Tipo de plataformas y sensores para el mapeo de … · Tipo de plataformas y sensores para el mapeo de áreas de sequía y desertificación: impacto de las plataformas y sensores

!"#$"%&#&

#

Tipo de plataformas y sensores para el mapeo de áreas de sequía y desertificación: impacto de las plataformas y sensores en la calidad del mapeo

II Escuela de primavera sobre soluciones espaciales para el manejo de desastres naturales y respuestas de

emergencias: sequia y desertificación

Paula D. [email protected]

UN-OOSA UN-SPIDER

Part I

Remote sensing systems: platforms and sensors

!Elements of remote sensing systems

!Remote sensing data collection

!Characteristics of some optical and radar sensors

!Sensors used vs. surface areas covered by land degradation studies

!"#$"%&#&

%

RECEIVING STATION

Earth surface

Spectral range:Visible, near and middle infrared

Spectral range:Thermal infrared

Spectral range:Microwaves

SOURCE(sun)

OBSERVATION SYSTEM(SAC-D)

OBSERVATION SYSTEM(SAOCOM)

Active sensors(radar)

Pasive sensors(radiometers)

Emitted radiation

AtmosphereEmitted

Electromagnetic radiationReflected radiation

Emittedradiation Backscattered

radiation

TargetsTargets

Platforms

Elements of remote sensing systems

Remote Sensing Data Collection

The amount of electromagnetic radiance, L (watts m-2 sr-1) recorded within the IFOV of an remote sensing system is a function of:

( )Ω= ,,,,, ,, PtsfL zyx θλRadiometric resolutionSpectral resolution

Spatial resolution

Temporal resolutionAngles among radiation source-target-sensor

Polarization

TÉRMICO

100 1 110 101000,1 0,1 1λ

Micrómetros Centímetros

MICRO-ONDAS

RADAR

INFRARROJO

NIR

MED

IOUV

0,4 0,5 0,6 0,7

SWIR

1,3 2,5

VIS

!"#$%&"'()

*"#+,%$"'()-"#,"'()

!"#$%"& '()*+,-)./)01

2345678#)*9:9,-)./)01

Ejemplos:

;!<=)>;!=;?)*!)-)@/)01

?"$"6%"&A@)*B-)C)01

!"#$%&'

;!<=)>?D=()*@,-)9EF)01'566"=;?AG)*G)-)@)01

()*)+

!"#$"%&#&

'

Satellite ERS-1 ERS-2 Radarsat-1 JERS-1 EnvisatALOS-

PALSARTerraSAR-

XCosmo/

SkyMED1SAOCOM

Band(wavelenght cm)

C (5.7) C (5.7) C (5.7) L (23.5) C (5.7) L (23.5) X (3) X (3) L (23.5)

Polarization VV VV HH HH HH/VV All All HH/VV All

GroundSwath(km)

100 100 45-500 75 56-400 40-350 5-100 10-200 30-350

Resolutionrange-

azimuth(m)26-28 26-28 9-100 18 30-150 7-100 1-16 1-100 10-100

Characteristics of some optical sensors

Satellite Landsat 4, 5Landsat 7

(ETM)TERRA (ASTER)

SAC-C ALOS PRISM ALOS AVNIR2

HYPERION IKONOS-2

Temporal resolution

16 16 16 16 46 46 - 1-3

Spectralresolution

Multispectral(7)

Panchromatic (1)Multispectral

(7)

Multispectral(14)

Multispectral(5)

Panchromatic(1)

Multispectral(4)

Hyperspectral(220)

Panchromatic(1)

Multispectral(4)

GroundSwath

(km)185 185 60 360 70 70 7.6 11.3

Spatialresolution (m)

30-12015

30-6015-90 175 2,5 10 30 1-4

Characteristics of some radar sensors

Metternicht, G.I., J.A. Zinck, P.D. Blanco y H.F. del Valle. 2009. Remote Sensing of land degradation: experiences from Latin America and the Caribbean. Journal of Environmental Quality 39: 42-61.

Sensors used vs. surface areas covered by land degradation studies

!"#$"%&#&

!

Blanco, P.D., G.I. Metternicht y H.F. del Valle. 2009. Improving the discrimination of vegetation and landformspatterns in sandy rangelands: a synergistic approach. International Journal of Remote Sensing 30:2579-2605.

Assessment of Terra-ASTER and Radarsat imagery for discrimination of aeolian degradation features in Northeastern Patagonia:

an object oriented approach

Part II Application example

!Why Object Oriented Image Analysis (OOIA)?

!Objectives

!Study area

!Research Approach

• Data sets

• OOIA conceptual model

• Classification rules

• Map Outputs

• Efficiency of synergistic approach

• Multi-resolution segmentation

• Fuzzy classification

Outline

!"#$"%&#&

(

Why Object Oriented Image Analysis (OOIA)?

SCALE

• Average heterogeneity of image objects weighted bytheir size should be minimized.

HETEROGENEITY CRITERION

Compactness

Smoothness

Shape

Color

HC

• intrinsic features

• topological features

• context features

Object features

Multi-resolution segmentation

!"#$"%&#&

)

Soft classifiers, e.g. fuzzy systemsUncertainties

Vagueness

Fuzzy rulesOne-dimensional membership function

Nearest-neighbour multidimensional feature space

If (layer mean(object) LLM) AND (length/width(object) HL/W),

then land cover(object) = river

2) Fuzzy rule base

If (layer mean(object) LLM), then land cover(object) = water

3) Defuzzification0.3

0.60.8

µ0.8

µ

µ=1.0

µc(x)

µF(x)

µ=0.5

µ=0.0

0.20.4

0.8µ

70 200 255

1) Fuzzification

Fuzzy logic:

Fuzzy classification

The current study focused on the cartography of aeolian degradation features overa sandy rangeland by combining multiresolution image segmentation and objectoriented image classifications of VNIR and microwave satellite data.

Objectives:

(1) to assess whether information content can be increased by specific SARenhancements, and select optimal textural measures for the discrimination ofaeolian features,

(1) to investigate and implement object-oriented image analysis algorithms andfuzzy logic techniques for the recognition and classification of aeolian featuresfrom radar-derived textural measures and VNIR data, and

(1) to assess the effect that the synergy of textural and optical data exerts on theclassification accuracy of aeolian degradation features of sandy rangelands.

!"#$"%&#&

*

!"#$%&'%!

VB./0123456"785997:7;36""""""""""""""""""""""""

<7/2476

=954856";5"59362>4"

?74836"5>/2'36

?74836"5>/2'36";5/@7;36

A0476

B957";5"5680;23

!"#$%

!"#$#%

!"#$#%&'()

*('+,%

!

!39;3456"/28397/56&'()( &'*)*

&(+)"

&(*)"

• Active dune• Reg• Grassland,• Grass Stabilized Lineal Dune• Scrubland• Scrub Stabilized Lineal Dune

Two vegetation types are considered as dune stabilizers: scrub (A) and grass (B)

Interest Classes

!Study area and interest classes

1. ASTER and Radarsat images pre-processing.

2. Creation of a geo-spatial soil database to store field observations andspectral characteristics of soil degradation features in the optical andmicrowave regions of the spectrum.

3. Segmentation and object-oriented classification using eCognition software.

4. Assessment of the synergistic approach using error matrices and Kappastatistics.

!Research Approach

!"#$"%&#&

+

• PCA => VNIR + Band 4

• Soil Adjusted Vegetation Index (SAVI)

VNIR (bands 1-3, SR=15 m)SWIR (bands 4-9, SR=30 m)TIR (bands 10-14, SR=90 m)

Pre-processing

Terra-ASTER

C-band, HH pol., SR=8 mRadarsat-1 ASAR

• Despeckle (Frost filter)

• Selection of textural measures derived from theGrey Level Co-ocurrence Matrix (mean,dissimilarity, contrast and variance)

Pre-processing

• Data sets

• OOIA conceptual model

!"#$"%&#&

$

AD=Active Dune, GSD=Grass Stabilized Lineal DuneSSD=Scrub Stabilized Lineal Dune

BR=brightness, L to W = ratio Length/Width RB = relation of border to a certain class

• Classification rules

Object-oriented classifications based on optical data from ASTER (a), textural

measures derived from Radarsat (b), and merged optical and texture image data (c).

!"#$"%&#&

#&

Class AD REG GRASSLAND GSD SCRUBLAND SSD

Table 6(a). ASTER data input.

AD 666 (94.87)(91.36)

36 (5.13) (8.70) 0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

REG 63 (13.46)(8.64)

360 (76.92)(86.96)

0 (0)(0)

27 (5.77) (3.23) 0 (0)(0)

18 (3.85) (2.27)

GRASSLAND 0 (0)(0)

0 (0)(0)

324 (67.92)(87.80)

153 (32.08)(18.28)

0 (0)(0)

0 (0)(0)

GSD 0 (0)(0)

0 (0)(0)

45 (6.41)(12.20)

657 (93.59)(78.49)

0 (0)(0)

0 (0)(0)

SCRUBLAND 0 (0)(0)

18 (3.17) (4.35) 0 (0)(0)

0 (0)(0)

333 (58.73)(92.50)

216 (38.10)(27.27)

SSD 0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

27 (4.62) (7.50) 558 (95.38)(70.45)

Table 6(b). Radarsat-derived texture data input.

AD 630 (97.22)(86.42)

18 (2.78) (4.35) 0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

REG 90 (18.87)(12.35)

387 (81.13)(93.48)

0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

GRASSLAND 9 (1.61) (1.23) 0 (0)(0)

324 (58.06)(87.80)

225 (40.32)(26.88)

0 (0)(0)

0 (0)(0)

GSD 0 (0)(0)

0 (0)(0)

45 (6.85)(12.20)

612(93.15)(73.12)

0 (0)(0)

0 (0)(0)

SCRUBLAND 0 (0)(0)

9 (10.00)(10.87)

0 (0)(0)

0 (0)(0)

324 (72.00)(90.00)

117 (10.00)(5.68)

SSD 0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

36 (5.06)(10.00)

675 (94.94)(85.23)

Table 6(c). ASTER combined textural data input.

AD 684 (95.00)(93.83)

36 (5.00) (8.70) 0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

REG 45 (7.84) (4.94) 378 (82.35)(91.30)

0 (0)(0)

18 (7.84) (4.30) 0 (0)(0)

18 (7.84) (4.55)

GRASSLAND 0 (0)(0)

0 (0)(0)

306 (77.27)(82.93)

90 (22.73)(10.75)

0 (0)(0)

0 (0)(0)

GSD 0 (0)(0)

0 (0)(0)

63 (7.95)(17.07)

729(92.05)(87.10)

0 (0)(0)

0 (0)(0)

SCRUBLAND 0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

342 (82.61)(95.00)

72 (17.39)(9.09)

SSD 0 (0)(0)

0 (0)(0)

0 (0)(0)

0 (0)(0)

18 (2.50) (5.00) 702 (97.50)(88.64)

• Contingency matrix for the accuracy assessment of the object-oriented classifications.

The diagonals are pixels correctly classified with user accuracy (in brackets bold) and producer accuracy (in

brackets bold italics) for each class. User accuracy shows the error of commision and producer accuracy shows the error of omission. Non-diagonals represent errors with commission percentage (in brackets) and ommissionpercentage (in brackets and

italics).

• Efficiency of synergistic approach

Spatial assessment of similarity using the two-way fuzzy map comparison: (a) ASTER data input vs. ASTER combinedtexture information classifications, and (b) textural measures vs. ASTER combined texture data classifications. Areasmapped identically have values close to or equal to 0, while areas of total disagreement show values close to orequal to 1.

Class AD REG GRASSLAND GSD SCRUBLAND SSDAD 1 0.5 0 0 0 0REG 0.5 1 0 0 0 0GRASSLAND 0 0 1 0.6 0.4 0.4GSD 0 0 0.6 1 0.4 0.4SCRUBLAND 0 0 0.4 0.4 1 0.6SSD 0 0 0.4 0.4 0.6 1

Fuzzy Similarity Matrix.

AD=Active Dune, GSD=Grass Stabilized Lineal Dune, SSD=Scrub Stabilized Lineal Dune.

!"#$"%&#&

##

¡Muchas gracias!

¿Preguntas?

Corresponding author: [email protected]

• From an ecological point of view, it is more appropriate to analyzeobjects as opposed to pixels because landscapes consist ofpatches that can be detected on the imagery like objects.

• Multi-resolution segmentation allows integrating data with differentspatial resolution and different radiometric characteristics (e.g.optical + radar).

• Objects are connected by a hierarchical network that allows theefficient propagation of many different kinds of relationalinformation.

• Some target classes are only distinct through topology and shapefeatures which are not available in traditional pixel basedclassification approaches.

• Complex “knowledge base” about classes can be directlyformulated in classification fuzzy rule sets.