Have you ever had the fun experience of trying to order some satellite data only to be confused over all the "Raw", "Level 1B", and "Path-Oriented Plus" techno-babble? In spite of some concerted efforts made by most of the image distribution agencies around the world to standardize their use of these terms, ambiguities still abound. In this month's column I hope to level the playing field so you can order your data with confidence.
All of these terms refer to the amount of processing which is done to your imagery between the time it is received on the ground and the time you load it into your computer. Since most remotely sensed data require similar amounts of basic processing before they are usable, image distribution agencies have adopted a common set of processing "levels" to describe the types of processing done to their images before they send them out the door. As you will see from the descriptions below the processing levels are hierarchical; that is Level 2 data start with the processing included in Level 1 imagery and add more features to it.
Level 0 imagery is raw instrument data, just as they were collected at the sensor. Since there are some fundamental corrections that should be applied to the data before they are usable (these corrections are applied at Level 1), most agencies will not distribute Level 0 imagery. Unless you are a real hard-core remote sensing type who is specifically studying the sensing device and not necessarily the earth features being sensed, you don't want Level 0 data.
The next step up from Level 0 is Level 1A. Level 1A data have been corrected for detector variations within the sensor. To understand what this means, imagine that you were doing an illumination experiment and lined up ten 60 watt light bulbs all in a row. Now as part of your experiment, you needed to be sure that each light bulb is emitting exactly the same amount of light. Even though they may look like they have the same brightness, if you measured the light output from each bulb using a precision photometer you would notice some slight variations in the lumens produced by each bulb. The same principle applies to a remote sensing instrument, except that instead of having light emitting devices, a typical remote sensor has many light (i.e. electromagnetic radiation)detecting devices, or detectors. The number of detectors depends on the instrument design and can vary from 16 for the Landsat Thematic Mapper to 6000 for the SPOT HRV. Since much of remote sensing image interpretation depends on looking for small changes in pixel values, it is imperative that the brightnesses measured by each detector on a sensor is recorded uniformly across the scene. Satellite operators try to equalize the detector responses through a series of data calibrations involving pre-launch values and in-flight measurements. The in-flight readings come from pointing the sensor at targets with known brightness levels: an on-board calibration lamp, the blackness of deep space, and/or the brightness of the sun.
Thus Level 1A processing involves applying inter-detector equalization s and is sometimes referred to as radiometric correction. This is highly desirable if you wish to do any sort of classifications with your imagery. Also, absolute calibration coefficients are posted in the ancillary data and can be used to convert the pixel values into real irradiance measurements.
Another type of systematic error inherent in space-borne data involves the geometry of the image product. Consider, for a moment, some of the things that are happening as a sensor is acquiring an image: it is moving in orbit above the Earth's surface; the Earth itself is rotating beneath the satellite; and the sensor's view of the surface can be considerably more oblique towards the edges of the scene than it is directly below. All of these variations (and a few more) create distortions in the geometry of the imagery such as mis-aligned scan lines and non-uniform pixel sizes. Since all of these distortions are predictable and measurable systematic corrections can be applied to the imagery to improve its geometric qualities. This is the operation of Level 1B processing and is also highly desirable. The systematic modelling is not precise, however, so the position accuracy of Level 1B data is still in excess of 100 m. RADARSAT International refers to Level 1B processing as a Path-Oriented product.
Following along the processing hierarchy, Level 2A images are an "improvement" to Level 1B images. Level 2A images have been systematically mapped into a standard cartographic map projection based on a prediction of where the satellite was when the image was acquired. Since we are still dealing with systematic geometry correction approximations, however, these images still have expected location errors of more than 100 m. RADARSAT International refers to Level 2A processing as a Map-Image product.
Level 2A images are frequently labelled as "geo-referenced" which I find misleading when you consider their poor location accuracies. While it is true that you can input a Level 2A image directly into a GIS using the supplied projection information, I question how much spatial analysis you can really do given its poor positioning accuracy.
If you really want accurate spatial positioning on remotely sensed imagery, you need to have your images processed to at least Level 2B (RADARSAT: Precision Map Image). Prior to this Level, the bulk of the processing can be done in an automated fashion. In order to create a precision geo-referenced product, however, considerable user input is required. Through a process called geometric correction or image rectification, the image analyst "registers" the image to an existing base map by selecting pairs of well-defined points (known as ground control points) from both the image and the map. When a sufficient number of ground control points have been accurately identified the image can be geo-referenced so that its geometry will match that of the base map to which it was registered. The position accuracy of Level 2B images generally matches the spatial resolution of the original data (e.g. 30 m for Landsat Thematic Mapper; 10 m for SPOT panchromatic), except in areas of high local relief.
If the region you are trying to map is quite mountainous you will need to account for relief displacement in order to obtain consistently high position accuracies. In addition to manually locating ground control points (as in Level 2B), a digital elevation model (DEM) must be supplied to the procedure so that the relief displacement at differing elevations can be accounted for. This process is generically called ortho-rectification. The position accuracy of Level 3A images generally matches the spatial resolution of the original data, including areas of high local relief.
If your region of interest is really big and you need several scenes to be mosaicked together, then Level 3B is for you. These images have all the same attributes as Level 3A scenes, but cover a larger area.
Tidbits of note
Interestingly, Landsat and SPOT processing systems reverse the roles of Levels 1A and 1B (Landsat corrects geometry in Level 1A and applies detector equalizations in 1B while SPOT does it the other way around). This does not matter when ordering Level 1B imagery, however, since at this Level both corrections have been applied and the order of application is inconsequential.
Although each Level increment produces a more map-like product, this is at the expense of radiometric detail (not to be confused with spatial detail). This means that more and more of the subtle variations in pixel values will be averaged out of the imagery at higher processing levels. While this might not be important if all you want is a background picture for your other spatial data layers, it may inhibit the ability to achieve detailed land-cover discriminations during classifications.
There is also the cost factor to consider. Data from any of the automated processing levels (Level 0 - Level 2A) generally costs the same price, since the extra computer time required between increments is insignificant. When the processing begins to require analyst intervention, however, the prices begin to jump. Expect about a 25% premium between Levels 2A and 2B, and an additional 25% increase when going from Level 2B to 3A.
So, what will it be?
What is the right level for you?
If you have some image experience and access to a remote sensing image analysis system or robust GIS with geometric correction capabilities, then Level 1B data are for you. Level 1B imagery has had both inter-detector equalizations and systematic geometry corrections applied. It forms a good base for most remote sensing applications. You can't just plunk these images into your GIS, however, since they lack geo-referencing. But, image rectification procedures are not hard to use on most systems and (with some initial guidance) even a novice can create their own Level 2B geo-referenced imagery in half a day.
If you don't have the necessary software and/or time to process the imagery, then you should consider purchasing your data at Level 2B or 3A, depending on the amount of relief in your area. These images will import directly into a GIS and be ready for use. Be aware, however, that detailed land-cover classifications may not be possible at these Levels.
Now that you know what I think, let me know what you think! E-mail me at firstname.lastname@example.org.
I recently came across an interesting offer for colleges and universities from SPOT Image Corporation. For only US$200 (regularly US$1500), you can purchase a recent (1999 or newer) SPOT image (PAN or Multispectral) of your campus. Called the "Campus Scene" offer, it is only supposed to be available in the Continental U.S., but they will consider requests for Canadian campuses if they have the imagery. Check out http://www.spot.com/HOME/NEWS/campusscene.htm for details.