Team members:

Aurelie Voisin
FRANCE

Attila Borcs
HUNGARY

Szilard Szalai
HUNGARY

Krzysztof Filipczak
POLAND

Introduction

The growth of the size of cities, often at rates exceeding the population growth rate, and the accompanying loss of agricultural lands, forests and wetlands, escalating infrastructure costs, increases in traffic congestion, and degraded environments, is of growing concern to citizens and public agencies responsible for planning and managing growth and development.

Remote sensing in the form of aerial photography has been an important source of land cover and land use information. However, the cost of aerial photography acquisition and interpretation of cover types is prohibitively expensive for large geographic areas. An alternative is to acquire the needed information from digital satellite imagery such as Landsat etc.

This approach has several advantages: (1) the synoptic view of the sensor provides coverage of large geographic areas, (2) the digital form of the data lends itself to more efficient analysis and the classified data are compatible with geographic information systems, eliminating the need to digitize interpreted information, and (3) land cover maps can be generated at considerably less cost than by other methods.

The task

Input: Landsat images of terrain Sample images of fields, sea, forest etc.

Aim: Segmentation of scene based on texture and colour including registration and matching of mutiple scenes

Additional goal: Identification of key features (cave openings, tanks)

Output: Labeled or segmented scene

State-of-the-art

For more effective use of the satellite remote sensing, land-use managers should be aware of the limitations and advantages of satellite data and should choose from their available land-use mapping options accordingly. Remote sensing is especially proper for initial reconnaissance mapping and continued monitoring of land-use over large areas. Selection of the most proper satellite image, band combination, and appropriate classifier are very important. Additionaly, the image processing is relevant and different stages of it such as band filtering and principal component analysis should be applied before evaluation.

In the frame of satellite image classification, the maximum likelihood method is found to be more applicable and reliable. While the minimum distance method has given more reliable results than the linear discriminant procedures, the parellelpiped method is found to give less reliable results when compared to the other methods.

Our solutions

We developed two different solutions to solve the previously mentioned problem.

The first method is divided into two steps, a training step and a test step, which we are going to detail in this section. The considered satellite images are optical Google Earth images (IKONOS satellite). They are described according to a RGB (Red-Green-Blue) color model. The training step aims to determine parameters related to each training sample. First, we have to create a training database with different kinds of classes such as mountains, lands, forests… Given the complexity of urban areas, we will not consider this specific class. For each training set, we apply a 2-D Gaussian convolution operator. The size of the convolution window is chosen by trial-and-error. Typically the size varies between 5 and 15 pixels. For each RGB channel, we estimate the mean of the convoluted training samples. Finally, we gather the means by averaging. The test step aims to determine the labels corresponding to each pixel of the optical input image. The same convolution as previously is applied to each pixel (same window size) and a RGB mean is estimated. For each pixel, the mean is compared to the mean of the training set, and the smallest Euclidian distance corresponds to the best class for the considered pixel. The classification results are assessed both qualitatively via a classification map and quantitatively via a confusion matrix. The confusion matrix is estimated thanks to a manually annotated ground truth.

In the second solution specific parameters (mean and standard deviation) are covered from local neighborhood histograms, on each channel (RGB and HSV channels). Then, they are compared to the input channels. After comparing, results are modified by well-known relations (thresholds) between samples. Each class of classified pixels are stored in binary images, then in case of need they are dilated and/or eroded. The final image contains the collected classes with different colors for each class. Classes are collected in the following order: forest, field, lake and city class. Black pixels represent the unclassified pixels.

Training images

Training images have been created from screenshots of Google Earth application, Google Maps pages and have been downloaded from different gallery databases (Please, see the References). The orientation of the images are different but usually top-north, bottom-south while the orientation of the downloaded satellite images depend on the actual content. The altitude of the Landsat images are usually around 20 kms above the surface of Earth. We would like to emphasize that the input and output images will not be used for commercial purposes. Small sample images have been used to teach the program how to recognize specific landscapes.

Input and output images

Like training samples, input images are taken from the sources mentioned above. The elevation is similar to the training images. Different types of images have been selected. These images were collected to a database. In these examples, the results of our algorithm can be presented on different types of surfaces. The output images contain some false colors representing the same type of surfaces.

These images show the results obtained with our first approach.







And these are the results obtained with our second approach:



References

Zoltan Kato, Ting-Chuen Pong: A Markov random field image segmentation model for color textured images

Satellite Imaging Corporation, http://www.satimagingcorp.com/

Google Earth: http://earth.google.com

Google Maps: http://maps.google.com