Calibrating Images from Different Dates

A report from the LWRRDC project

"Detecting and Monitoring Changes in Land Condition

Through Time using Remotely Sensed Data"

June 1994

Norm Campbell, Suzanne Furby, Brian Fergusson

CSIRO Division of Mathematics & Statistics

Department of Agriculture Western Australia

1   Background

With funding from the Land and Water Resources Research and Development Corporation, the CSIRO Division of Mathematics and Statistics is undertaking a project to develop methods for detecting, measuring and monitoring change through time in land condition using remotely sensed data integrated with landform maps and other spatial data. The particular issues being studied are:

• the change in remnant vegetation status and condition;
• the change in area affected by salinity;
• tracing the temporal response of crop and pasture in successive seasons; and
• the extent of wind erosion through time in relation to season and management.

This work is being carried out in collaboration with the WA Department of Agriculture and the Remote Sensing Applications Centre of the WA Department of Land Administration.

Landsat TM data are being used in the study. Approximately 48 scenes have been purchased to date, covering the whole of the agricultural area and focusing on regions in the northern and eastern wheatbelt, the Upper Great Southern and the south coast for the initial study. These scenes range in time from 1987 to the present. Extended sequences within particular years have been acquired to determine which times of the year are most appropriate for detecting and monitoring the different types of land condition and degradation.

2   The Importance of Calibration for Monitoring

An important step in the development of methods for monitoring change is the ability to compare images from different dates and sites in different scenes. These comparisons require the digital counts from each scene to be calibrated to common reference values. The raw data values recorded by the sensors on the satellite are not consistent in this way.

The sensors record the energy from the sun reflected by the Earth's surface. The amount of incident energy varies throughout the year with changes in the solar zenith angle. Both the incident and reflected energy are affected by atmospheric scattering. The amount of scattering varies with factors such as temperature, humidity and haze. Sensor decay through time and on-ground preprocessing also alter the raw data values.

The calibration process is applied to the raw data to remove these time and scene-dependent effects. It is sufficient to convert the raw digital counts to be consistent with the counts for a chosen reference image when comparing images to detect change. This is the approach that has been implemented for the data used in this project to date.

Ideally, all images would be calibrated to reflectances. Reflectance, which is defined as the percentage of incident radiation reflected by the surface material, is a physical property of every substance. It can be measured under laboratory conditions. If the images are calibrated to reflectances, image pixel values can be compared directly with field and laboratory measurements, and changes can be related to physical properties of the ground cover.

In practice, various methods are used to estimate a calibration line to convert the digital values of an image to a reference image (See Appendix I for the calibration equation). The slope and intercept of the line are referred to as the gain and offset.

3   Calibration to Reflectances

Two methods are being examined to calibrate image data directly to reflectances.

The first approach is based on a two-stage spectral mixture analysis (see, e.g. Smith et al., 1990). The digital counts are first related to image endmembers. In the second stage, these image endmembers, calibrated by a gain and offset, are decomposed as mixtures of reference reflectance spectra.

Experience to date has indicated that this approach cannot, by itself, estimate gains and offsets for calibrating to reflectances. It appears that only the ratios of the gains can be estimated uniquely, unless there is very careful choice of starting values and image endmembers. This means, in practice, that at least two targets with known reflectances must be available for an image in order to estimate the gains and offsets uniquely.

The second approach which is being examined is a scene-based radiometric calibration of remotely sensed data developed by Dr. Joachim Hill of the Joint Research Centre at Ispra, Italy (see, e.g. Hill and Aifadopoulou, 1989). This approach makes a number of corrections based on known physical parameters (including sun elevation and distance to the satellite), calibration coefficients relating to at-satellite calibration and an estimate of aerosol optical thickness. The estimation of the latter parameter requires a known dark target, preferably water.

Operational details of these procedures are still being studied. As part of this study, field reflectance spectra of invariant and vegetated targets are being measured at monthly intervals throughout the year to provide ground information to verify the calibrations.

4   Calibration to Like-Values

Invariant targets are being used to calibrate the digital counts of coincident or overlapping scenes to the digital counts of a reference image. The digital counts of these targets are extracted from both images and robust regression techniques are used to estimate the gains and offsets.

The invariant targets used include quarries, gravel pits, lakes, reservoirs, bare ground, salt pans, well-maintained town ovals and sporting facilities and bare rocks. They are selected based on ground information and features identified on 1:100 000 map sheets.

4.1   Using Like-Value Calibrated Images

Simple displays of sequences of calibrated images with the same band combinations and stretch limits immediately convey changes in the images. Training sites extracted for various condition classes can be applied from image to image. Areas which are poorly classified on images other than the one(s) on which the classifier was trained will be indicative of areas which have changed spectrally to a class not covered by the original training data.

Figure 1 shows a raw and calibrated image. The darker winter image has been calibrated to the spring reference image. Differences between the calibrated image and the reference image are due to the progression of the growing season.

Figure 2 shows a sequence of images throughout a growing season from sowing (June) through anthesis (August) and curing (October) to a post harvest (December) image.

5   Guidelines for the Calibration of a Sequence of Images

5.1   Selection of reference image

The reference image is the scene to which the other scenes are related. It is important that it is cloud free and contains no white outs (i.e. values >= 255). Below is a list of steps for the selection of a reference image.

• Determine what time of year is appropriate to the application.
• Check data values for 0 and 255 in all bands, particularly band 5, to avoid white outs.
• Check for cloud cover.
• Select image with minimum cloud cover and 0

5.2   Selection of Invariant Targets

Invariant targets are features which have constant reflectance over time. The data values are used to define linear functions to transform each overpass image to the reference image by assuming these targets should have the same digital count values in each image. Targets must be selected for a range of bright, mid-range, and dark values and you must have a balanced number of bright and dark targets. Possible targets are listed below.

Vegetated targets should usually be avoided as they show seasonal trends. Very dense forest has been used successfully, however, in the Collie - Pemberton region between September and November images when few other targets have been available.

Targets can be located by viewing the reference image on the screen and using 1:100 000 hardcopies and maps. When obtaining targets, it is important to locate targets over a uniform section of the feature. The enhancement is important when determining uniform areas. The stretch must be taken over the feature as many features will appear to be uniform with a standard stretch over the whole image. The size of the targets must be no greater than 3x3 and for small features it is important to locate targets in the centre to avoid errors due to misregistration. When obtaining targets, it is important to record information about the target. A list of such information is below.

• Target type
• Size
• Line and pixel location
• Geographic location
• Map sheet

When the images to be calibrated are not registered, the line and pixel coordinates of the target should be recorded separately for each image. With such images, you should try to avoid using pixels from very small targets or near the edge of larger targets.

The following are some hints on the selection of invariant targets.

• Bright and dark targets can be determined by combining band 3 of three dates into a 3 band image and quantizing to reveal 5% of the scene.
• Lakes in the wheatbelt are unreliable as water and sediment levels change. The darkest and or largest lake is rarely the least variable.
• Rock outcrops appear as a gold color with bands 3,5,7 in b,g,r.
• For airfields, take the intersection of runways.
• Avoid features near cloud and cloud shadow.

5.3   Calculation of coefficiants

The next step is to calculate the regression coefficients which relate the overpass images to the reference image. Coefficients for least squares, s-estimation and weighted least squares estimation procedures are calculated. The s-estimation method fits a line to 58% of the data in each band separately and assigns a weight to each point. The weighted least squares method uses the minimum weight from all bands of each point to determine the weighted least squares line.

5.4   Coefficient Examination

Once the calibration functions have been calculated, the next step is to examine and select the best calibration line for each band. This is done by plotting the lines and data. The y-axis is used for the reference image data and the x-axis is used for the overpass image data. The linear functions are then plotted using a different linestyle for each plot. In the following examples, a solid line is for least squares, the dotted line is for s-estimation and the dashed line is for weighted least squares. A legend should be attached to each plot.

The following figures show examples of a good fit, and several common problems encountered.

Plot 1, above, shows an acceptable line.

Whiteouts may be evident in band 5 (values = 255) and may cause incorrect fits due to the lack of points, as whiteouts are automatically omitted from the fit. More mid-range targets must be obtained in this situation. This is shown in plot 2, above.

With cloud-affected dates, bands 5 and 7 tend to have more scattered plots, due to moisture in the air from clouds. This is shown in plot 3, above.

Plots 4 and 5, above, show cloud-affected plots. With plot 4, most of the data is cloud-affected, causing the valid data set to be down-weighted and the estimated functions to be incorrect. Due to the robust nature of s-estimation and weighted least squares, cloud-affected points had no effect in plot 5 as less than half the points were cloud-affected.

The final plot shows the possible consequences of having more dark targets than bright. The robust estimation procedure has fitted the trend within the dark pixels and ignored the bright pixels completely. Where none of the lines adequately represent the data, you should examine the targets selected and select additional targets.

Once the plots have been examined, the line of best fit is then chosen. The line of best fit will be either s-estimation or weighted least squares, least squares is never used. The line for weighted least squares is nearly always adopted, except when the s-estimation of the other bands has caused valid points to be down-weighted.

References

Hill, J. and Aifadopoulou, D. (1989). Scene-based atmospheric correction of Thematic Mapper imagery acquired during 1988 over the Ispra/Novara region, Italy. JRC Ispra, Institute for Remote Sensing Applications, Technical Report.

Smith, M. O., Adams, J. B. and Gillespie, A. R. (1990). Reference end-members for spectral mixture analysis. Proceedings 5th Australasian Remote Sensing Conference, Perth, WA. pp 331-340.

More Information

For more information on image calibration contact Suzanne.Furby@csiro.au.

Appendix I

To top

last updated 05/06/02
Xiaoliang.Wu@csiro.au