How to Make More Out of Satellite Images
Hamed Hamid Muhammed
One of the most important fields of image analysis research is analysing remotely sensed images acquired by satellite and airborne systems. Remote sensing offers the opportunity to study a relatively large region on the Earth by a single image. In addition, satellite and airborne remote sensing analysis systems offer efficient and environmentally friendly non-destructive techniques.
My research aims to analyse these kinds of images in a way that takes into account the entire content of the available data set (which is an image) to extract as much useful information as possible from these data. The statistical method is one of the most commonly used and efficient techniques for this purpose, where it is allowed to have data affected by some kind of disturbing factors. This is important because it is impossible to make sure that we have exactly the same conditions when acquiring satellite images at different times, even when using the same instrument. Weather conditions and variations in the sensitivity of the instrument are the main disturbing factors.
Remote sensing systems acquire very large image data sets, which require fast and efficient analysis techniques. What the statistical methods are good at is for looking at the data and trying to know which part of it contains the useful information that we are interested in. This little part is what we actually need to focus on, because by using only this part, we still get as much useful information as possible, even if we had used the whole data set.
It is obvious that less computational and memory requirements are needed when reducing the amount of data we have to process. Statistical methods look at the statistical properties of the data samples to decide which of them are useful. Many statistical properties can be utilised here, such as the mean value of the data samples and how much each of them deviates from the mean value, etc. The following example illustrates the basic idea of using statistics. Suppose that we have two apples beside each other on a plate and an ant on the same table. The task for the ant is to take a photo in which the two apples can be recognized clearly. Most likely, the easiest way for the ant to do this is to go as far as possible from the plate and around it until the two apples can be recognized. Another way would be for the ant to be at a certain place on the table and move the plate away from itself (a translation) and then rotate it until it becomes possible for the ant to recognize the apples. The latter way is what we try to do when using statistical methods. But the hard task here is to find the suitable statistical properties of the data that can be used to get a better view in order to decide which of the data samples are useful for our analysis.
In the case of the plate with the apples, moving the plate away from the ant corresponds to multiplying the view by a factor smaller than one (a translation factor). Rotating the plate corresponds to multiplying the view by a rotation factor, which is more complex than the simple translation factor. However, it is possible to combine the two factors into one which can be used to produce a new view, one showing both the translation and rotation effects. Multiplying the view by a factor is called "projecting the view" on that factor. In other words, the new view is the projection of the view on the factor, or is what can be seen from the factor's point of view. An even better factor would be one that not only gets a better view, but also goes further and finds the interesting part of the view.
Many methods exist for finding such factors, called "basis vectors" in this field, and these methods are referred to as "transformations." The difference between these methods is in the choice of a suitable set of statistical properties of the data samples that are considered when finding the basis vectors. Finally, our image data set is projected on these basis vectors. So far, the resulting projections correspond to obtaining better views. The next step is to study these basis vectors to try to build smarter ones that can be used for both obtaining the best view and extracting the useful information from it.
|Modified 15 January 2003 * Contact Us|