ERDAS : RADARSAT Processing Case Study

What is RADAR and how is it collected?

RADARSAT uses synthetic aperture radar (SAR) which is a microwave sensor capable of imaging the earth using different imaging modes including Fine, Standard, Wide, ScanSAR (narrow and wide), and Extended Beam (some examples shown in the image below). By sending microwaves with a different fingerprint at the earth at different angles, it can provide good coverage and see through cloud cover (an effect of the long wavelength). RADARSAT orbits Earth fourteen times a day and has a 24 day return to location cycle.

What is unique about RADARSAT versus other imaging methods (Landsat) is that it can image off nadir. This means that although it’s return time is 24 days for the exact same beam coverage of the area, every place on earth is imaged once a day using at least one of the beam modes. In the example below from MDA RADARSAT documentation it shows the Nadir track of the satellite and the different off nadir imaging beams. As the Satellite Ground Track shifts, different areas are imaged using a different beam with a different angle to the earth (closest is steepest angle). What this means is that even though the track shown below won’t be travelled again for 24 days, the area under the track will be imaged daily by the off-nadir imaging.

When choosing which beam you are going to use you need to consider the affect of the angle of the beam, what the necessary scale is, terrain, requirements for stereo viewing and repeat coverage.

Processing RADARSAT

RADARSAT products can be provided at different stages of completion including Raw Data, Path-Oriented, Map Image, and Precision Map Image.

Raw data cannot be viewed as a scene as it is an unprocessed matris of time delays packaged into CEOS format.

Path-Oriented products can be provided at three levels, the Single Look Complex, Path Image, Path Image Plus. Single Look complex data is assigned latitude and longitude positional information and has satellite reception errors corrected. This data my not be viewable. Path Image products are processed to the point where the scene aligns parallel to the satellites orbit path and gives latitude and longitude positional information for the first, mid and last pixel position of each line of data. Path Image Plus uses smaller pixel spacing and retains full resolution. This is good for identifying point targets and linear features but results in a much larger file.

Map-Oriented products can be processed to three stages: Map Image, Precision Map Image, and Ortho-Image. Map Image products are provided in the client-requested map projection (ideal for people who don’t want to do any processing).

Precision Map images utilize ground control points and map projection to align the scene and provide greater accuracy. Ortho-Image removes terrain distortion using a DEM.

 

Image depicts off nadir imaging and different beam modes.
Image from MDA Corporation documentation
http://gs.mdacorporation.com/products/sensor/radarsat/rsiug98_499.pdf
Image depicts off nadir imaging and different beam modes. Image from MDA Corporation documentation http://gs.mdacorporation.com/products/sensor/radarsat/rsiug98_499.pdf

Why use RADARSAT?

RADARSAT is great if you need current data that is easy to reference, if you need a range of scales and resolutions or flexible viewing geometry, or if you require frequent revisit time. Applications include but are not limited to determining surface roughness, moisture, land/water boundaries, anthropogenic features or topography. For more information refer to the below table or the website is was borrowed from http://gs.mdacorporation.com/products/sensor/radarsat/rsiug98_499.pdf

Image shows several uses for RADARSAT
Image from MDA Corporation documentation
http://gs.mdacorporation.com/products/sensor/radarsat/rsiug98_499.pdf
Image shows several uses for RADARSAT Image from MDA Corporation documentation http://gs.mdacorporation.com/products/sensor/radarsat/rsiug98_499.pdf

Case Study

The purpose of this project is to take two RADARASAT-1 SAR scenes in CEOS (Acres) format (provided by MDA Geospatial) and convert them to positioned mosaicked images (Ortho-Image level processings as described above but without the use of GCPs). The scenes were captured in Standard 7 Beam Mode and have been processed to Path Image processing level. This means that the SAR signal data is turned into the SAR equivalent of DN values and orbital information is included for georeferencing. The scene has orbital ephemeris information which gives a rough georeferencing model (approximately 100m accuracy). The data for this assignment is of Nova Scotia as shown in the location map (COGS, 2015).

Source and resolution information is available below.

 

RADARSAT Imagery Copyright [2001-2005] Canadian Space Agency
RADARSAT Imagery Copyright [2001-2005] Canadian Space Agency
Source information for the files used in the case study
Source information for the files used in the case study

Overview of Processing Required to Take RADARSAT Data from a Path Image to Ortho-Image Product

Importing RADARSAT Data

The first step is to use the Import Data menu and import RADARSAT CEOS data (in this case RADARSAT (Acres CEOS)). All default settings were used during the import. The header information in the CEOS file was used for positional accuracy. This is a setting that had to be toggled on. For more information look at the tutorial provided in the Tutorial section of this website or click here

Georeferencing Imported RADARSAT Data

Although manual georeferencing can be done, with RADARSAT it is easy to use the AutoSync Georeferencing Wizard. IMAGINE AutoSync uses a point matching algorithm to generate thousands of tie points on a roughly georeferenced image to an accurately georeferenced image using edge matching. The AutoSync Wizard uses Automatic Point Measurement (APM), a software tool for image matching, and applies it to the two images. One way to aid in this process is to make sure the two images have as much contrast as possible. One way to do this for RADARSAT georeferencing to Landsat images as being done here is to utilize the Landsat NIR band which creates a large contrast between land and water. The IMAGINE AutoSync generates tie points between images automatically.  If the data is not accurate enough, 3 to 6 GCP points can be added in to decrease errors. This is good for rough georeferencing, but if comparing to an image with ortho-rectification this process may not be accurate. Mosaicking and ortho-rectifying will further refine the accuracy. Information regarding the AutoSync Wizard was from the Imagine AutoSync User’s Guide available here.

An accuracy report is generated that supplies and RMS value. This RMS value is not based on a projection error but instead on the pixel displacement of the model. For this case study the RMS value was 4.77 pixels. Since the scene metadata tells us that a pixel is 11×15.5m (in the metadata there is an option to switch from degrees to metres), the error is therefor 4.77 * (11×15.5m). If your RMS values are poor or your image clearly does not line up properly with the reference image, try using different models (affine, polynomial, rubber sheeting) for the warping methods of the new output image. For more information on AutoSync Georeferencing refer to this guide.

OrthoRADAR – Orthorectifying RADARSAT Data to a DEM

Open the OrthoRadar tool. The purpose of this tool is to warp the image to fit a DEM model. It can also georectify an image using satellite orbital information. Data will be more accurate if previously georectified.

You must then choose the image you would like to orthorectify and a dem to use for the warping of the image. Using the output image from the autosync georeferencing wizard is recommended.

The next step is to choose the geographic international RADARSAT projection. Otherwise default values were used. For a step by step guide refer to this site.

  • Landsat

    Landsat

  • RADARSAT

    RADARSAT

  • Georeferenced

    Georeferenced

  • OrthoRectified

    OrthoRectified

Mosaicking using MosaicPro

Mosaicking is a puzzle. It involves playing with colour correction settings until one is found that makes the data join seamlessly. This settings include exclude areas, use illumination equalizing, use image dodging, use color balancing, and use histogram matching. You can use these options in combination or just pick one. For this specific case study I found that the histogram matching worked best. Another variable that can be adjusted is which image is on top. Sometimes it is easier to mosaick without a seam if one of the images is on the surface (less edge colour variation etc).

A second variable that can be set is the overlap/seamline function. This tells the model to either take the pixel values from the top image, to average to two, always select the minimum value, always select the maximum value or to feather them. Feathered images takes the pixels from the most appropriate image on a gradational scale based on position. Examples of some results are shown below. For this case study I used overlap.

SS_Mosaic

The above images all show one image overlapping the other. They all use the Overlap seam function unless otherwise stated. The image on the right was not the final used image. In the final image shown in the location map, the image order was flipped leading to the best result.

For more information on this refer to this tutorial or the ERDAS help files.

Image Comparison

Two ways to compare the georeferenced image and the orthorectified image are to look at change detection  between the two images including image difference and highlight changes. By measuring the change on the two images, it seems reasonable to suggest that the average difference between the two mosaics is 100-150m. On certain edges of the image this error can increase up to 450m. This does not help with determining absolute error of the image as this is just the difference between the two mosaicked images.

  • Change

    Change

  • Highlight

    Highlight

Results and Concerns

It is difficult to determine the overall error associated with the georeferencing/orthorectifying process. Visually the RADARSAT images line up with the landsat images. Considering that all of the referencing and orthorectifying was done automatically, the positioning is quite good. The error associated with the georeferenced image was expressed in the shift of light and dark pixels from where they should be based on the landsat imagery used for georeferencing. Points of error came from changed land use and water levels including lake and stream heights as well as tidal change. These errors could be a factor in the difficulties with the georeferencing of the images in the case study because the RADARSAT image was taken in 1996 and the landsat imagery was from 2001-2005. This affect could be minimized by using imagery that was gathered on the same date at similar times.

The off-nadir nature of RADARSAT can cause difficulty with the image match algorithms due to the shifted contrasts associated with viewing angle (shade relief). This is decreased when the beam is closer to the satellite. Using lakes, rivers and coastlines for edge matching reduces this affect due to high contrast and shape recognition.

The orthorectified data seems to be a slightly better approximation than the straight georeferenced data. This is difficult to see in a side by side comparison. This is likely because the landsat data is orthorectified, making it difficult to warp the RADARSAT data appropriately without DEM data.

Overall I would recommend that this method is useful for a small scale map that is looking for large land use change or doing a broad classification. For large scale work or detailed temporal comparison this method of referencing and orthorectifying is not going to be sufficient without ground control points and/or user input.