Geographical Information System (GIS) to Knowledge

1
Geographical Information System (GIS) to
Knowledge
Peter Bajcsy, Ph.D. Research Scientist, ALG,
NCSA Adjunct Assistant Professor, CS and ECE
Departments, UIUC
2
Outline

• Introduction to GIS Related Decision Making
• Decision Making Scenario
• Input Information Extraction and Representation
• Georeferencing, Registration and Raster
  Information Extraction
• Boundary Aggregation and Evaluation
• Error Evaluation of New Boundary Aggregations and
  Decision Making
• Other Interesting Problems
• Summary

3
Acknowledgement

• Project Team Members Peter Bajcsy, Peter Groves,
  Sunayana Saha, Tyler Alumbaugh, Sang-Chul Lee
• Support Michael Welge, Loretta Auvil, Dora Cai,
  Tom Redman, David Clutter, Duane Searsmith, Lisa
  Gatzke, Andrew Shirk, Ruth Aydt, Greg Pape, David
  Tcheng, Chris Navaro, Marquita Miller.

4
GIS Decision Making

• Decision Making Driven by applications
• GIS Application Examples
• Urban planning Where should we build new fire
  stations? Hospitals? Schools?
• Agriculture Should we apply pesticides ? How
  much per acre?
• Forestry Should we cut trees ? How many ?
• Wildlife What should we do to help endangered
  species ?
• Watershed/Catchment management Would we have an
  surface erosion problem after cutting trees?
• Archeology Where should we excavate to find
  prehistoric sites?
• Geology and mining Is is safe to build houses in
  the location (lat/long)?

5
Decision Making Motivation/Process/Goal

• Motivation Seek Optimal Solutions
• How?
• Evaluate Multiple Solutions
• Assess Quality and Significance of Input
  Variables
• Predict Output Variables
• Impose Application Specific Constraints
• Construct Optimality Metric
• Goal To Make the Most Optimal Decision
• Planning Decision
• Management Decision

6
GIS Applications Natural Resources

• Agriculture land conservation, market analysis,
  farm planning
• Forestry timber assessment and management,
  harvest scheduling and planning, environmental
  impact assessment, pest management
• Wildlife habitat assessment and management,
  rare species studies
• Catchment management runoffs and erosion
  modeling, sedimentation and water quality
  studies, integrated catchment management
• Archeology prediction of prehistoric sites,
  site vandalism studies
• Geology and mining geologic hazard mapping,
  oil, gas and mineral studies, open pit mine
  design
• Watershed management - guidelines for regulation

7
GIS Applications Urban Resources

• Spatial Distribution of
• Utilities
• Hospitals
• Schools
• Fire Stations
• Management of Storm Water
• Crime Analysis
• Waste Collection Routing
• Hazardous Waste Transportation
• Disease Outbreak Patterns

8
Feature Analysis Decision Making Scenario

• Problem Formulation
• Analyze FBI Crime Reports and Other Features per
  County, Zip Code, US Census Bureau Track or
  Blocks, and Support Decision Making Process about
  Crime Prevention
• Features
• FBI Crime Reports, Various Boundary Definitions,
  Forest Cover Maps, DEM, Satellite and Aerial
  Data, Prison Point Information

9
Problem Statement

• Problem Statement search for the best partition
  of any geographical area that is
• (a) based on raster or point information,
• (b) formed by aggregations of known boundaries,
• (c) constrained or unconstrained by spatial
  locations of know boundaries and
• (d) minimizing an error metric.
• Raster or Point Information
• Grid-based information, e.g., from satellite or
  air-borne sensors
• Geographical point information, e.g., from GPS or
  address data base
• Boundaries (Vector Data)
• Man-made, e.g., Counties, US Census Bureau
  Territories
• Defined by environmental characteristics, e.g.,
  Eco-regions, Historical iso-contours
• Spatial Constraints and Error Metric
• Defined by applications

10
Proposed Approach
Feature Analysis and Decision Making
Data Preprocessing
11
Input Information Extraction and Representation
2D or 3D
12
Input Information Extraction and Representation
13
Data Types and Representation Examples

• Raster Information GeoImage Object
• Boundary Information Shape Object
• Tabular Information Table Object
• Neighborhood Information NBH Object

14
Raster Information File Formats

• USGS Digital Elevation Data (DEM) Files
• Header file with georeferencing information
• Floating point values, 30 m spatial resolution,
  IL coverage, published in 2002
• TIFF Files
• Georeferencing information from
• One or more standardized files are distributed
  along with TIFF image data as .tfw and/or .txt
  files.
• The metadata is encoded in the image file using
  private TIFF tags.
• An extension of the TIFF format called GeoTIFF is
  used.
• Forest labels, 1km spatial resolution,
• Forest Cover Types 29 labels, USA coverage,
  published in 2000
• Forest Fragmentation Index Map of North America,
  8 labels, USA coverage, published in 1993
• Land use labels, 1km spatial resolution, world
  wide coverage, published in 2001

15
Boundary Information File Formats

• Computational Tradeoffs Between Vector
  Information Retrieval and Data Storage
• US Census Bureau TIGER Files
• Elaboration of the chain file structure (CFS)
• Used record files 1, 2, I, S, P
• Environmental Systems Research Institute (ESRI)
  Shapefiles
• Location list data structure (LLS)
• shp, shx, dbf files
• TIGER to ESRI Shapefiles

16
Point Information File Formats

• FBI Crime Reports
• United States Crimes Database, years 94-98, USA
  states, reports per county, published in 2001
• United States Crimes Database, years 98-00, IL
  state, reports per county, published in 2002
• Entries
• Theme_Keyword crime, arrests, murder, forcible
  rape, rape, robbery, aggravated assault, assault,
  burglary, larceny, motor vehicle theft, theft,
  arson
• Challenges
• Multiple Files
• Varying notation
• Association with geographical boundary
  information

17
Data Size

• Data size driven
• operations
• Sub-setting
• Sub-sampling
• Cropping
• Zooming

18
Formation of Vector Data

• Iso-contour extraction from historical maps
• Segmentation and clustering of raster data into
  homogeneous regions

19
(No Transcript)
20
Georeferencing Data Sets and Raster Information
Extraction
21
Georeferencing Data Sets and Raster Information
Extraction
22
Why Registration Georefencing ?
23
(No Transcript)
24
Georeferencing Based on Data Types

• Raster and Raster
• Vector and Vector
• Raster and Vector

25
Georeferencing Based on Coordinate Systems
26
Image Registration Without Georeferencing
Information

• Registration the act of correct alignment
• Registration Steps
• Determine locations of salient features in
  multiple data sets (spatial correspondence)
• Select registration transformation (geometric
  transformation)
• Evaluate accuracy with a metric
• Is my registration correct?

27
Image Registration Computation

• Extract features from multiple images
• select feature space
• Search for a finite set of parameters for a
  transformation function (also called deformation,
  warping function)
• Select transformation function
• Select similarity metric
• Choose a search technique which can reduce the
  computational cost search strategy

28
(No Transcript)
29
Raster Information Extraction Continuous Variable





Elevation Statistics Per County
Standard Deviation
Sample Mean
Skew
Kurtosis
30
Raster Information Extraction Categorical
Variable
Frequency of Occurrence
31
Feature Driven Boundary Aggregation and Evaluation
32
Feature Driven Boundary Aggregation and Evaluation
33
Aggregation of Territories

• Functionality Aggregate territories based on
  similarity of attributes
• Example if auto theft in County 1 is similar to
  auto theft in County 2 then aggregate County 1
  and 2
• Aggregation Constraints
• Segmentation aggregation with neighborhood
  constraint
• Clustering aggregation without neighborhood
  constraint
• Hierarchy of results
• Desired Number of Aggregations
• Maximum Internal Dissimilarity of Aggregations
• Number of Layers in Hierarchy
• Resulting Labels Saved in ESRI Shape Files

34
Territorial Aggregation Input Parameters

• Tools Main Frame Menu/ImLabels/SegGeoPts and
  ClustGeoPts
• Input Parameters
• "input .shp or .shx filename ESRI shape files
  including .nbr file.
• select features Any combination of numerical
  features in .dbf file.
• select weights Any numerical feature for
  weighting territories during aggregation.

Output Aggregations Results
Aggregations Results in dBASE (.dbf) File
IL Counties
Statistics
35
Spatially Unconstrained Boundary Aggregation

• Hierarchical clustering of crime data with the
  exit criterion being the number of clusters and
  the clustered feature being auto theft in 2000
  leads to six aggregations.

Tabular Display
Geographical Display
Boundaries
Boundary Aggregations
36
Spatially Constrained Boundary Aggregation

• Hierarchical segmentation and hierarchical
  clustering of oak hickory feature with the exit
  criterion of 18 numbers of county aggregations

With Spatial Constraint
Without Spatial Constraint
Boundaries
Boundary Aggregations
37
Boundary Aggregation With Hierarchical Output

• Hierarchical segmentation of extracted forest
  statistics (oak hickory occurrence) with two
  output partitions.

43 aggregations
21 aggregations
Boundaries
Boundary Aggregations
38
Visualization of Aggregation Results

• Functionality Visualization of Labels and
  Hierarchy of Labels
• Data to Visualize Labels in dBASE (.dbf) file,
  Hierarchy of Labels in dBASE (.dbf) file
• Input Parameters None (ShowResults and
  ShowGeoResults)

IL Counties
39
Error Evaluations of New Territorial Partitions

• Error evaluation of partitions obtained by
  clustering and segmentation of mean elevation
  feature per Illinois county with Variance error
  metric

40
Territorial Partition Error Evaluation

• Error Metrics
• Variance
• Normalized Variance
• City Block
• Normalized City Block

41
Geographical Error Evaluations and Decision Making

• Geographical error evaluation of partitions
  obtained by clustering and segmentation of mean
  elevation feature per Illinois county with
  Variance error metric

Partition Index
Eval0
Eval1
Eval2
Eval3
42
(No Transcript)
43
Example of Feature Analysis

• Crime Data Over Multiple Boundary Definitions
• Evaluation Parameters
• Features Auto Theft, years 1999, 2002
• Error Evaluation using Variance
• Number of aggregations 50
• Spatial Aspect
• Boundary type
• Spatial constraints on aggregations (contiguity
  ?)
• Optimal number of aggregations
• Temporal Aspect
• Temporal changes of crime report features
• Optimality Metric
• Feature weighting
• Cost/evaluation function

44
Example Results Auto Theft
Clustering00
Clustering Error
Clustering99
Segmentation99
Segmentation00
Segmentation Error
45
Example Results Auto Theft

• Crime data evaluations
• AutoTheft 00, 99
• Seg
• Evaluation Results
• Evaluation Metric Variance
• Number of Evaluations 2
• Eval 0 Seg_0
• Number of Aggregations54
• Error5051.924242424242
• Eval 1 Seg_0
• Number of Aggregations54
• Error420688.6572463768

• Crime data evaluations
• AutoTheft 00, 99
• Clust
• Evaluation Results
• Evaluation Metric Variance
• Number of Evaluations 2
• Eval 0 Clust_0
• Number of Aggregations37
• Error5575.483333333334
• Eval 1 Clust_0
• Number of Aggregations36
• Error4387.848695652174

46
Decision Making

• Which global partition minimizes a chosen error
  metric?
• Which partition minimizes a chosen error metric
  at a selected boundary definition?
• What is the geographical error distribution given
  a territorial partition?

47
Additional Interesting Problems

• Feature Selection
• High-Dimensional Data Visualization
• Data Fusion

48
Feature SelectionProblem Formulation

• General Formulation
• Given a set of candidate features, select the
  best subset in a classification problem.
• Specific Formulation with the Focus on
  Multispectral and Hyperspectral Image Analysis
• Given a set of spectral bands and ground
  measurements, select the best subset of bands
  with the best predictive accuracy for the ground
  measurements

49
Hyperspectral Band Selection

• Spectral band selection problem with ground
  measurements
• Soil conductivity continuous variable
• Grass label categorical variable
• Input Data sets
• Regional Data Assembly Centers Sensor (RDACS),
  model hyperspectral (H-3), which is a 120-channel
  prism-grading, push-broom sensor developed by
  NASA Plus ground soil conductivity from
  Agriculture department, UIUC
• AVIRIS Sensor developed by NASA Plus Gramma grass
  labels from UCSD

50
Input Hyperspectral Data
RDACS Sensor Data
Ground Measurements e.g., Soil Conductivity
51
Input Hyperspectral Data
AVIRIS Sensor
White spatial location of valid labels grass
categories
Three bands
52
Hyperspectral Band Selection

• Objectives
• What is the optimal number of selected bands?
• Which bands should form the optimal subset?
• What methods should be used for band selection?
• Challenges
• Computational complexity
• combinations of bands with each supervised
  method, where nb is the number of bands
• No free lunch (NFL) theorem

53
Proposed Approach

• Use multiple unsupervised methods to rank the
  spectral bands
• Test those band choices with multiple supervised
  methods (Wrapper Method)
• Select those band combinations that minimize
  classification error given a particular set of
  ground measurements
• Select that pair of unsupervised and supervised
  methods that minimize classification error

54
Cross Validation

• How to measure accuracy?
• Testing on same data set as trained on is
  cheating.
• Overfitting model becomes lookup table, cannot
  deal with new examples
• Use n-fold cross-validation
• Split data set into n subsets
• Train n models, with one subset held out for
  each
• Find average error rate over the holdout sets
  with predictions made by the appropriate model.

55
Analysis of Continuous Variable Soil Conductivity
Unsupervised Ranking
56
Analysis of Continuous Variable Soil Conductivity

• Linear Regression
• Prediction is a linear combination of the inputs
  (Xe)
• Learning becomes finding the coefficients (Find
  that satisfies g(T)).

57
Analysis of Continuous Variable Soil Conductivity
Supervised Linear Regression Based Classification
58
(No Transcript)
59
Data Fusion of GIS Data

• Map Mosaicking Challenges Heterogeneous Data
  Sets
• Multiple Data Sets with Different
• Data Types
• BYTE, SHORT, INT, LONG, FLOAT, DOUBLE
• Spatial Resolutions
• HORIZONTAL AND VERTICAL RESOLUTIONS
• Spectral Resolutions
• NUMBER OF BANDS
• OVERLAPPING ALIGNED OR MISALIGNED WAVELENGTHS
• Geographic Projections
• DATUMS (3D GLOBE MODEL)
• 2D PROJECTIONS

60
Map Mosaicking Task

• Task Automatic Data Fusion that Resolves
  Dissimilar Projections, Spatial Resolutions, Data
  Types, and Number of Bands
• Example of a Simplified Task
• Puzzle-Piece Problem
• Identical projections, resolutions, data types,
  number of bands
• Different geographic locations

61
ArcMap vs I2K Reprojection Results

• Tentative Results
• Resampling Method Highly Accurate
• On-the-fly projection not quite as good
• I2K in between

62
(ArcMap on-the-fly reprojection shown)
63
Table of Results
64
Summary

• Applications of GIS tools
• Remote Sensing
• Agriculture
• Hydrology
• Water Quality Survey
• Atmospheric Science
• Military
• Socio-Economics
• Interested ? Useful ? Let us know.
• Email pbajcsy_at_ncsa.uiuc.edu
• Reading http//alg.ncsa.uiuc.edu/do/documents
• Teaching http//cee.uiuc.edu/people/kumar1/cee498
  HI/lectures.htm

65
Documentation
66
References

• Journal Papers
• Peter Bajcsy and Peter Groves, Methodology For
  Hyperspectral Band Selection, Photogrammetric
  Engineering and Remote Sensing journal, accepted
  June 2003.
• Conferences
• Peter Groves and Peter Bajcsy, Methodology for
  Hyperspectral Band and Classification Model
  Selection, IEEE Workshop on Advances in
  Techniques for Analysis of Remotely Sensed Data,
  Washington DC, October 27,2003.
• Peter Bajcsy and Tyler Jeffrey Alumbaugh,
  Georeferencing Maps With Contours, Proceedings
  of the 7th World Multiconference on Systemics,
  Cybernetics and Informatics (SCI 2003), Orlando,
  Florida, July 27-30, 2003.
• Peter Bajcsy, Automatic Extraction Of
  Isocontours From Historical Maps, Proceedings of
  the 7th World Multiconference on Systemics,
  Cybernetics and Informatics (SCI 2003), Orlando,
  Florida, July 27-30, 2003.
• ALG Technical Reports
• Tyler Alumbaugh, and Peter Bajcsy,
  Georeferencing Maps with Contours in I2K,
  Technical Report NCSA-ALG-02-0001, October 2002
• Peter Groves, Sunayana Saha and Peter Bajcsy,
  Boundary Information Storage, Retrieval,
  Georeferencing and Visualization, Technical
  Report NCSA-ALG-03-0001, February 2003
• Peter Bajcsy, Peter Groves, Sunayana Saha, Tyler
  Alumbaugh, and David Tcheng, A System for
  Territorial Partitioning Based on GIS Raster and
  Vector Data, Technical Report NCSA-ALG-03-0002,
  February 2003