How to extract road networks from highly and semi-urbanized areas without detecting the edges of buildings and other nonroad features?

How to extract road networks from highly and semi-urbanized areas without detecting the edges of buildings and other nonroad features? I can't seem to group and "filter out" the buildings such as in the photo below:

Basically, the goal is to pre-process the image first before I use edge detection in order to extract the road network. But the pre-processing technique that I use always yield unacceptable results because I can't "filter out" the cluttered buildings that closely resemble the road surface. Any matlab code to help me group and "erase" these cluttered buildings in the image? Many thanks.

Answer 1

[Image Analyst]

There is nothing in MATLAB to do that at a high level. You'll have to build it up from lower level functions. Go here to see algorithms: http://www.visionbib.com/bibliography/contentscartog.html#Cartography,%20Aerial%20Images,%20Remote%20Sensing,%20Buildings,%20Roads,%20Terrain,%20ATR Sorry, I have not read any of those papers nor do I have code for them.

[Jannie Sy]

I have read a preprocessing techique that may be executed in matlab. wherein first, we obtain a road network of a specific width. Extract elongated road regions using its linear structure by using a linear operator (predefined/user-defined length) to filter the road segments that are not part of the global road network. Then, a median filtering technique is applied to the resultant image. In this median filter, each output pixel is set to an average of the pixel values in the neighborhood of the corresponding input pixel. The median of the neighborhood pixels determines the value of an output pixel. But I can't seem to put it into the matlab code correctly. :(

[Image Analyst]

Your definition of median filter is not correct - it's not the average, it's the median. In MATLAB it's done by medfilt2(). You might also be interested in hough() or houghlines().

I don't know what the first two sentences you describe do. What is the road network or global road network, and how does that differ from the actual roads present in your image? And how do you get that first (like you said it needed)???

[Jannie Sy]

I'm sorry I wasn't clear about it and didn't define it prior to my question. But here's the thing actually.

First, my input is a high resolution satellite image. I set a specific threshold value for which the image will be grouped into clusters in an unsupervised classification. A level one classification is carried out by dividing the image into two classes: Roads and non-roads.

Then, I'm supposed to apply a nearest neighborhood grouping (NNG) operation to the previously classified image. This should smoothen the spectral response within the pixel’s local neighborhood. Then choose a pixel from which its surrounding eight neighbors will be subject to voting. The chosen pixel will be assigned to a specific class if that class gets four or more votes. Otherwise, the pixel will maintain its class.

The last step would be filtering, and this was what I described previously.

I'm really sorry for the trouble but I hope you can help me. Thanks. :(

[Adedoyin Atanlode]

Jannie what unsupervised classification method did you use? My work is similar but input is panchromatic HR satellite image but I've been having similar unacceptable result due to absence of preprocessing step. If through with your work, how did you get over the problem pls? Many thanks