4_Paragominas.ppt

ROAD EXTRACTION IN URBAN AND RURAL ENVIRONMENTS

EXPLOITING A DUAL-BAND SAR SYSTEM

P. Gamba(1), G. Lisini(2), D. Luebeck(3)

(1) Università degli Studi di Pavia(2) IUSS, Pavia(3) Orbisat, Sao Jose dos Campos

• What’s the problem?• The Orbisat system• The proposed algorithm• Experimental results• Conclusions

Outline

Common problems of road extraction

Road networks automatically extracted from remotely sensed data are often incomplete

Common disadvantages are:

• due to the backscattering effects on the edge of the road• buildings and trees often shadow some parts of the roads• shape of the road is often biased by buildings

In highly vegetated areas:

• shape of the road is often incomplete

Does radar frequency matter?

Green arrows point to roads in P-band data that are not visible in X-band data.

P-band X-band

Dual-band SAR may help ...

Dual-band SAR systems are well suited for vegetation analysis

X-band is more prone to be scattered by trees and foliage

P-band, due to its longer wavelength, is able to pass through vegetation and to better detect underlying roads and other man-made or natural objects

• OrbiSAR airborne-RFP, designed and built by Orbisat da Amazônia Company, consists of a SAR sensor in the X (9.65 GHz) and P (0.415 GHz) bands, installed on board an aircraft TURBO COMMANDER, a navigation system with measurement equipment of the position with absolute accuracy of 1 m in real-time, and an Inertial Measurement Unit with angular accuracy of one hundredth of degree.

Orbisat system

The general idea

… A fusion methodologies able to exploit road extraction from an airborne dual-band SAR

acquisition …

Multi-frequency SAR data

Nth band road extraction

1st band road extraction

Road network fusion

2nd band road extraction

…

Dual-band SAR data analysis

SAR road extraction methodology

Roaddetection

Roadextraction

Road networkMRF optimization

Networkregularization

HRSAR

Final roadnetwork

Multi-scale feature fusion Junction-aware MRF model

Perceptual grouping

M. Negri, P. Gamba, G. Lisini, F. Tupin, “Junction-Aware Extraction and Regularization of Urban Road Networks in High Resolution SAR Images”, IEEE Trans. on Geoscience and Remote Sensing, vol. 44, n. 10, pp. 2962-2971, Oct. 2006.

Road detection

Roaddetection

Roadextraction

HRSAR

Multi-scale feature fusion

• In high resolution SAR images, roads are no more a subset of image edges.

• Instead, they usually appear as dark, elongated areas, with bright lateral edges.

• Parallel edges, however, identify other artificial structures also (buildings) and low reflectance areas have very similar spectral response.

Multiple feature extraction

• Low reflectance is provided by minimizing the mean value along any given direction:

• … and retaining the direction of interest

• Contrast is computed, too.

Remote Sensing Group

Road candidate area extraction

“min radiance” thresholded

“min radiance” output

Binarization?

• Binarization is achieved by local thresholding

Road candidate extraction

• Aim: from pixels to segments

Roaddetection

Roadextraction

HRSAR

Multi-scale feature fusion

Tracking process

Road candidate extraction

Final segments

The approach is efficient, with the only drawback of reducing curvilinear roads to chains of

linear segments.

Perceptual grouping step

The procedure is based on Perceptual Grouping Concepts and allows connecting segments where reasonable, based on their mutual positions

1°

2°

3°

4°

5°

6°

Network optimization Some extracted segments are not connected along the entire path

34

5

2

1

7

6segment extreme

found segment

added segment

Markovian approachMarkovian approach

Energy function defined by:

probability density function of amplitude SAR image

prior knowledge about the road shape

Roads are long (they should almost never stop) roads have a low curvature intersection are rare (at least in no urban areas) crossroads with either “cross” or “T” shapes are frequent crossroads with more than four segments are rare

Remote Sensing Group

Experimental Results

The area around the town of Paragominas (state of Parà, Brazil).

•OrbiSAR RFP sensor - single look ground range image• High spatial resolution (2.5 m)• The scene covers an area of nearly 5000 m x 6750 m • Both P-band and X-band image available.

First test areaSecond test area

Remote Sensing Group

First test site results

Original P-band image P-band extraction

Original X-band image X-band extraction

Remote Sensing Group

First test site results

final results after the fusion step (option 1b)

final results after the fusion step (option 1a)

final results after the fusion step (option 1c)

Quantitative evaluation

X-band extraction results

P-band extraction results

Fusion results

(1a)

Fusion results

(1b)

Fusion results

(1c)

Completeness 0.16 0.67 0.73 0.73 0.64

Correctness 0.78 0.79 0.74 0.78 0.93

Quality 0.15 0.56 0.60 0.61 0.54

Redundancy -0.07 -0.03 0.06 0.06 0.03

Main outcomes:+ due to the presence of vegetation, P-band results are better

than X-band ones;+ the fusion of both extractions increase in the completeness and correctness index values;+ the real improvement in the road extraction results is for the roads outside the human settlement.

Second test site results: Paragominas

P-band image X-band image

Best result for Paragominas

Final results after the fusion step

(option 1b)

P-band extraction results

Fusion (option 1a)

Fusion (option 1b)

Fusion (option 1c)

Completeness 0.64 0.92 0.92 0.84

Correctness 0.55 0.41 0.45 0.95

Quality 0.42 0.41 0.43 0.63

Redundancy -0.02 0.05 0.05 0.03

Quantitative indexes for the road extraction results in the Paragominas urban

area test site

Conclusions

• Exploiting dual band system, there is a chance to obtain in an automatic way nearly 90% of the overall road network, with a 50% redundancy of the extraction.

• The experimental results validates in two different areas the choice for a procedure that fuses information extracted from the data sets at different wavelengths at an intermediate step of the whole road network extraction chain.

• Next steps for this work will be related to the definition of accurate pruning approaches and refinement steps able to overcome the current limitations of the algorithm.