Change detection. Which algorithm is implemented?


#1

Hello,

I found a tool for change detection in S-1 toolbox. I wanted to read more about it - especially about an algorith that is used - but for that tool there is no information in help available from the software.
Would it be possible to get info about some references, e.g. articles, books, websites?

Thank you in advance,
Helena


SAR Time-series Analysis tool"
#2

I think its just a ratio between the second and the first image.

So at pixels where the intentensity increased you get high values and where the intensity decreased you get low values.

Please see here:
http://www.isprs.org/proceedings/XXXVII/congress/7_pdf/10_ThS-18/12.pdf
2.2 image ratio method


2.2 Image rationing


#3

Thanks a lot for information!


#4

It’s based on the log likelihood from this paper.
www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA458288


Sentinel 1 GRD change detection workflow
Change detection and ratio
Change detection and ratio
#5

Thank you for a correcting!


#6

After reading the article I think I need to use polarimetric covariance matrix as an input data. Am I right? That requires Sentinel-1 SLC data. Which type of algorith should I use to compute change detection based on Sentinel-1 GRD data?

Thank you for any comments.


#7

Just to be sure, the “log_ratio” image that is an output of “Change Detection” algorithm implemented in Sentinel-1 toolbox is computed according to eq. 84 - 85 on page 41 (Preiss, M., & Stacy, N. J. (2006). Coherent change detection: Theoretical description and experimental results (No. DSTO-TR-1851). Defence Science And Technology Organisation Edinburgh (Australia)). Is it correct?


#8

Hi hlos

I am quoting the following paragraph from Preiss and Stacy 2006:
The mean backscatter power ratio is sensitive to changes in the average backscattered energy in the transduced imagery. The sample coherence on the other hand is sensitive to changes in the speckle noise pattern in the repeat pass image pair. Scene disturbances arising from man-made changes however, can potentially cause changes over a broad range of scattering properties. In [11] the sample coherence and mean backscatter power ratio were used to detect changes in repeat pass ERS-1 SAR imagery. It was found that the areas of disturbance identified by each method did not necessarily agree and each method gives complementary characterisations of scene changes. Therefore both change statistics should be considered to provide a complete description of scene changes. In the context of change detection this presents problems in fusing the detections from the two change statistics to achieve a single combined detection list in which the probability of detection is maximised whilst minimising the false alarm rate.
An alternative approach to discriminating between those regions affected by man-made scene changes and those that are not can be achieved by formulating the detection problem in an hypothesis tesing framework. In this approach the change detection problem is to determine whether pixel pairs Xk = [fk,gk]T,k = 1···N in a local area are realisations of a null (unchanged scene) hypothesis H0 or an alternative (changed scene) hypothesis H1…
Given a local neighbourhood of N independent pixels Xk = [fk,gk]T,k = 1···N a simple decision statistic for determining whether the pixels are realisations of the unchanged hypothesis or changed hypothesis is the likelihood ratio defined as,

The end is based in two hypothesis…:blush:

I was processing a data set from ALOS PALSAR 1 with three methods for change detection:

1.- Composition multitemporal in RGB,
2.- Log Ratio,
3.- Segmentation


#9

Hi Gabriela,

Thanks for the information.
I got an image of changes and the image histogram, which looks like this:

Does X axis (horizontal axis) of histogram show “z” value from eq. 113?


#10

According to the following quoting from Preiss and Stacey 2006: Discrimination between the two hypotheses is achieved by evaluating z over the local N pixel neighbourhood and applying a threshold T. For z > T the pixels are labelled as being realisations of the H1 hypothesis and for z < T the pixels labelled as being realisations of the H0 hypothesis.

I think that is right, z values are in X axis.

Please someone have an opinion different???


#11

Example for Log Ratio of a area in Lara state in Venezuela (2009-2010), the probability index was 0.64