Decision trees to classify multitemporal imagery Thales Sehn

Science challenges How are ocean, atmosphere and land processes coupled? Understand patterns of change

Changes in different time-scales (Heas, 2005)

(Goodchild, 2007) Changes in geographical objects

(Goodchild, 2007) We focus on stationary objects.

How to model changing patterns in land use/cover?

SITS – Satellite Image Time Series 06/2008 SITS 07/2008 08/2008

Detect changes t 2 t 3 What? When? Where? re fo tio n sta

SITS example 2007 2006 2008 2007 2009 2008

What attributes that best describe changing patterns? Image objects • Pixels • Cells •

Segmentation in first, or multiple times?

NDVI Variations in image attributes define temporal signatures. ≠ Temporal resolutions

Similar signatures define changing patterns. Signature for deforestation (Freitas, 2008)

Transformation Classification Algebra Visual Interpretation

Statistics Data Mining AI, M. Learning, P. Recognition Data Bases Hypothesis Classification methods based

Decision trees to classify changes Independence of number of attributes amplitude of attributes Easy

C 4. 5 Algorithm Entropy Information Value Gain Advantage of using one attribute in

C 4. 5 classification example 2 attributes pixel_mean, area 3 training classes 4 forest,

Fuzzy Decision Trees Extension of decision trees to include nonrigid limits for the thresholds

Objective Provide a technological framework to identify land use/cover changing patterns.

Extended Geo. DMA framework Timeline Visualization Mining

Slides: 26

Download presentation

Decision trees to classify multitemporal imagery Thales Sehn Korting http: //www. dpi. inpe. br/~tkorting/

The Earth is constantly changing.

Science challenges How are ocean, atmosphere and land processes coupled? Understand patterns of change in local and global scale (Kumar, 2001) Where are changes taking place? Who is being impacted by the change? (Câmara, 2008)

Changes in different time-scales (Heas, 2005)

(Goodchild, 2007) Changes in geographical objects

(Goodchild, 2007) We focus on stationary objects.

How to model changing patterns in land use/cover?

SITS – Satellite Image Time Series 06/2008 SITS 07/2008 08/2008

Detect changes t 2 t 3 What? When? Where? re fo tio n sta re fo de t 1 de n tio ro co ad ns tru c SITS

SITS example 2007 2006 2008 2007 2009 2008

What attributes that best describe changing patterns? Image objects • Pixels • Cells • Regions (Kumar, 2001)

Segmentation in first, or multiple times?

NDVI Variations in image attributes define temporal signatures. ≠ Temporal resolutions

Similar signatures define changing patterns. Signature for deforestation (Freitas, 2008)

Transformation Classification Algebra Visual Interpretation

Statistics Data Mining AI, M. Learning, P. Recognition Data Bases Hypothesis Classification methods based on data mining are efficient to identify temporal signatures.

Decision trees to classify changes Independence of number of attributes amplitude of attributes Easy to understand the result

C 4. 5 Algorithm Entropy Information Value Gain Advantage of using one attribute in despite to another info for all classes minus info per branch

C 4. 5 classification example 2 attributes pixel_mean, area 3 training classes 4 forest, 3 clear_cut, 2 road

Split in attribute pixel_mean …

Split in attribute area …

Fuzzy Decision Trees Extension of decision trees to include nonrigid limits for the thresholds

Objective Provide a technological framework to identify land use/cover changing patterns.

Classification scheme

Extended Geo. DMA framework Timeline Visualization Mining

Decision trees to classify multitemporal imagery Thales Sehn Korting http: //www. dpi. inpe. br/~tkorting/