Landsat Gap Fill

The newest Landsat satellite (Landsat 7) was actually launched in 1999. Many Landsat users know that that this valuable earth science asset has been seriously broken since 2003 when its Scan Line Corrector (SLC), which compensates for its forward motion as it scans the earth's surface failed. This has resulted in systematic gaps in every Landsat image collected since that time.

To understand the problem, it is necessary to know a little about how the Landsat sensor works. As Landsat traverses an (approximate) north-to-south orbit, a mirror scans east-to-west collecting data across the orbital path. If uncorrected, a zigzag path would be described with wedge-shaped gaps in between the data. In order to make the transverse data rows parallel and remove these gaps, a mechanical corrector is attached to the mirror. This is the part that failed in 2003. All of NASA's attempts to fix the problem have been unsuccessful.

As a result, instead of nice parallel rows (like a well-mowed lawn) the uncorrected gapped pattern is transmitted to the ground station. This cannot be compensated by post-processing at the Landsat ground station because gap data is physically absent from the scan. Approximately 22% of the data is lost, with more lost information occurring at the edges of the image than at the center, where there is a vertical strip of essentially complete data.

Although there is no way to recover the missing gaps, in many cases it is possible to make these images more useful by filling in the gaps with overlapping data or data collected before 2003. This is possible because various Landsat satellites have orbited the earth for so many years that there are multiple many scans of the exact same patch of the earth's surface (these are called scenes). This legacy data can be used to fill in the gaps in more current scans. PANCROMA^TM has the necessary utilities to perform such gap-filling operations.

The gap filling procedure requires a Landsat Reference image (the one you are trying to correct and which has the gaps) and at least one corresponding Adjust image (the one you will extract data from to fill the gaps in the Reference image.) The Adjust image is usually gap-free, but it does not have to be. Because the gaps in two images may not coincide exactly, at least some useful data may be transferred from the Adjust image to the Reference image despite possible gaps in both.

NOTE: If you are using two gapped images, rather than one gapped and the other gap-free, there are two ways to combine the images. It is often adventageous to try both ways as one way may yield significantly better results than the other.

There are some fundamental problems with a gap filling strategy. The first is image registration. In order to transfer data from one image to the next, the two images must match exactly. The corner latitudes and longitudes must match and the number of rows and columns must be the same in the two images as a minimum. Constructing a set of such matched images is a vital part of the pre-processing work that must be done before the gap filling can begin.

Another problem is inconsistencies in lighting, season, snow cover and cloud cover between images. Two images of the same area, even when properly matched, may differ so significantly in these properties as to make them unattractive for gap fill. Of course great changes can occur over time in urban areas and deforestation may occur in remote areas. Other secondary factors such as radiance and color cast due to lighting and atmospheric effects are also important. However these can be somewhat corrected by PANCROMA^TM.

This procedure will describe the process for gap filling a 30m resolution Landsat RGB image from three band files using PANCROMA^TM. I have chosen an image of southern New Zealand for this tutorial to illustrate the process for producing a useful gap filled image. Start the gap filling process by selecting and downloading your Reference (gapped) Landsat scene. Record the Reference image Path and Row numbers when you download the image. Download the blue, green and red bands of the Reference image.

The next step is to select your Adjust image. An easy way to do this is to use the GLOVIS website as it archives many Landsat scenes of the same Path and Row number. After navigating to the site, select 'Landsat Archive' and 'Landsat 7 SLC-on (1999-2003)' from the drop down menus. Click the 'View Images' button. When the Visualization Viewer screen appears, type the Path and Row numbers that you recorded earlier into the text boxes provided. Now click the 'Prev. Scene' and 'Next Scene' buttons to browse through all of the available scenes at your selected Path and Row numbers.

You can use the 'Max Cloud' drop down menu to eliminate scenes with excessive cloud cover. The ideal case is when both the Reference and Adjust images have 0% cloud cover. Selecting this will of course limit the number of scenes that you can select from. However for many parts of the earth there are multiple cloud-free scenes available. Once you have found an appropriate Adjust image, download the blue, green and red bands. It is recommended that you set up a file structure to keep everything straight during the subsequent operations.

The next step is to subset and register the three pairs of Landsat band images. First load the three reference bands (blue, green and red). Select 'Band Combination' | 'Subset Images' | 'Subset Three Bands' from the main menu. After a moment a single grayscale image and subset form will appear. Select the subset area using any of the three methods provided (pixel coordinates, latitude/longitude, or rubber band box). Once you have selected the corner coordinates, select 'Enter'. The three band files will be subsetted and will be displayed on the screen. Save the images by selecting 'File' | 'Save Subset Images' and selecting either '.tif', 'jpg' or 'bmp. Select 'Close Graphics Windows and Reset'.

Now open your three Adjust (non-gap) file bands. Select 'Band Combination' | 'Subset Images' | 'Subset Three Bands' as before. This time when the subset coordinate box appears, the corner latitude and longitude coordinates that you chose for the Reference image are already entered into the appropriate text boxes and the 'Select by Coordinates' radio button is also selected. You need only select 'Enter' and the Adjust file will be subsetted to the same latitude and longitude coordinates as the Reference subset images that you previously created. I named my subsetted files 'Reference' and 'Adjust'. (Band identification suffixes are automatically attached. The band1 Reference and Adjust images are shown on the right.

The Reference and Adjust subset images may not be exactly the same size after this operation, i.e. the rows and columns may be off by one. It is necessary for the images to not only have corresponding latitude and longitude corner coordinates, but they must have the same number of pixels as well. PANCROMA^TM has a utility for resizing images in order to fix this problem. The resize utility uses a simple x and y scaling algorithm to achieve the fit. In general it is more correct to co-register both images using control points but since the images will be nearly the same size after subsetting the scaling approach will yield the same result for matching the images to each other.

PANCROMA^TM will decrease the size of the larger image to match the smaller one. As a result it is necessary for you to load the larger set by opening the three band files. Then select 'Band Combination' | 'Subset Images' | 'Resize Images'. The three images will appear on the screen along with the 'Resize Data Input Form'. Enter the row and column values for the smaller set into the text boxes and select 'OK'. The images will be scaled down. You can save the resized images when the process is complete.

At the end of this operation, you should have six processed Landsat band files: two blue, green and red band files for the Reference image and two blue, green and red files for the Adjust image, all the same size and with the same corner coordinate in the same projection to the same datum. One set will have gaps and the other will be whole. You are now ready to gap fill and process the images using PANCROMA^TM.

The process is fairly straightforward. The six band files are loaded in pairs: first the two blue bands, then the two green bands, and finally the two red bands. Each pair is gap filled immediately after loading by selecting 'Gap Fill'|'ETM+ Gap Fill' from the menu. Each gap-filled grayscale image must be successively saved, and then the next pair is loaded, repeating the process. PANCROMA^TM uses the non-linear histogram matching algorithm to automatically modify the 'Adjust' image in order to get a close match to the 'Reference' image. Without the histogram match, the radiance differences between the Reference and Adjust images would result in a severely striped image in most cases.

My example gap filled images are shown to the right. There is some faint striping visible but much of this is due to seasonal differences in vegetative covering. Both these image pairs show the benefit of gap filling and the satisfactory results that can be achieved.

The process outlined can also be used for pan sharpened images. In this case the images would be pan sharpened first., then the grayscale band files would be generated by loading a pan sharpened image while checking the 'Generate Band Images' check box. After the band files are thus obtained for the Reference and Adjust images, they can be gap-filled exactly as described above.

Note: see other articles on the Landsat gap fill problem and other subjects by viewing the White Papers on this site.