Automatic Road Feature Recognition and Extraction from Remote Sensing Imagery

1
Automatic Road Feature Recognition and Extraction
from Remote Sensing Imagery

• E.F. Granzow
• Iguana Incorporated
• David Fletcher
• Geographic Paradigm Computing
• -------------------------------

2
Presentation Overview

• Research Context
• Basic Approach
• The IPaver Toolkit
• An Example
• Findings and Reflections

3
Research Context

• White paper prepared as resource document for
  NCRST - Safety, Hazards and Disaster Assessment
• Research and software developed as part of NASA
  supported ARC project Development and Automation
  of High Resolution Image Extraction Methodologies
  for Transportation Features

4
Research Context

• Problem Statement
• Feasibility of automating extraction of
  transportation features and potential degrees of
  automation
• Development and use of microcomputer based and
  specially developed software in component
  environment
• Economic rationale for commercial applications in
  this area

5
Basic Approach

• Key Concepts
• Roadway network considered as single object in
  feature identification process
• Image elements regarded in terms of estimated
  probability of inclusion in solution set
• A priori assumption elements are not part of road
  network object
• Approach has roots in pattern recognition and
  computer vision

6
Basic Approach

• Key Concepts (Continued)
• Based on user directed iterative application of
  tools
• Provides immediate feedback on progress
• Scaled for interactive usage

7
Basic Approach

• Roadway Network as Object - Benefits
• Flexible Problem/Image Segmentation
• Processing Efficiencies
• Global/Reusable Classification/Processing Model

8
The IPaver Toolkit

• Functions
• Recomputes DNs based on mean and offsets into n
  equally spaced classes
• Merges two images with user specified weightings
• Calculates given statistic for specified kernel
  and replaces kernel DN with value
• Deletes image features based on a combination of
  size and morphology

9
The IPaver Toolkit

• Functions (continued)
• Identifies and eliminates tenuous connections
  between features based on pixel strings of
  varying types
• Uses pixel distance map to develop single string
  representation of linear features
• Uses width/member seeds to trace and draw road
  elements

10
The IPaver Toolkit

• Support Software
• IguanaSpace - Implements custom IPaver interface
  and parameter management
• ScionImage - Implements macro procedures to
  view images and produce DM and seed files
• IParse - Tiles images to specified
  size and overlap
• Evidence - For known solution reports false
  and true positives and negatives

11
IPaver Interface (IguanaSpace)

• Supports both menu and flowchart access to the
  IPaver toolkit
• Allows direct editing of each functions input
  options and parameters through windows dialogs
• Automatically logs program states and sequences
  for review and reuse
• Will easily accepts changes/additions to IPaver

12
IPaver Interface (IguanaSpace)
13
IPaver An Example

• 1 Meter resolution USGS DOQ
• Residential Area
• Central Albuquerque
• Panchromatic (0-255)
• 1/2 square km

14
IPaver An Example

• Classification by Road Material Type

Parameters DN Mean - 135 Group Interval -
15 Number of Groups - 4
15
IPaver An Example

• Statistical Projection in 3x3 Kernel Neighborhood

Parameters Statistic - Std Deviation Kernel Size
- 3
16
IPaver An Example

• Merging two images

Parameters Weight I1 - 1.0 Weight I2 - 1.0
17
IPaver An Example

• Deletion with Morphological Constraints

Parameters Max Object Size - 300 Max H/W Object
Size - 1500 Min H/W Ratio - .8
18
IPaver An Example

• String Filtering

Parameters Type - cul-de-sac DN Threshold -
255 Cul-de-sac depth - 2 pixels
19
IPaver An Example

• Centerline Development from Distance Map

Parameters None
20
IPaver An Example

• Road Edge Tracing

In development Uses seed to identify both
essential line and road edge Constrains trace
based on degree of curvature and aberrant section
length Controls degree of deviation between EL
and road edge path Uses DM based and source
image together
21
IPaver An Example

• Superimposition of Solution on Base Image

Parameters None
22
Findings Reflections

• The Evidence Model

Model was developed to measure success in
delineating image elements both within and
outside the travelway Solution template was
developed by hand for 256x256 image thumbnail and
compared on pixel by pixel basis to IPaver
derived solution Evaluation phrased as true and
false postives and negatives
23
Findings Reflections

• The Evidence Model (Continued)

image size is '65536' total road rasters
'9517' percent of image '14.5' total true
positives '7219' pp/tp '75.9' total true
negatives '53621' nn/tn '95.7' total false
positives '2398' fp/tn ' 4.3' total
false negatives '2298' fn/tp '24.1' total
efficiency '71.6'
24
Findings Reflections

• Project Conclusions - Technical

• It is possible to automate portions of the
  transportation feature recognition and
  extraction process
• Its feasible to do this without use of legacy
  commercial products (i.e. ERDAS Imagine) and
  large scale hardware
• The probable minimum spatial resolution for
  IPaver is probably about 1 meter

25
Findings Reflections

• Project Conclusions - Economic

• Our original conclusion to not pursue commercial
  options may be obsolete
• New interest and funding for transportation
  feature and centerline extraction may present
  new commercial potentials
• Changes/evolution of image provider licensing
  policies have enhanced these potentials
• Spaceborne imagerys near total reliance on
  defense applications and procurements creates
  continuing commercial uncertainties

26
Findings Reflections

• Some General Observations

• New multi- and hyperspectral high resolution
  imagery offers avenues to enhance the extraction
  process
• Urban scenes present greatest challenges due to
  oblique shadow effects of urban canyons and
  other urban specific issues
• Likely applications are for suburban/rural high
  growth and unmapped areas