  testing
• Classification results were compared against
  expert labeled images
• Confusion matrix, classification accuracy and
  experiment run times were calculated

36
GOES Cumulus Cloud Classification Sample Result
Original
GLRL
Association Rules
GLCM
Expert Labeled
Sobel
Sobel Laplacian
Laplacian
37
GOES Cumulus Cloud Classification Conclusions

• Accuracy
• Best results using texture features
• GLRL (78) with a filter size of 11x11
• Association Rules (75) with a filter size of 5x5
• GLCM gave the worst results (51-55)
• Best results using edge detection filters
• Sobel Filter (78) with a filter size of 11x11
• Laplacian (73) with a filter size of 9x9
• Laplacian and Sobel (75) with a filter size of
  9x9
• Timing Results
• Times were calculated on an 933MHz Pentium III
  processor PC with 512 MB memory
• Texture feature techniques in general required an
  order of magnitude more time than edge detection
  filters

38
Dimensionality Reduction Mesoscale Convective
System (MCS) Detection
Scientists
Populating Knowledge Base (reducing data volume )
Scientists

• Define the Experiment
• Select algorithm (Devlin)
• Automatic extraction of MCSs from SSM/I
• data

Mining Results MCSs
Knowledge Base Event/ Relationship
Search System
SSM/I Data
39
Dimensionality Reduction Research Analysis

• Reduced amount of data
• Allow scientists to pose questions
• and get results
• Allow easy visualization
• Maximize knowledge discovery/
• minimize data handling
• Scientists can refine their
• knowledge repository
• Answer the science questions

• Analysis
• Find MCSs over river basins in Middle East?
• Data Sets
• MCSs
• River basin data set
• Political boundaries

Scientists
Mining Results MCSs
Event/ Relationship Search System
Knowledge Base Event/ Relationship
Search System
SSM/I Data
40
Dimensionality Reduction Knowledge Reuse

• Climatological Study of MCSs
• What is the latitudinal distribution of
• MCSs?
• Which continent has more MCSs?
• What is the size distribution of the
• MCSs for JUN-JUL-AUG?
• What is the relationship between the
• number of MCSs and their intensities?
• Do results vary for El-Nino years?

Scientists

• Knowledge Reuse

Mining Results MCSs
Event/ Relationship Search System
Knowledge Base Event/ Relationship
Search System
SSM/I Data
41
Event/Relationship Search System

• Allows users to conduct coincidence searches and
  relationship tests between mined phenomena and a
  variety of parameters
• Parameters include geographic regions,
  political boundaries, or other named phenomena
  for a specific time period

42
References

• Graves, Sara J., Thomas Hinke, Shanlini Kansal,
  "Metadata The Golden Nuggets of Data Mining",
  First IEEE Metadata Conference, Bethesda,
  Maryland, April 16- 18, 1996
• Hinke, Thomas, John Rushing, Shanlini Kansal,
  Sara J. Graves, Heggere S. Ranganath, "For
  Scientific Data Discovery Why Can't the Archive
  be More Like the Web", Proceedings Ninth
  International Conference on Scientific Database
  Management, Evergreen State College, Olympia,
  Washington, August 11-13, 1997
• Hinke, Thomas, John Rushing, Heggere S.
  Ranganath, Sara J. Graves, "Techniques and
  Experience in Mining Remotely Sensed Satellite
  Data", Artificial Intelligence Review 14 (6)
  Issues on the Application of Data Mining, pp
  503-531, December 2000
• Hinke, Thomas, John Rushing, Shanlini Kansal,
  Sara J. Graves, Heggere S. Ranganath, Evans
  Criswell, "Eureka Phenomena Discovery and
  Phenomena Mining System", AMS 13th Intl
  Conference on Interactive Information and
  Processing Systems (IIPS) for Meteorology,
  Oceanography and Hydrology, 1997

43
References

• Hinke, Thomas, John Rushing, Heggere S.
  Ranganath, Sara J. Graves, "Target-Independent
  Mining for Scientific Data Capturing Transients
  and Trends for Phenomena Mining", Proceedings
  Third International Conference on Data Mining
  (KDD-97), Newport Beach, California, August
  14-17, 1997
• Keiser, Ken, John Rushing, Helen Conover, Sara J.
  Graves, "Data Mining System Toolkit for Earth
  Science Data", Earth Observation (EO)
  Geo-Spatial (GEO) Web and Internet Workshop,
  Washington, D.C., February 1999
• Rushing, John, Heggere S. Ranganath, Thomas
  Hinke, Sara J. Graves, "Using Association Rules
  as Texture Features", IEEE Transactions on
  Pattern Analysis and Machine Intelligence, Vol
  23, No. 8, 845-858, 2001
• Nair, Udaysankar J., John Rushing, Rahul
  Ramachandran, Kwo-Sen Kuo, Sara J. Graves, Ron
  Welch, "Detection of Cumulus Cloud Fields in
  Satellite Imagery", The International Symposium
  on Optical Science, Engineering and
  Instrumentation, Denver, 1999
• Nair, U., J. Rushing, R. Ramachandran, R. Welch,
  and S. J. Graves, Detection of boundary layer
  cumulus cloud fields in GOES satellite imagery,
  submitted to Journal of Applied Meteorology,
  September, 2001