An Image Information Mining Framework for Rapid Assessment and Response in Coastal Disaster Events

1
An Image Information Mining Framework for Rapid
Assessment and Response in Coastal Disaster Events

• Surya S. Durbha, Roger L. King, Nicolas H.
  Younan, David Shaw, and Ioana Banicescu

2
Outline

• Motivation/Background
• Feature Selection/generation
• Wrapper-based Genetic Algorithm
• Semantic Model
• Results

3
Motivation

• In a coastal disaster event, it is necessary to
  obtain information about water level (depth),
  winds, currents, waves, temperature-salinity
  stratification in real time and predictions of
  water level (12-24 hrs), storm surge(48-72 hours)
  in advance.
• reduce average costs of storm-related disasters
  by 10.
• Calls for systems that will facilitate quick
  assessment of the scenario from multiple
  perspectives.
• Study focus is on the rapid retrieval of the
  status of different land covers using satellite
  remote sensing data.
• Such information is normally made available after
  a lengthy process of manually identifying the
  affected areas, classifying the image data and
  then updating the GIS (maps) for the area.

4
Background

• Previous efforts in image information mining have
  focused mainly on the reduction of features using
  clustering approaches
• Little has been reported on the selection of
  best feature subsets.
• In our view, this is of more importance than
  clustering of the data features
• Feature data reduction, irrespective of
  understanding which features are optimal for the
  prediction of a particular semantic class or a
  set of classes, does not enable maximum
  utilization of the hypothesis space.
• Predictive model development should go in
  conjunction with feature selection and feature
  generation approaches.

5
Central Problem In Machine Learning

• Deciding which features to use in describing the
  concept
• Deciding how to combine those features
• Selection of relevant features and elimination of
  irrelevant ones

6
Feature Selection

• Feature extraction is an important and integral
  part of image mining systems
• feature extraction is computationally intensive
  and produces huge amounts of data that is
  difficult to manage in practice.
• Focus on the selection of a minimal subset of
  features that can give optimal performance for
  the identification of a particular land cover
  class.
• Propose a genetic algorithm (GA) based wrapper
  approach for this purpose.
• In addition to feature selection, models of
  feature generation enrich the hypothesis language
  with additional constructed and derived features
• A hybrid approach for both feature selection and
  feature generation using GAs is presented.

7
Feature Selection and Generation for IIM
Applications

• Feature selection is defined as the selection of
  a subset of features to describe a phenomenon
  from a larger set that may contain irrelevant or
  redundant features
• Feature selection techniques usually involve a
  criterion function and a search algorithm.
• Criterion fuction consists of evaluating the
  separability of classes for a given subset of
  features
• Several separability indexes have been proposed
  in the remote sensing literature
• Criterion functions based on the average pair
  wise distances without taking into consideration
  the costs associated with classes are not
  appropriate for selecting features that minimizes
  the total classification cost. (L. Bruzzone,2000)
  

8
Wrapper-based Approach

• Recently, it has been noted that the feature
  selection stage and classification stage are not
  independent because the goal is correct
  classification with a corresponding feature
  pattern extracted with the intermediate step of
  feature extraction and dimensionality reduction
• it is recommended to couple feature selection
  with effective classification techniques.
• In an wrapper-based approach, the feature subset
  selection algorithm exists as a wrapper around
  the induction algorithm.
• The feature subset selection algorithm conducts a
  search for a good subset using the induction
  algorithm itself as a part of the function
  evaluating feature subsets

9
Genetic Algorithm-based Wrapper Approach

• The Rapid Image Information Mining (RIIM) system
  adopts a Genetic algorithm-based wrapper approach
  for feature selection and generation.
• Genetic algorithms (GAs) are randomized search
  and optimization techniques guided by the
  principles of evolution and natural genetics.
• They are efficient, adaptive, and robust search
  processes, producing near optimal solutions and
  have a large amount of implicit parallelism.

10
Current Framework
11
Current Workflow
12
Rapid Image Information Mining (RIIM)

• Genetic algorithm- based wrapper approach
  for
• Feature selection
• Feature Generation
• Model creation
• Performance evaluation

13
Data Description

• Landsat ETM data corresponding to post and pre
  hurricane Katrina are used
• primitive features from 7117 segmented regions
  extracted from 60 tiles (each of 967 x 915
  dimension) Post Hurricane
• primitive features from 4592 segmented regions
  extracted from 60 tiles (each of 719 x 575
  dimensions) Pre Hurricane
• bands 4, 3, 2 corresponding to near infrared,
  red, and green were selected and the false color
  composites was derived from these bands

14
Training samples used in the study each sample
corresponds to a region in the image.
15
USGS Wetlands Classification
16
Results - Rapid Image Information Mining (RIIM)

• Primitive features extraction based on color,
  texture and shape.
• Feature selection, and feature generation using
  Genetic Algorithms (GA).
• Predictive models generation through Support
  Vector Machines (SVM)

17
Features Selected by GA
18
Accuracy, precision, recall and F-measure
obtained using only feature selection by GA
RSet of returned regions SSet of regions
relevant to the query
(proportion of relevant regions to all the
regions retrieved)
(proportion of relevant regions that are
retrieved, out of all relevant regions )
(weighted harmonic mean of precision and recall)
F-measure
19
Accuracy, precision, recall and F-measure
obtained using both feature selection and
generation by GA
20
Results of a Semantic Query (Flooded Vegetation)
21
Summary

• A fast and reliable RIIM system for tracking
  coastal disaster events has been developed
• makes use of GAs based wrapper for features
  selection and model generation, thus reducing the
  associated computational cost
• provides capabilities for a first assessment of a
  disaster situation through querying of the actual
  content of remote sensing imagery
• Can be easily modified to account for data from
  different sensors