General principle

General Principle

The method proposed by BESSETTES and DESACHY intends to extract and classify urban areas from remote sensing images by means of combination of an edge detector and a classification algorithm. This study improved the urban area extraction algorithm by defining a fuzzy method of urban/non-urban classification.

The algorithm is divided into two main steps:

First of all, built up areas are extracted from the image. Pixel by pixel classification is combined with an image segmentation by edge detection to automatically extract the pixels that belong to a building. Then these pixels are grouped to define the urban areas with the use of an iterative segmentation process based on edge detection.
Then, urban density is computed for each built up area previously obtained. The urban density is defined as the percentage of pixels that belong to the street class into a urban district (each region fixed by wide thoroughfares).

On SPOT images, the buildings appear as small rectangles from 1 to 16 pixels. The idea lies in the detection of the angles of buildings, then in the gathering of zones with strong density of angles, to recreate the urban zones.

A SOBEL gradient operator is used horizontally and vertically on the image. The two resulting images make it possible to calculate the angles formed by the gradient vectors.

For a given pixel, the angles formed with its eight neighbours are calculated. Starting from the fact that a building angle is a right angle, a membership function expresses how much a angle looks like a right angle and therefore how much it is likely to be the angle of a building. The expression is given by:

$\begin{displaymath}\mu_{ij}= \sqrt{g_i* g_j}* (\frac{\theta_{ij}}{\pi/ 2})^\alpha \end{displaymath}$

with:

g_i is the gradient norm of vector i, given by the two components of SOBEL;
g_j is the gradient norm of vector j;
$\theta_{ij}$ is the angle formed by the gradient vectors i and j;
$\alpha$ is a positive coefficient used to weight the angle participation. If the value of $\alpha$ is close to zero, the value of the angle will have less importance in the calculation of $\mu_{ij}$ , whereas if the value of $\alpha$ is very high, $\mu_{ij}$ will only have a high value when a_ij is close to $\pi/2$ .

The new value of the pixel will be the average of the values obtained for its eight neighbours. The image thus obtained is then filtered by a median filter.

The original image is then segmented into regions by means of the ISEF filter, followed by a hysteresis thresholding, then by an edge closing. The small regions are finally merged according to their average radiometry.

Each region defines a sample in the image of membership to angles. The samples make it possible to calculate by dichotomy a threshold to delimit the classes ``urban'' and ''non-urban'', as well as possible. The value of this threshold makes it possible to obtain areas of which the number of urban pixels is higher than $60\%$ , and others of which the number of non-urban pixels is higher than $60\%$ . The final value of the threshold is calculated starting from the histograms of the most urban area and the least urban.

To obtain a less noisy binary image, the regions which are not sufficiently homogeneous are burst according to two classes. Then they are merged again: urban with urban, non-urban with non-urban.

The final result gives a binary mask of the urban areas.

IRIT-UPS