In this
post, we will be talking about finding circular irrigation fields. The image
was download from here. If you want to try, download it and play around. The approach
I am using is detection of edges and finding edges that form circular edges.
If you have heard of Hough’s transformation for detection of straight lines,
the method is just an extension of it to detect circles. For Hough’s transformation,
please go through here. If you have this excellent book, it has an elaborate
explanation of Hough’s transformation to detect straight lines. Peoples have
used Hough’s transformation for many different purpose. Many people use it for detecting straight building edges that can used to reconstruct buildings for 3-D
buildings, generating city GML models etc.

For Hough’s
transformation for straight line detection, you can use either PYTHON based
Scikit-image or MATLAB. Circular Hough's transformation is also available in both scikit-image and openCV. The algorithm (

A circle is represented mathematically as:

where xcenter and ycenter are center of the circle and r is the radius of the circle. As you can see, there are 3 parameters to be fitted for cicular hough transformation.

Read the documentation of MATLAB, to get an idea about

*imfindcircles)*is available in MATLAB since 2013b version with image processing toolbox. I could not find the algorithm in any remote sensing software so far. My personnel view is that remote sensing software are very behind on incorporating state-of-the-art algorithms. So, knowing some coding either PYTHON, MATLAB or R will take you to greater heights in your professional path.A circle is represented mathematically as:

where xcenter and ycenter are center of the circle and r is the radius of the circle. As you can see, there are 3 parameters to be fitted for cicular hough transformation.

Read the documentation of MATLAB, to get an idea about

*imfindcircles*function parameters. So here basically, we are going to use*imfindcircles*function to detect center pivot irrigation fields in the image. In this image, there are only dark pivot circular fields surround by bright objects. So, we will be using only ‘dark’ mode of*imfindcircles.*The minimum circle and the maximum circle radii are image dependent so you need to provide those information with a little bit of data exploration.
Here is a code in MATLAB.

Original RGB image |

RGB image with detected pivot irrigation fields |

As you can see, 7 fields out of 9 fields were correctly detected. Two undetected fields( middle -top of the image) are also darker circles but due to its low contrast with surrounding, they were not detected. Even with the higher value for parameter '

*sensitivity*' and the low value of '*edgeThreshold*' paramter, those two fields were undetected. You can further increse sensitivity and lower edgeThreshold parameters, to find those undetected circles but then you risk of finding many false alarms as well.
I have a gut feeling that with OBIA with ecognition, the process of finding center pivot irrigation fields would be not a straight forward procedure as with using the function

*imfindcircles*and the process would be much complex. Nevertheless, i will try it with eCognition in near future and report back.
Pixel-based based classification using any machine learning classifies will fail miserably for this case as center pivot irrigation fields ares spectrally similar to vegeation in other rectangular plots.

__UPDATE__

I spent some time in eCognition exploring a newer algorithm " template matching". The technique is not new but it has been incorporated with the last release of eCognition. The concept is given a template, the template moves over the image (a single layer) in a sliding window and calculates normalized cross-correlation simalarity between the template and the pixels within the sliding window. The result is an cross-correlation image. Subsequently a threshold value is used to find position of pixels with higher cross-correlation value.

The concept of normalized cross-correlation is shown below taken from a good presentation. Study the presentation in detail if you want.Notice border effect of the cross-correlation image below. This can be avoided if padding with replicated pixels are added in the image (commonly done in MATLAB) for any convolution procedure.

The concept of normalized cross-correlation is shown below taken from a good presentation. Study the presentation in detail if you want.Notice border effect of the cross-correlation image below. This can be avoided if padding with replicated pixels are added in the image (commonly done in MATLAB) for any convolution procedure.

The detail explanation for performing the template matching in eCognition will follow sometime in future.

Cross-correlation concept |

A genarated template with many samples. |

Cross-Correlation image |

Result of template matching in eCognition |

Shailesh, this is great work on identifying crop circles!

ReplyDeleteHI Li, thanks for dropping few words.

ReplyDeleteHello,

ReplyDeleteI found this very useful. I need to detect irrigation circles, with an automated process. I'm proficient with Python and familiar with Machine Learning (sklearn, etc). I understand your comment (and agree) about pixel-based classification not being a good idea. How would you recommend that I start?

if you are familiar with sklearn, you should of course use circular hough transformation as shown in first approach.

Delete