Remote Sensing, Image Classification, and GIS.

The term Remote Sensing includes all sorts of methods for gathering information from afar. Our eyes, for example, common snapshot cameras, sonar, radar and scanners of all types including those carried on satellites. The latter are most interesting from the perspective of this course for several reasons:

• Their broad-scale geographic coverage
• The record that they produce are accessible to systematic information processing.
• Their remote, automated operation makes their information available despite local bureaucratic data access problems.
• Since satellites have been collecting data since the 60's, their imagery is often the best if not the only record of the geograqphic history in some areas or subject areas.

As we often say n this course, that much of the matter of information systems boils down to these four basic capacities:

• Structures for storing references to entities and their attributes
• Procedures for selecting and associating entities based on their references.
• Procedures for deriving new entities or attributes by transforming existing ones
• Interfaces for accessing and combining procedures, and visualizing data

Data Structures

The data structures created by satellite imaging systems are images, like our familiar jpgs and tiffs. As such, they are by definition rasters. The fundamental unit of analysis is a cell that records a the intensity of reflected light. Our familiar tools and procedures such as those provided by Raster GIS and Photoshop will operate on our satellite images. The primary difference between most satellite images and the tiff and jpg images that we collect with our pocket cameras, is that the scanners that we send up into space are not so much concerned with recording the same sorts of light that our eyes do. While our visual system normally makes use of combinations of the Red Green and Blue slices of the spectrum, satellite scanners divide the spectrum into finer divisions, and also collect reflectances of radiation that we can't see. So, while we may still be discussing pixels and cells (our familiar raster data structures) extracting all of the meaning that multispectral images have to offer takes us through a discussion of the electromagnetic spectrum.

• 1 Bit black and white
• 8 Bit Gray Scale
• 8 Bit color mapped
• 24 Bit Color (8 bits assigned to each Red Green and Blue)
• Multispectral can represent other bands of the spectrum)
• Hyperspectral (lots and lots of finely segmented spectral bands)

Here is a nice page of illustrations from Dr. Ze-Nian Li of Simon Fraser University

Procedures for Associating and Transforming Multispectral Images

Like the earth itself, images of the earth's reflectivity is very unstructured. Even though our eyes may be able to see useful patterns in these images if we color and overlay the various components, as in the two images of Las Vegas taken in 1972 and 1992. We aren't able to systematically access this information as a land cover map as we would in Raster GIS until we have associated groups of pixels together, identifying classes of more or less homogenous land cover. See Picture This classification process is accompilshed through a class of software known as Multispectral Image Classification Software. This software uses statistical techniques and information we may know abour the reflective properties of things to develop reflective signatures that can be fit to unstructured, fuzzy, data.

First a Little History

THe history of remote sensing is like the history of photography and the history of the space program. NASA's has a very good history of remote sensing in their Remote Sensing Tutorial A couple of interesting bits that aren't highlighted in the NASA history include the history of the recently unclassified Corona project (this page by Kieth Clarke at UCSB) which began collecting space images in 1962. In 1972 NASA sent up the LANDSAT satellite which has been systematically sending back snapshots ever since. This data, available as part of NASA's Mission to Earth is one of the greatest data values on the internet. Especially since you may now download these images FREE from a the Global Land Cover Facility at the University of Maryland WE should also bring you up-to-date on the recent history including the emergence of private-sector providers of space imaging such as: DigitalGlobe and SpaceImaging Which can get you very crisp color images with nearly half-meter resolution of nearly anywhere for about $2,500. The private sector involvement in this industry has introduced some interesting new parameters including a check on previously secret information (or mis-information) and also healthy competition in terms of quality and price.

The Electromagnetic Spectrum

It is beyond the scope of this class to discuss photons, waves and beams, but suffice it to say that the sun is a huge source of radiation that is absorbed and reflected at different rates by different materials. Some Active remote sensing devices such as RADAR create their own radiation. Other imaging systems record long-wave infrared (heat) that is emmitted from objects. The images we will be discussing -- most of all useful geographic images -- are recorded by passive imaging systems that record reflected radiation that initially comes from the sun.

Different materials are distinct, not only in the amount of radiation that they reflect, but in the different wavelengths. For example, your blue pants are reflecting a lot of blue light and absorbing relatively more red and green light. Red, green and blue are spectral bands or slices of the spectrum, if we organize it by wavelength or frequency.

Thanks to Utah State University and NASA for these images.

A great page on the electromagnetic Spectrum at the NCSA Science Expo.

Here's a reference with more detail in the visible spectrum (By Doug Ramsey, Utah State University).

Imaging Systems

Here is a detailed overview of how data are returned from the Landsat Thematic Mapper from NASA's Remote Sensing Tutorial.

Visible and Invisible Radiation

False Color Images

In order to understand a photograph that includes information on reflectivity in the near infrared forces us to think in a new way about the way we look at photographs. Because of the limitations of our eyes and the computer displays that we use to interface with data, we have to learn to see these images, not as records as the way things look, but as records of invisible aspects of things. For an excellent treatment of the subject, see this page from the USGS.

Spectral Resolution
Thinking about dividing specific parts of the electromagnetic spectrum into slices, and alotting 255 levels of intensity to each slice gives us a new area of resolution to worry about -- Radiometric Resolution which refers to the relative narrowness of each band, and therefore the specificity of each of the 1/255th slices.

This image is from the web site for France's Spot Image Corp

A good blurb about radar

Spectral Signatures
Certain types of objects reflect and absorb different parts of the spectrum. This is why we see things as different colors. A red shirt is reflecting a lot of red light, and absorbing most of the blue and green part of the spectrum. In fact, for the visible part of the spectrum, you could say that the spectral signature of a red shirt is 90% to 100% reflectance in the red band, and 0% to 10% in the blue and green bands. Look at the line for water in the chart above: the spectral signature for water is about 10% in the green, 3% to 5% in the red, and practically 0% reflectance in the Infrared. The fact that water absorbs practically all the infrared light that hits it makes the IR band particularly useful for detecting water (wet areas reveal themselves as dark areas in infrared images).

Spectral signatures for various types of objects is best by setting up calibration sites on the ground as was done in this Hawaiian remote sensing project

While it is possible to determine the exact spectral signature of a blade of grass, the resolution of most sattelite or even airborne sensors (that we know about!) will usually be picking up refelctance from more than one sort of material, yielding what some people call mixed pixels (or mixels) . The process of trying to sort out the involves lots of mathematics and physics.

The annoying Details

Measuring spectral signatures of objects can yield ideal reflectivity information. Of course, the measurements made from space are often muddied up by cluds and haze. This is why hard-core remote sensing engineers often discuss radiometric correction which attempts to compensate for the atmospheric conditions that may be filtering the reflectance.

Image Processing Software

An understanding of the reflectivity of various objects is key for serious interpretation fo remotely sensed images. But very useful information can be derived by harvesting spectral signatures from images when you have prior knowledge of what was on the ground. This technique is known as Image Classification. In the simplest sense, the object is to classify pixels according to their intensity. To make it a bit more complicated, you can classify pixels in multi-spectral images into groups of similar intensities accross several bands.

The idea of classifying areas with similar signatures is important because of the inherent non-exactness of spectral signatures and of remote sensing technology. Assignment of a location to a particular class is accomplished through the use of one of several Clustering Algorithms which can be illustrated as mapping each location in a multi-dimensional graph with one axis for each image band, and breaking up the multidimensional space into envelopes of 'color-space.'

Unsupervised Classification
There are two primary ways of classifying the locations of an image. The first approach is normally to turn the GIS loose with its clustering algorithm to find a predetermined number of 'natural classes' in the image. This approach usually yeilds a set of classes that may contain most of the detail you wanted, and more. Classes can then be merged to fine-tune the classification.

Supervised or Trained Classification
An alternate, or supplemental method of classifying an image involves the explicit identification of areas in the image as prototypes, or Training Samples to seed the clustering algorithm. Normally, a Maximum Liklihood algorithm is used to categorize each location with the group that it is most similar to.

Here are some examples of image classifications that were done in previous years of this course:

• Sissy Willis
• Luis Bresciani
• Our next lab is another example.

• Natural Resources Canada
• American congress on Photogrammetry and Remote Sensing
• NASA's Remote Sensing Tutorial.