Lecture Index: Introduction. / Imagery use in the geosciences. / Commonly used types of imagery in geoscience. / Basic principles behind image filtering/analysis. / Classifications. / Exercise 7
Reading:
The basic goal of image analysis is to extract
information from an image. Remote Sensing is a discipline
that focuses on this goal in much greater depth, and we have a
course devoted to learning this topic. The goals of this weeks
exercise are: a) to introduce you to Adobe Photoshop, a
software package that manipulates raster images, b)
teach you some ways that images can be "enhanced", and
c) teach you some of the basic ideas behind such image
analysis, including the concept of classification.
This was altered in Adobe
photoshop from an original black and white photo of the Elkhorn
River near Scribner NE in an attempt to emphasize the river channel
forms. Old channel forms show somewhat selectively in green with
dark outlines. Notice the large amount of noise.
It is helpful to think about the chain of events in image analysis. It might be summarizes as having the following steps: an illuminator sends out radiation -> atmospheric effects occur -> reflectance occurs (surface effects) -> atmospheric effects occur -> sensor acquisition -> computer capture and processing (including rectification) -> analysis and classification -> end use. Thus, there are a lot of variables that influence the image and final product. We will focus on the last three stages in this exercise.
A Holy Grail of image analysis is that a distinctive feature has a distinctive reflectance signature. For example, does contaminated soil have some unique spectral or distribution signature that will allow for its distinction from all the other features in an image, permitting analysis of the extent and pattern of occurrence of that feature in the image. A major goal is to identify such unique signatures in the imagery. If a series of distinctive signatures can be determined, then the image can be classified. You can think of classification as assigning a pixel an attribute trait instead of just a reflectance value. For water vs. land as classes this can be pretty easy to envision. For contaminated soil vs. uncontaminated soil that could be much more difficult. Fuzzy classifications can be used here (i..e. it belongs 80% to one class and 20% to another). More on fuzzy set theory later. Once you have figured out a scheme that works you can automate analysis. The challenge is non-unique signatures, meaning that two different types of features 'look' the same. This is why multi-band satellite data is used. For black and white imagery you can see that developing a unique signature could be a real problem. For each pixel all you have is a gray value. For a color image for each pixel you can have a color and an intensity, and with multiband for each pixel you may have different band values for each pixel. But the more information you collect (a multi-band approach, where bands refer to different portions of the electromagnetic spectrum), the better the chance for determing a unique signature for a unique surface. If for each pixel location you consider more reflectance information exists, the possibility of identifying a unique response for a feature increases.
Below is a list of how imagery is commonly used in the geosciences.
Lineament analysis: this is a special case of a type of analysis that maps linear and curvilinear patterns.
Below is a list of
commonly used types of imagery. This list is far from complete
and is constantly evolving.
The image to the left is of part of the Mississippi River. It is clearly a false color image, and the bands involved are actually radar bands. The point bar sand deposits and channel scars stand out clearly in this image. Details can be found at NASA's Visible Earth link.
One standard approach is to look at the distribution of values (via a histogram of pixel value frequency) and then modify it. This consists of recomputing the values in the array on the basis of some modification algorithm. You could stretch, condense, remove, replace all or portions of the histogram. An asymmetric distribution could be changed into a more symmetric and centered one. The possibilities are almost endless - so, the question as to which possibility will be most helpful is a crucial one.
Thinking of a digital image as an x-y array of values (where multispectral then imagine stacked layers of arrays of numbrs) where the z value is the intensity of luminosity. If you take one value and then compute a new value for it on the basis of its neighbor values, you can create a new image.
Consider each of the below transformations for the z variable 'image' and describe what it should do to your image:
This is basically matrix manipulation (which is why many remote sensors love C++ and similar programming languages). The possibilities are almost endless. In the cases above you are comparing a pixels value to its neighbors and making a modification on the basis of the result. The window size, or how far away you look for neighbors, is also an important consideration
We will explore in class some simple matrix manipulation in Excel to see how it changes the 'picture/pattern'.
Classifying an image is an attempt to map out the distribution of features of interest, such as grasslands versus woodlands, or mineralized versus unmineralized rock, or simply different rock units. The mapping is based on identifying a spectral signature for the feature and then all the pixels (or other 'window' sizes) with that signature are assigned a common pixel value that represents that feature. The lure of classification is to automate mapping in the computer environment. It is a very challenging endeavor because the world is complex with lots of fuzzy boundaries and non-unique signatures, and it gives one a little more respect for the brain, that we so easily classify features in an image.
Modeling spectral responses: When classifying you could build a theoretical-physical model of how a surface type reflects, but it can be challenging. Consider the factors that determine spectral reflectance of granular material at the earth's surface (e.g. of a point bar):
You could make another list for a type of vegetative cover that would be much longer. In addition, there are illumination factors, such as the intensity and angle of illuminaitno. If you can develop a working model then you can use this to analyze and classify your image. In practice it is best done empirically, matching a known site of surface type with its spectral reflectance.
Sites where spectral responses for minerals and other materials exist. These can serve as a basis for building spectral response models.
There are very sophisticated programs for image analysis and classification. However, the associated learning curve is long, and beyond the scope of this class. Instead we will explore Adobe Photoshop, which can digitally manipulate raster files, and which you can learn to use significant portions of in an hour to several hours time. Adobe Photoshop provides no real capacity for true analysis (e.g. to measure percentage of coverage by a feature, to classify an image, to map gradients), but it will introduce you to basic and 'canned' raster manipulation.
Potentially useful links for further exploration (let me know if you find other relevant links):
Copyright by Harmon D. Maher
Jr.. This material may be used for non-profit educational purposes
if proper attribution is given. Otherwise please contact Harmon
D. Maher Jr.