Satellite Data Correction and Calibration

The SPOT PAN scene was purchased as a level 1B (6) processed product. Pre-processing was applied by the SPOT Image Corporation, France. The process involved detector equalisation and geometric corrections. Detector equalisation consisted of compensation for the differences of sensitivities between the elementary detectors of the Charged Coupled Device (CCD) arrays, using a linear model, ie the raw data were radiometrically (7) normalised to reduce striping due to variable performance of the detectors. The pixels of the raw product had been re-sampled to 10 metres.

The geometric corrections were performed with the aim of compensating for the internal distortions of the image caused by the imaging conditions (variations of the attitude of the spacecraft, panoramic effect, earth curvature, rotation of the earth during the imaging of the scene, etc). The transformation model used for the level 1B is such that two consecutive scenes along a same data strip match together (ie there is a perfect registration between the overlapping portions of the two images). The orientation of the processed image is the same as the one of the raw image (no rotation of the lines is performed). When displayed, the Spot Scene image is a parallelogram. The ancillary data (coordinates of the scene centre and of the four corners, location model) allow to locate any point of the image with an accuracy better than 500 m (rms). The distortion of lengths is less than 2 x 10-3 (for a flat terrain).

Further processing was applied to the SPOT image to correct for terrain distortions and improve the locational accuracy. Terrain distortions were more pronounced in the SPOT image, whose angle of incidence (relative to the earth's surface) was 98 for the selected scene and were corrected by ortho-rectification. First order polynomial was used to carry out the rectification. Ten control points were selected from the 1:20,000 topographic data (Lands (1997 1:20K)) with a RMS (root mean square) error of no more than one pixel (10 metres). Image co-ordinate residuals (computer-observed) after correction for elevation displacement is shown in Table 4.2a. A digital elevation model was used as an input to the ortho-rectification process.

(6) Most satellite imageries are supplied at different pre-processing levels. The lowest level is the raw data (level 1A for SPOT image) without corrections of any kind. The higher levels (e.g. levels 2A and 2B) may require the use of local data for the purpose of geometric registration and ortho-rectification.

(7) Radiometric correction refers to the adjustment of the digital numbers (values inside each pixel) of the imagery.

The Landsat TM imagery was purchased as a level 4 product. Processing applied by the data vendor was the same as the SPOT PAN imagery. Each of the two sub-scenes was cropped to the approximate extents of the study area for further processing. Geometric registration of both Landsat scenes was improved by registration to the ortho-rectified SPOT PAN image. Around 45 control points were identified in each scene with a RMS error of no more than one pixel (30 metres). Rectification of the Landsat TM imagery was then performed using a polynomial Ground Control Point (GCP) warp, using nearest neighbour re-sampling and a linear polynomial order.

Atmospheric corrections were applied to one of the Landsat TM scenes (121/45), which was partly affected by haze. Atmospheric corrections are necessary in haze affected imagery as particulates in the atmosphere (aerosols) mask the objects being imaged or alter their spectral signature (Lavreau 1991). The haze correction methodology employed in this study is that designed by Chris (1984).

Atmospheric Corrections were performed in ER Mapper. The sequence of steps to achieve the atmospheric corrections was as follows:

Determination of Haze-threshold: The haze-free threshold (between hazy TC4 and haze free TC4o pixels) was determined for each band in turn by plotting TC4 against TMx for pixels in the training polygon. A line of best fit was drawn through the lowest DN values, thus allowing the determination of the between haze (TC4) and TM values (Lavreau, 1991). However, it was noted stage that the range of TC4 values was small, making it difficult to calculate the regression line. Similar results were reported by Lavreau (1991) for one of the TM scenes discussed in that paper. Nonetheless, regression lines were fitted and thresholds calculated to determine if the average threshold was appropriate. The following values were determined:

When the this threshold (27.33) was applied to the image, visual inspection revealed an excess of correction had been made. Similar results were reported by Lavreau (1991) due to a threshold that was too low. Thus, the threshold was gradually raised until the correction appeared, by visual inspection to be more effective. In this manner haze-free threshold of 60 was arrived at.

Application of the Haze Correction: Once the haze-free threshold was calculated, a TC4 image was added to the dataset as a new band so that it could be included in the correction formula given by:

(TMx,i)c = TMx,i - (TC4i - TC4o) Ax

where:

(TMx,i)c = the corrected value
TMx,i = the uncorrected value (x for each spectral band, i for each pixel).

This was implemented in ER Mapper for each of the six TM bands (1-5,7) as Input 1 - ((Input 2 - x)*y).

x = TC4o (or the haze-free threshold)
y = Ax (or the slope of each band/TC4 relation)

The Ax values used to correct each band, calculated from the TC4/TMx relationship in each band, were shown below:

Band 1 = 1.90
Band 2 = 1.00
Band 3 = 1.40
Band 4 = 0.85
Band 5 = 1.40
Band 6 = 0.71

As there is a strong likelihood of different atmospheric conditions between multi-date images, a comparison of digital numbers (values inside each pixel of the imagery) in an area common to the Landsat TM scenes was conducted to quantify this difference. The values in each band of scene 122/45 were then adjusted relative to scene 121/45 (to which a haze correction had been previously applied), by adding or subtracting from each pixel in each band, the average difference in digital numbers.