Introduction

   Satellite imagery analysis has played a key role in environmental monitoring and modelling over the past few decades. Repeated observation of a given area over time yields the potential for many forms of change detection analysis. These repeated observations are confounded in terms of radiometric consistency due to changes in sensor calibration over time, differences in illumination and observation angles, and variation in atmospheric effects (Eckhardt et al., 1990).

   Radiometric correction of satellite imagery falls into two broad categories, absolute and relative. Absolute radiometric correction converts the digital number of a pixel to a percentage reflectance value using established transformation equations (Richter et al., 1990). Relative radiometric correction normalizes multiple satellite scenes to each other. For both categories, the majority of methodologies developed require ancillary data or the subjective selection of pseudo-invariant features (PIFs) in the imagery. A more exhaustive comparison of radiometric correction methods which require the use of ancillary data or ground targets can be found in Yuan and Elvidge (1996), Yang and Lo (2000), and Song et al. (2001). Most methods of radiometric correction are not applicable to all change detection studies and also require substantial time inputs and subjectivity on the part of the image analyst.

   Most forms of absolute radiometric correction rely on sensor calibration coefficients, atmospheric correction algorithms, and illumination and observation geometry coefficients. These data are used in a radiative transfer model to correct the imagery. While a considerable amount of research into radiative transfer models has been conducted, the application of these models to a satellite scene often requires atmospheric and sensor properties for the acquisition date of that scene. For the majority of archived satellite data, these properties are not available.

   Relative radiometric correction has several advantages over absolute radiometric correction. The methodology is usually simpler, requires less computer operating time and less theoretical understanding. More complicated algorithms do not necessarily perform better, and for most studies relative radiometric correction is recommended (Song et al. 2001). Relative radiometric correction usually involves the selection of ground targets whose reflectance values are considered constant over time, otherwise known as PIFs, and relating these targets to all imagery in the study. Selection of such ground targets results in a subjective radiometric normalization and also relies on the assumption that the reflectance properties of the ground target haven’t changed over the time interval. There are five generally accepted criteria for a PIF or PIF set (Eckhardt et al., 1990). These are:

1. The targets should be approximately the same elevation so that the thickness of the atmosphere over each target is approximately the same.
2. The target should contain only minimal amounts of vegetation because vegetation spectral reflectance is subject to change over time.
3. The targets must be in relatively flat areas so that changes in sun angle between images will produce the same proportional increases or decreases in insolation to all normalization targets.
4. The spatial pattern of the normalization target should not change over time.
5. A set of targets must have a wide range of brightness values for the regression model to be reliable.

   Features used as PIFs in previous studies have included lakes, beaches, new asphalt, old asphalt, concrete, and gravel (Caselles and Garcia, 1989; Coppin and Bauer, 1994; Elvidge et al., 1995; Pax Lenney et al., 1996; Yuan and Elvidge, 1996; Michener and Houhoulis, 1997; Yang and Lo, 2000). In many studies, the selection of appropriate PIF sets is not problematic and high quality radiometric correction is possible. However, in some studies the presence of suitable PIFs can be confounded by any combination of variable cloud cover, variable climate leading up to the date of image capture, high topographic complexity in the imagery and lack of urban development. Additionally, the longer the time interval between satellite images the higher the probability that any given PIF will have experienced change. It has been noted that, in some areas, PIFs with constant reflectance do not exist (Du et al., 2001; Du et al., 2002).

   The time period covered in this study is eighteen years, represented by a series of eight satellite images. Due to this long time interval, the study area has experienced a high degree of change. While many terrestrial surfaces exist that conform to two or three of the above criteria, none conform to all five criteria and only one type of area conforms to four of the above five criteria, i.e. deep lakes. Deep lakes fail to meet all five criteria because they are the only suitable terrestrial surface and therefore the resulting PIF set would have a very low range in digital numbers. The manual selection of ground targets in this scene would therefore introduce a high degree of subjectivity.

   Five methods of radiometric correction which require little or no ancillary data or subjectivity were performed on the data. The first method of radiometric correction was a form of absolute radiometric correction available from PCI Geomatics named ATCOR, based on the Richter model (Richter, 1990; Franklin, 2001). The second method uses the same model and computes the impact of surrounding pixels on the target pixels radiance, referred to as the adjacency effect. The third method adjusted the origin and scale for all images to a common origin and common scale. The fourth method used the major axis of scatterplots from image pairs to determine PIFs and adjusts each image based on the mean and standard deviation of the PIFs. The fifth method performed median based directional scaling on the transformation from the fourth method.

   Quality control of radiometric correction is essential to obtaining a high quality final result. The best method for assessment of the fidelity of radiometric correction is through field measurements of reflectance. As already stated, data such as this are rarely available. Comparing the visual appearance of multitemporal imagery is the most common method of testing the fidelity of radiometric correction methods. While it is useful for large differences between images, this method is highly prone to subjectivity when differences are more subtle. Another common method is to compare the outputs of a simple classification performed on each radiometrically corrected image in the time series. As an example of this method, training statistics are derived from one image and used to classify all imagery. Similar classification accuracies signify quality radiometric correction (Song et al., 2001; Heo and Fitzhugh, 2000). This method has proven successful for coarse class structures, such as features that would not be expected to change over time, i.e. deciduous, coniferous, wetland, lake, etc. However, change would occur for more discrete class structures, such as classification of a forest by levels of biomass. Additionally, multitemporal classification similarity, while not qualitative, statistically compares the imagery after reducing the heterogeneity of the data. Therefore it is less capable of determining discrete differences between images, i.e. even if image b is 2 digital numbers brighter than image a overall, it is still possible for the two images to have high classification similarity. While classification creates datasets that are often of more use due to reduced heterogeneity, they also negate the continuous distribution of data which can be of more use to researchers (Cohen et al., 2001).

   Another method is to compare the root mean square error (RMSE) between images (Yang and Lo, 2000). This method can determine discrete differences if care is taken to ensure that the data used in the RMSE calculation have not experienced change, otherwise that change is incorporated as error. The assessment methods used in this study are similar to this method as they measure error, but specifically for areas that are either expected to experience no change, or are expected to experience linear change.

   Three forms of assessment of the radiometric correction methods were performed. First, edge detection was analyzed for each image pair. Second, statistics were derived from the imagery over old-growth spruce forests assumed to be in equilibrium in terms of reflectivity. Third, regenerating clearcuts were examined to determine the consistency of increasing NDVI over time.

   One purpose of this study is to develop a method of radiometric correction that requires no ancillary data and minimal user processing time while maintaining high radiometric consistency for multi-temporal satellite images. While some of the methods outlined in this paper may require significant user processing time, these methods could be easily automated and incorporated into image processing software. The second purpose of this study is to develop quantitative techniques for accuracy assessment which provide more detail of radiometric consistency than that provided by classification similarity across a satellite image time series.

Methods