Loading presentation...

Present Remotely

Send the link below via email or IM

Copy

Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.

DeleteCancel

Make your likes visible on Facebook?

Connect your Facebook account to Prezi and let your likes appear on your timeline.
You can change this under Settings & Account at any time.

No, thanks

Thesis

Slides for the presentation of my phd thesis.
by

Thales Sehn Korting

on 17 January 2014

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of Thesis

Input data
MOD13Q1 product
MODIS 16-day EVI2 profiles
23 images
250 m pixel size
08/02/2008 to 07/20/2009
30 neighborhoods
Micro-regions with 250m²
800 cells
Heterogeneous landscape:
Gardens
Impervious surfaces
Paved roads
Residential Areas
Shadows
Sheds
Vegetation
Water bodies
entropy(samples)
entropy(nodes)

gain(threshold)
Feature extraction
Data mining to detect land cover and change patterns
Evaluation of classification
Defining the input data
GeoDMA
Geographic Data Mining Analyst

The biomass condition along an entire phenological cycle is approximately 1 year.

(Esquerdo et al., 2009)
Example of NDVI profile
How to detect cycles?
Example of high spatial resolution imagery
Segmentation obtains objects from an image
Segmentation-based spectral features
Multi-temporal features
In many cases, hundreds of independent features need to be simultaneously considered in order to accurately model system behavior.

(Goebel, 1999)
GeoDMA: A toolbox integrating data mining with object-based and multi-temporal analysis of satellite remote sensed imagery

Thales Sehn Korting

Ph.D. thesis at the Remote Sensing program, INPE

Advised by
Dr. Leila Maria Garcia Fonseca
Dr. Gilberto Câmara

Collection rate for IKONOS is ~890 megapixels/minute
Collection rate for CBERS-2B was ~120 megapixels/minute
Cycle 2007/8
Cycle 2008/9
Cycle 2009/10
Available segmentation algorithms
(Bins et al., 1996). Satellite imagery segmentation: a region growing approach.
(Baatz and Schape, 2000). Multiresolution segmentation: an optimization approach for high quality multi-scale image segmentation.
(Korting et al., 2011). A geographical approach to Self-Organizing Maps algorithm applied to image segmentation. Lecture Notes in Computer Science, v. 6915.
(Korting et al., 2011). A resegmentation approach for detecting rectangular objects in high-resolution imagery. IEEE Geoscience and Remote Sensing Letters, v. 8.
Example: Each cell is described by the deforestation polygons inside it
Segmentation-based spatial features
Landscape-based features
The standard computational models of time do not consider that certain events or phenomena may be recurring.

(Hornsby et al., 1999)
To support cycle's visualization, Edsall et al., (1997) proposed a time-wheel legend, resembling a clock face, divided into several wedges according to the data instances.
How to visualize a remote sensing time series using this cyclical approach?
Given the shapes, various shape and linearity metrics are extracted:
area, perimeter,
main direction, bounding ellipse,
eccentricity, radius

A cycle with constant values outcomes a circle, and different cycles draw different shapes according to their properties.
Polar Features
Examples
Some possible features:
area per quadrant (season)
polar balance (differences between seasons)
Unsupervised classification using Self-Organizing Maps (SOM)
Supervised classification using Decision Trees
Building a decision tree
When the set contains only samples belonging to a single class, the decision tree is composed by a leaf.
If the set contains samples of different classes, it is refined into single class subsamples.
threshold
...
A purity measure of each node improves the feature selection
How to define the thresholds?
Suppose a set of possible events p1, p2, ... pN

The entropy aims to answer "how uncertain we are of the outcome?"
The gain aims to answer "how much entropy of the training set some test reduced?"
Accuracy assessment
Error matrix
Kappa coefficient
Monte Carlo simulation
Validation has become a standard component of any land cover map. Knowing the accuracy of the map is vital to any decision making.

(Congalton, 2005)
No result will convince, even if quantitatively proofed, if it does not satisfy the human eye.

(Baatz et al., 2000)
A strong and experienced evaluator of segmentation techniques is the human eye/brain combination.

(Gamanya et al., 2007)
Sample selection for supervised classification
Case
Studies

QuickBird image
523 x 445 pixels
11 bits
4 bands
blue
green
red
infrared
Acquisition date: March 30, 2002
(Baatz et al., 2000) segmentation:
Scale 15
Compactness 40%
Color Factor 40%
2254 objects

Samples per class:
15 training
10 validation
Accuracy of 85%
Kappa 0.842
Kappa 0.91
Northern region of Rio de Janeiro, RJ
São Paulo, southeast Brazil
Detection of urban landscape patterns
Land cover using intra-urban imagery
Analysis of deforestation patterns in the Brazilian Amazon
Multi-temporal land cover mapping
Multi-temporal detection of single and double cropping
São Félix do Xingu, southeastern Pará, Brazil
PRODES deforestation maps
2006 and 2009
Cellular grid of 10 Km²
1440 cells
Diffuse
Geometric
Consolidated
Forest
Kappa
2006: 0.87
2009: 0.81
Samples per class
10 training
10 validation
Satellite imagery provides data about land cover, which does not translate exactly into land use or land change information.

(McCauley and Goetz, 2004)
Combined with ecosystem models, remotely sensed data provides tools for predicting and understanding the planet's behavior.

(Tan et al., 2001)
Remote sensing data is the only source capable of providing a continuous and consistent set of information about the Earth’s land and oceans.

(Bradley et al., 2007)
How to retrieve rich information from remote sensing images?
TerraClass derived typology (year 2008)
Croplands
Pasture
Dirty pasture
Clean pasture
Urban area
Deforestation
Forest
Forest
Secondary vegetation
Pasture with regeneration
Mato Grosso, Brazil
Training and Classification using Basic multi-temporal features
Training and Classification using Basic + Polar multi-temporal features
x
Samples
100
Input data
MOD13Q1 product
MODIS 16-day EVI2 profiles
23 images per year
250 m pixel size
08/02/2008 to 07/26/2011
Using Croplands class (TerraClass)
Single croppings
Double croppings
Others
Mato Grosso, Brazil
Samples
Monte Carlo simulation evaluated the Polar features, varying the parameter min
obj
in [1, 50], 100 simulations per min
obj
value
Training and Classification using Basic multi-temporal features
Training and Classification using Basic + Polar multi-temporal features
x
The polar visualization of croppings
single x double
2008
2009
2010
Monte Carlo simulation evaluated the Polar features, varying the parameter min
obj
in [1, 130], 100 simulations per min
obj
value
Measuring land change
(S)ingle, (D)ouble
Constant (SSS, DDD)
Intensification (SSD, SDD)
Reduction (DDS, DSS)
Interchange (DSD, SDS)
(Reis, 2011)
(Saito, 2011)
In this thesis, two objectives were pursued
The versatility of the approach was tested in 5 different applications
Frohn e Hao (2006)
Metzger et al. (2009)
Novack et al. (2011)
Pinho et al. (2012)
Silva et al. (2008)
Southworth et al. (2002)
Ribeiro et al. (2009)
To interpret remote sensing images, end users still lack effective and operational tools
Data integration becomes a great challenge
conversion of data format
knowledge of the software to be used
files replication
image segmentation
feature extraction
feature selection
multi-temporal analysis
data mining
GeoDMA
Geographic Data Mining Analyst
blue roofs
bright roofs
ceramic roofs
dark roofs
gray asbestos roofs
grass
swimming pools
shadows
trees
Functionalities:
Spatio-temporal analysis and visualization
Simulation methods for accuracy assessment
Different geographic data types
Detection of multi-temporal changes
GIS integration (TerraView)
The construction of a scheme for multi-temporal analysis
The development of a free and open source toolbox for satellite image interpretation
When comparing the Polar features to Basic, the classification accuracy improved about 5.12% and 36.84% when coupling both feature sets.
Further work on multi-temporal analysis
automatic change detection
cycle detection (seasonal or biannual)
evaluation of trends for land change
Toolbox improvements
Multi-resolution analysis
Definition of new features
Storage and retrieval of large images
Documentation available
toolbox
source-code
GeoDMA 0.2.2 with TerraView 4.2.0
QT, QWT
TerraLib (C++)

Largest data tested
50000 Polygons
Raster with 10000 x 10000 pixels, 4 bands
Multi-temporal Rasters
1500x1500 pixels, 75 bands (3 years)
5000x5000 pixels, 23 bands (1 year)
image segmentation
feature extraction
feature selection
multi-temporal analysis
data mining
(Novack, 2009)
Kappa = 0.81

Average increment of 5.12% in accuracy when compared to Basic features
Kappa = 0.8

Average increment of 36.84% in accuracy when compared to Basic features
http://geodma.sf.net/
References
SILVA, M. P. S. Mineração de padrões de mudanças em imagens de sensoriamento remoto. 2006.
GAVLAK, A. A. Padrões de mudança de cobertura da terra e dinâmica populacional no Distrito Florestal Sustentável da BR-163: população, espaço e ambiente. 2011.
REIS, I. C. Caracterização de paisagens urbanas heterogêneas de interesse para a vigilância e controle da dengue com o uso de sensoriamento remoto e mineração de padrões espaciais: um estudo para o Rio de Janeiro. 2011.
SAITO, É. A. Caracterização de trajetórias de padrões de ocupação humana na Amazônia Legal por meio de mineração de dados. 2011.
Object-based image analysis
Full transcript