Change Detection Techniques for ERS-1 SAR Data

Several techniques for detecting temporal changes in satellite synthetic aperture radar (SAR) imagery are compared using both theoretical predictions and spaceborne SAR data collected by the European first Remote Sensing Satellite, ERS-1. In a first set of techniques, changes are detected based on differences in the magnitude of the signal intensity between two dates. Ratioing of the multidate radar intensities is shown to be better adapted to the statistical characteristics of SAR data than subtracting, and works best when the number of looks is large. In a second set of techniques, changes are detected based on estimates of the temporal decorrelation of speckle. This method works best with one-look complex amplitude data, but can also be used with intensity data provided that the number of looks is small. The two techniques, ratioing of the signal intensity and decorrelation of speckle, are compared using actual SAR data collected by ERS-1. The results illustrate the viability as well as the complementary character of these techniques for detecting changes in the structural and dielectric properties of remotely sensed surfaces. Finally, change detection using the ratio method is applied to mosaics of repeat-pass S A R imagery to illustrate the potential of this method for monitoring applications at the regional scale.


I. INTRODUCTION
ETECTING temporal changes in the state of remotely D sensed natural surfaces by observing them at different times is one of the most important applications of Earthorbiting satellite sensors because they can provide multidate digital imagery with consistent image quality, at short intervals, on a global scale, and during complete seasonal cycles. A lot of experience has already been accumulated in exploring change detection techniques for visible and near infrared data collected by Landsat [l]. In the case of spaceborne synthetic aperture radar (SAR) imagery, change detection techniques have been developed for the temporal tracking of multiyear sea-ice floes using Seasat S A R observations [2], and rainfall events have been detected based on spatial radiometric variations in multidate Seasat SAR imagery [3]. Seasat SAR, however, did not provide calibrated radar measurements, and multidate observations were produced in limited quantity due to the short duration of the mission. Change detection techniques for spaceborne SAR data have not yet been fully explored.
With the launch of the European Space Agency (ESA) first Earth Remote Sensing Satellite (ERS-1) in July 1991, the Japanese first Earth Resources Satellite (J-ERS-1) in February 1992, and the Canadian RADARSAT in 1995, multidate SAR Manuscript received September 8, 1992;revised March 1, 1993. This work was supported by a contract with the National Aeronautics and Space Administration.
The authors are with the Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109. data are now generated and archived on a regular basis, at short intervals, over the entire planet, and for many years to come. This new generation of SAR sensors is characterized by an enhanced image quality (in particular a stable and reliable calibration of the data), and a significant increase in data volume, data rates, and duration of the mission. As a consequence, there is an important need for developing automated procedures of change detection for SAR data that can help human interpretors or subsequent computer algorithms analyze the data and relate the observed changes in radar backscatter with changes in the structural and dielectric properties of the remotely sensed areas.
Change detection techniques for SAR data can be divided into several categories, each corresponding to different image quality requirements. In a first category, changes are detected based on the temporal tracking of objects or stable image features of recognizable geometrical shape. Absolute calibration of the data is not required, but the data must be rectified from geometric distortions due to differences in imaging geometry or SAR processing parameters, and the accurate spatial registration of the multidate data is essential. Applications include sea-ice monitoring and motion tracking, monitoring of glaciers, landslides, and oceanic features. In a second category, changes are detected based on temporal differences in radar backscatter. The requirements are a stable calibration accuracy of the data, and an accurate spatial registration of the multidate data. Typical applications include monitoring of crops, volcanic activity, snow extent and conditions, glacial melt, soil moisture, and vegetation water content [4]. Finally, change detection techniques may be applied to data collected by more than just one sensor, for example, combining Seasat SAR data collected in 1978 with J-ERS-1 SAR data collected in 1992.
In this paper, we limit ourselves to the study of SAR-specific change detection techniques based on detecting changes in radar backscatter from land surfaces using repeat-pass imagery. Two sets of techniques adapted to the statistical characteristics of SAR data are presented and compared using both theoretical predictions and actual data collected by ERS-1 SAR. The paper is organized as follows. Section I1 presents the ERS-1 mission, and summarizes the characteristics of repeatpass ERS-1 SAR imagery, their calibration, and their registration. Knowledge of these images characteristics strongly influences the selection of candidate change detection techniques. Section 111 presents change detection techniques based on differencing of the multidate radar intensities. Section IV discusses change detection techniques based on the temporal decorrelation of speckle. Section V shows several examples of change detection applications using ERS-1 SAR, and conclusions are given in Section VI.

REPEAT-PASS ERS-1 IMAGERY
ERS-1 currently follows a sun-synchronous orbit with an orbital inclination of 97", and had a 3-day repeat cycle from August to December 1991 during the commissioning phase, followed by a 35-day repeat cycle during the multidisciplinary phase started in January 1992. ERS-1 operates a SAR instrument at C-band frequency (A = 5.7 cm), W-polarization (that is, vertical transmit and vertical receive polarization), 23" look angle, 100-km swath width, and with a spatial resolution better than 30 m [5]. As ERS-1 does not carry an on-board SAR data recorder, S A R data are only acquired in view of an Earthbased receiving station. The ERS-1 SAR data presented in this paper are from NASA's Alaska SAR Facility (ASF), at the University of Alaska, Fairbanks, AK [6].
During the 1991 commissioning phase, ERS-1 followed very accurate satellite orbits, with a 3-day repeat cycle which for some parts of the world, including Alaska, provided exact repeat coverage every 3 days. Within the ASF receiving station mask, ERS-1 overflew the same areas, at the same time of the day, with the same imaging geometry. The incidence angle of the radar signal onto the remotely sensed area was repeated within fractions of a degree, and the geometric distorsions and radiometric, calibration errors introduced in the imagery due to topographic variations were also exactly repeated. Repeatpass slant-range ERS-1 SAR images are therefore directly usable for change detection and monitoring applications, and there is no need for geometric rectification of the data and for projections and resampling to a common Earth-fixed grid. Registration of multidate repeat-pass slant-range images is a straightforward process limited to the determination of an azimuth shift (due to a different processing start time for the first range line, and slight differences in time of overpass of the satellite between two passes), and a residual slant-range shift (due to differences in SAR processing parameters and slight changes in the trajectory of the satellite), typically about 1 or 2 pixel elements, Le., 30 m. To determine this pixel offset with subpixel accuracy, we compute the correlation peak of the signal intensity between the two SAR images. This method is reliable even in the presence of large changes in the signal intensity because numerous and stable textural features are always present in the imagery.
Mosaicking of SAR data acquired along the same flight line or transect is more difficult because the data are corrupted by geometric distortions that vary along track due to relative changes in altitude, Earth velocity, and attitude of the ERS-1 satellite. The azimuth pixel spacing is higher in near-range by 0.04% and 0.11% due to differential Earth circumference and velocity, respectively. The slant-range pixel spacing varies between 12.493 and 12.507 m between the first and last record [7]. Since consecutive images along track are processed with slightly different processing parameters, the geometric distortions in their overlapping segments are not identical. Typically, two consecutive images overlap by 134 lines (over a total of 1024 lines), and the correlation peak varies by about f0.25 pixels in slant-range, as a result of differential geometric distortions. We computed the correlation peak of the signal intensity at midrange as a compromise, but registration errors are visible at the transition boundary between two scenes. These registration errors are difficult to identify by a human eye from a S A R image display, but they appear distinctly in change detection maps because the process of change detection is sensitive to subpixel registration errors, which are difficult to avoid in mountainous terrain where large transitions in pixel brightness occur within a single pixel and geometric distorsions are important.
One of the most important characteristics of ERS-1 SAR is the stability of its radar gain. The standard error of the radar gain is currently less than 0.4 dB according to ESA's active external calibration devices [8], and in good agreement with the predicted stability of 0.33 dB measured during ground testing [5]. The remarkable stability of ERS-1 radar gain, combined with the great accuracy of its repeat-orbit cycle and its short revisit interval, render ERS-1 repeat-pass SAR imagery very suitable to change detection applications. Since the calibration factors (that is, absolute radar gain, and antenna pattern removal) are already included in the ASF/SAR processor [9], no calibration work was required in this study.
In the next sections, we describe and compare several change detection techniques for ERS-1 repeat-pass SAR data.

CHANGE DETECTION USING THE DIFFERENCE
Differencing and ratioing are well-known techniques for change detection [l]. In differencing, changes in radar backscatter are measured by subtracting the intensity values, pixel per pixel, between two dates; in ratioing, changes are measured by dividing the intensity values, pixel per pixel, and are conveniently expressed in decibels (that is, taking ten times the logarithm in base ten of the ratio of the intensities). In this section, the two methods are compared based on their statistics and probability of error.
The multilook SAR intensities are assumed to be gamma distributed as where (Io) is the mean intensity of an homogeneous region at time to, and N is the equivalent number of looks of the SAR data. Since a premise of change detection is that radar intensities do change between time t o and tl, the intensities IO and 11 acquired, respectively, at times to and tl over the same resolution cell are assumed to be mathematically independent.
The joint distribution of 10 and I1 is therefore the product of their marginal distributions given in (1).
The distribution of the difference (d = I1 -lo) of the signal intensities is derived after a change in variables by integrating the joint distribution over Io, leading to [lo] The mean of the difference and the mean of the square difference are where E ( x ) denotes the expected value of the random variable x. The distribution of the difference given in (2) depends on both the relative change ( ( l 1 ) / ( 1 0 ) ) , and on a reference intensity level, Le., either (lo) or (11). As a consequence, changes will not be detected in the same fashion in high intensity regions compared to low intensity regions, which is not a desired effect. When the intensities (10) and ( 1 1 ) are multiplied by a constant factor a, (3) shows that both ( E ( d ) ) 2 and E ( d 2 ) are multiplied by a 2 , which means that the variance of the distribution of the difference increases with the intensity level. Since the probability of error in detecting changes increases with the variance of the distribution, this result shows that the difference method will produce more errors in the high intensity regions of the image than in the low intensity regions, so that the difference method is not adapted to the statistics of S A R images.
Similarly, the distribution of the ratio (T = 11/10) is derived after a change in variables by integrating the joint distribution over IO leading to [lo]-[I21 where F = ( l 1 ) / ( 1 0 ) . The mean and standard deviation of the ratio are n - In contrast to the distribution of the difference, the distribution of the ratio given in (4) only depends on the relative change F in average intensity between two dates, and detecting changes in the S A R imagery does not depend on the intensity level of the pixels. To evaluate in a more quantitative fashion the error produced by the ratio method, we now compute its probability of error as a function of the change in radar intensity between two dates.
If we assume a threshold TO to decide on whether to classify a given change in radar backscatter between class A characterized by a change T A = (I~)/(lo) or class B characterized by a change T E = (IB)/(~o) (Le., T > TO is classified as class B, and T 5 TO is classified as class A, and assuming T B > T A ) , the class A probability of error is given by Assuming equal a priori probabilities for classes A and B, the threshold decision is found using leading to which is independent of the number of looks N . The total probability of error is which is found to be equal to A plot of P E as a function of the change in radar backscatter intensity, Ano = 20 loglo(X), expressed in decibels, is shown in Fig. 1 for different values of N . The results show that P E decreases slowly with an increasing number of looks N . To detect changes in radar intensity less than 2 dB with a confidence level better than 90% (that is PE = 10 %), N must be greater than 64 looks, which means a large number of independent S A R samples. Detecting changes in radar backscatter of the order of one decibel requires a number of looks larger than a hundred. Hence, the ratio method is quite sensitive to the presence of image speckle in the sense that speckle patterns that are difficult to detect by a human eye in SAR images when the number of looks is large will still be visible in ratio images.
Another important reason for selecting the ratio method instead of the difference method is that the ratio method is very robust to calibration errors whereas the difference method is not. Radiometric errors are usually introduced in the imagery during SAR processing by assuming a flat Earth model, resulting in two errors: 1) an error in the computation of the scattering element size due to errors in the true incidence angle of the electromagnetic signal onto the surface; and 2) an error in the antenna pattern removal due to errors in the true depression angle of the signal onto the reflecting surface [13]. Both errors are multiplicative factors to the total radar intensity. These factors are eliminated when computing the ratio image because calibration errors are exactly reproduced in repeat-pass imagery. In contrast, in the difference method, radiometric errors directly modulate the difference of the radar intensities, typically yielding higher changes in very hilly terrain (e.g., slopes facing the radar) than in rather flat areas for the same type of surface, which is not a desired effect.
In most applications, it is necessary to threshold the ratio image to identify image pixels that change by more than fz dB, where z is a real number. Selecting threshold values depends on the application and on the nature and magnitude of the expected changes, and can generally be viewed as an optimization problem with a number of constraints. We will not address this topic in this paper. Given threshold values, image pixels are classified into classes, each of which corresponds to a well defined change in magnitude of the radar intensity. Equation (9) shows that an optimum threshold for classifying a pixel into class A or class B is given in decibels by T = ~(lO1oglO(TA) + lolOglo(TB)), independent of the number of looks N, because this operation corresponds to a maximum likelihood classification of the ratio image into a number of classes of changes. To improve the results, one possibility is to include the a priori probabilities of the classes in the classification process and perform a maximum a posteriori classification of the data. An example of maximum a posteriori (MAP) classifier for change detection applications was proposed in [ 121. The a priori distribution of the classes is modeled using a Markov random field, and is combined with the a priori distribution of the ratios (4) to yield an expression for the a posteriori distribution of the classes given the ratios over the entire image array. Bayes' optimal classification of the ratio image maximizes the a posteriori distribution of the classes over the entire scene. Although the exact solution to this optimization problem is extremely difficult to compute, approximate solutions can be obtained. Several techniques are possible, depending on computational cost versus quality of the solution, as discussed in [12], [14]. An example of MAP classification using ERS-1 SAR data is given in Section V.
Other techniques have been proposed to detect changes in digital satellite imagery, including multidate classification of the data, and principal component analysis [l]. Results obtained with Landsat data indicate that these more sophisticated techniques do not perform better than a simple differencing of the images. We did not test them on SAR data. In the next section, we will present a novel SAR-specific change detection technique based on the temporal decorrelation of speckle. In principle, temporal decorrelation of speckle may arise even when no changes in the radar intensity are detected, so that this change detection technique may detect small changes in radar backscatter that could not be detected using the ratio because of limitations imposed by image speckle (Fig. 1) or by the radar gain stability.

THE DECORRELATION OF SPECKLE
According to the random walk model of speckle [15], decorrelation of speckle may arise from a redistribution of the scatterers within a resolution cell, either as a result of random displacements of the scatterers due to external disturbances, or because the scatterers change totally in nature and distribution between the two dates. Hence, the temporal decorrelation of speckle provides information about possible changes in the surface structural and dielectric properties. The detected changes are independent of the calibration accuracy of the data.
The decorrelation of speckle is not measured directly by SAR, because there are other sources of signal decorrelation in repeat-pass SAR imagery. These sources are: 1) a decorrelation due to thermal noise; 2) a spatial baseline decorrelation; and 3) a temporal decorrelation of speckle [16] . The total correlation coefficient is modeled as the product of the correlation coefficients from these independent sources of decorrelation as P a = Pn PB P s where SNR is the signal-to-noise ratio, B is the antenna baseline separation (assumed to be in the horizontal direction), R, is the spatial resolution in azimuth, t9 is the look angle, X is the radar wavelength, R is the range distance to the radar, pa is the magnitude correlation coefficient of the SAR complex amplitudes (that is, amplitudes and phases of the signal expressed as complex numbers), pn is the thermal noise decorrelation coefficient, PB is the spatial baseline decorrelation coefficient, and ps is the temporal decorrelation of speckle. Knowing the thermal noise power level, and the antenna baseline separation between repeat-passes, one can estimate ps from pa. In the case of ERS-1 SAR, R, = 30 m, X = 5.7 cm, t9 = 23", and R = 830 km so that the effect of the spatial baseline decorrelation is significant when B 2 460 m (that is, p~ 5 0.5). The effect of thermal noise is significant for SNRS 1 dB (for which pn 5 0.6), which means radar intensities close to the system noise floor. Equation (14) shows that p~ does not vary significantly within a SAR scene, because R and cos(8) vary by 10% and 5% across swath, respectively, and R, and X are constant. Hence, if a large value of pa is detected in an area of large SNR, that value will set a lower limit on p~, and practically insure that the effect of the spatial baseline decorrelation can be ignored in the analysis of spatial changes in pa.
We estimate the temporal decorrelation of the signal using where 00 and 01 are the S A R complex amplitudes measured, respectively, at time t o and tl. A large number of samples is required to obtain an accurate estimate of pa because of speckle. When speckle is highly correlated between the two dates, the variance to mean square ratio of the estimate of the spatial correlation of one-look complex amplitudes is given in [14]. In this study, we used one-look SAR complex data, 4096 pixels in range by 12 000 pixels in azimuth, and added the different terms required to estimate the local averages in Equation (15) by taking 4 pixels in range and 20 pixels in azimuth to obtain an 80-looks image of pa that is 512 pixels by 600 pixels in size with nearly square pixels. As neighboring pixel samples are not rigorously independent because pixel spacing is usually selected equal to half the width of the radar impulse response, the effective number of looks N is only 55 when estimated correctly using the equation where N , and N, are, respectively, the number of pixel samples in azimuth and in range, p u o ( i , j ) is the spatial correlation coefficient of the single-look complex data acquired at time to, i is the displacement in pixel in azimuth, and j is the displacement in pixel in range. When N = 55, the variance to mean square ratio of pa varies between 13% for I pa I= 0.2 and 2% for I pa I= 0.9, which is reasonably low. Note that pa is independent of spatial variations in radar backscatter cross-section of the surface as discussed in [14].
Disadvantages of the decorrelation technique are the volume of data to be manipulated, the computational cost, and the difficulty to generate one-look complex amplitude S A R data products for a large community of users as these data are intermediate products that are not saved after S A R processing. An alternative is to consider the decorrelation of the signal intensity I =I a l2 between two dates. The correlation coefficient of the intensity, noted P I , is estimated using The phase of the radar return is not used. In the absence of a textural modulation of the SAR signal due to the natural spatial variability of the radar backscatter cross section of the reflecting surface P I =I Pa l2 (18) using a fundamental property of circular symmetric Gaussian processes. Equation (18) shows that taking the square root of P I yields an estimate of 1 pa 1. This equality does not hold, however, if textural variations of the radar backscatter from the surface are present and the number of looks is much larger than one. This effect is explained using a simple multiplicative model of the S A R intensity where T; is a normalized ((Ti) = 1) texture random variable that modulates the mean intensity return (Ii) of an homogeneous area at time ti, and Si is a normalized ( ( S i ) = 1) speckle random variable with a mean squared to variance ratio equal to the equivalent number of looks N of the data. Using (18) and (19), we obtain assuming that TO and TI have the same variance a$. When texture is not present, o$ is zero, and PI =( pa 12. When texture is present, and the number of looks N is very large, PI = p~, and texture dominates the correlation of the signal, which is not surprising since image speckle is not a dominant source of spatial variability of the signal for large N . As a consequence, the temporal decorrelation of speckle from intensity data can only be detected for N small, in practice with N less than 4. Examples are shown in the next section.

V. EXAMPLES USING ERS-1 SAR DATA
Several examples are shown in this section using multidate data from ERS-1 SAR. The objective here is not to interpret physically the observed changes in radar backscatter but to illustrate the use of change detection techniques with actual spacebome SAR data. Understanding the reasons for detected changes is a topic of ongoing research.  Table I. Air temperatures did not change significantly between the two dates and remained at about +9"C. Rain occurred one day prior to the imaging of the first scene, and one day prior to the imaging of the second scene. Extrapolation of these weather conditions indicate that air temperatures were probably below zero in mountainous areas above 900 m based on a dry lapse adiabatic rate of loC/10O m, and local rain storms probably covered the mountain tops with snow. We do not have weather data and surface observations within the scene itself.
An image of the magnitude of the complex correlation coefficient pa is shown in Fig. 2(d)  and 1; (c) Ratio image Aao between -2 and +2 dB; (d) Correlation coefficient pa of one-look complex amplitude data between 0 and 1. (bHd) are overlaid on the SAR image, which means that the amplitude of the SAR signal is used as a saturation factor between 0.5 and 1.0 on the red, blue, and green channels.
We overlaid the results of the change detection algorithms on the S A R data because topographic variations are totally eliminated in the change detection maps as predicted from the near perfect repeat-pass of ERS-1 over the same areas, and areas of change are difficult to position in the S A R imagery by a human interpreter without identifiable reference points. To construct this data overlay, we modulated the saturation of the red, green, and blue channels by a factor between 0.5 and l., depending on the magnitude of the S A R signal. In many areas of Fig. 2(d), speckle is highly correlated between the two dates. The maximum value of p a is 0.964 in an area of high SNR, indicating that p~ 2 0.964 and does not significantly modulate the results. Low values of pa are seen in valley bottoms in the lower right portion of the figure, and in nearly all mountain tops which include active volcanoes. Correlation of speckle decreases markedly above a certain altitude, perhaps revealing the contours of the tree line, of the snow line, or both. The ratio image shown in Fig. 2 Fig. 3(c) indicates a large increase in radar backscatter over the glaciers and mountain tops. The correlation coefficient of speckle computed from one-look intensity data is shown in Fig. 3(b), and is in good agreement with the results shown in Fig. 3(d), indicating that the decorrelation of speckle can be studied using one-look intensity data. Fig. 3(e) shows an image of the decorrelation of speckle computed using 4-look intensity SAR data. The effect of texture is more pronounced, (MINIMUM-MAXIMUM) RECORDED AT TALKEETNA AIRPORT radar backscatter larger than 2 dB in mountain tops and valley bottoms. In general, changes in uo are detected in areas where pa is low; and, no changes in 00 are found where pa is high. Both methods reveal important changes in the physical and electrical properties of the surface at the higher elevations, yet the decorrelation of speckle detects changes in the lower right portion of the figure that cannot be identified in the ratio image. The correlation of speckle computed using one-look S A R intensity data and 4 pixels in range by 20 pixels in azimuth, is shown in Fig. 2(b), revealing a good correspondence between fi and pa, although the contrast between regions is better in Fig. 2(b). Image features such as ridge tops and slopes facing the radar yield erroneously large values of fi in Fig. 2(b), because of texture.

(c) reveals changes in
Another example is shown in Fig. 3 (Table 11). Air temperatures at night remained close to O°C the whole period, suggesting freezing conditions at elevations higher than 200 m. Airtemperatures during the day were warmer at Talkeetna Airport, but probably below O'C above 1200 m. The higher portions of the glaciers seen in the top part of Fig. 3(a) are at elevations higher than 2000 m.
The complex correlation map shown in Fig. 3(d) indicate a large correlation of speckle between the two dates, even at the higher elevations. Correlation of speckle is low over the glaciers. Random motion of scatterers within one resolution cell results in a complete decorrelation of speckle after rms motion of 2-3 cm at C-band [16]. However, glacier motion may more likely result in a nearly uniform translation of all scatterers within a resolution cell instead, which would result in total decorrelation if the displacements are of the order of one resolution cell or 30 m. Glaciers do not move at such high rates in this region, and the decorrelation of speckle is probably due to changes in their surface physical and electrical properties. This is confirmed by examining the changes in radar backscatter over these same areas. The ratio image shown and the contrast between regions where speckle is correlated and those where speckle is decorrelated is reduced. Yet, areas where speckle is highly correlated are still visible, illustrating that 4-look intensity data can provide information about the degree of correlation of speckle. This result is of interest to SAR interferometry studies because it shows that multilook SAR intensity data could be used to select image pairs for which speckle is highly correlated and interferometric mapping is possible, before correlating and analyzing one-look S A R complex amplitude data, thereby offering a large data volume reduction in processing of the data.
In principle, when the structural and dielectric properties of the reflecting surface change, the correlation of speckle decreases, and the intensity of the radar returns changes. As mentioned earlier, there can also be areas where speckle decorrelates, yet no change in radar backscatter is detected by the SAR system. As an example of practical situation, a forest under varying wind conditions could undergo no change in radar intensity if the soil and vegetation moisture conditions are unchanged, yet temporal decorrelation of speckle could arise from a random motion of branches which efficiently scatter the radar signals at C-band frequency. Surprisingly, we also detected areas where speckle is highly correlated but the radar intensity changes by several decibels. One such area is identified in Fig. 3(a) with a white circle, with changes in radar backscatter larger than 2 dB and a correlation coefficient Pa larger than 0.8. A close inspection of both Figs. 3 and 4 reveal that other areas exist where a significant change in radar backscatter is detected and yet pa is greater than 0.5 (height information can still be retrieved when pa M 0.5 according to [16]). A value of pa greater than 0.8 suggests that the distribution of scatterers within each resolution cell has not changed, while a 2 dB change in radar backscatter indicates that the scattering properties of the scatterers have been modified. We do not have surface observations within those areas to help propose a definite explanation. One practical surface that could result in such scattering properties could be a recently burned forest. Scattering is there dominated by treetrunk/ground interactions (the branches have been burnt), and a change in the ground moisture conditions andlor in the treetrunk water content would change the magnitude of the radar returns without any detectable change in the distributions of the scatterers (i.e., the tree trunks), hence resulting in highly correlated speckle.
These first examples illustrate the practicality of ERS-1 repeat-pass SAR data for change detection applications, and demonstrate that numerous changes in the surface structural and electrical properties can be detected using the ratio method or the method based on the decorrelation of speckle. Comparison of these different change detection techniques  In the next two examples, we examine larger changes in the surface structural and electrical properties detected using the ratio method. We did not analyze the decorrelation of speckle in these examples because image speckle is not expected to remain correlated in areas of large changes separated by several weeks. flying from bottom to top, looking to its right. The scene comprises a mixture of various types of tree species intermixed with treeless areas of short vegetation of shrubs and sedges on a moss layer (see [17] for more details). The Tanana River is visible going across the center of the scene from east to west, corresponding to low pixel brightness values.
Weather data collected at this site by the people from the Institute of Northern Forestry in Fairbanks, Alaska, reveal air temperatures of about f4.8"C during the day, and -8.5"C at night, and a soil in the process of thawing (Tsoil = -0.loC during the day) on Apr. 25, 1992. Air-temperatures were warm (T = +19.0°C) with thawed soil (Tsoil = +l.O°C during the day and above zero at night) on May 30, 1992. Changes in radar backscatter larger than 6 dB are detected in the imagery. The color coded ratio image shown in Fig. 4(c) shows areas where the radar backscatter increased by more than 3 dB (yellow), decreased by 3 dB (red), or did not change by more than 3 dB (blue), overlaid on the SAR image acquired in April.
Comparison of the observed changes in radar backscatter with vegetation types reveals that areas colored yellow in Fig. 4(c) are treeless areas (apart from a few points along the river), areas colored red correspond to the river (the disappearance of the rough water ice probably caused the observed decrease in radar backscatter values) and in smaller areas of standing water, and areas colored blue are the forests. For reasons explained earlier on in this paper, the color ratio image shown in Fig. 4(c) is also a maximum likelihood classification map of the area into three classes of changes since the color coding is reduced to three colors only. Fig. 4(d) is an illustration of subsequent improvements in classification quality that can be obtained using a maximum a posteriori classifier [12]. The ratio image was classified into three classes of change (AUO 2 f 3 dB, Aao 5 -3 dB, and I Auo (< 3 dB), and the results are using the same color coding as in Fig. 4(c). The different areas outlined in Fig. 4(c) are better defined in Fig. 4(d). This result indicates that a maximum a posteriori classifier provides a worthwhile improvement in segmentation quality for change detection applications compared to a simple thresholding of the ratioed data, which is consistent with the fact that ratio images are very sensitive to image speckle. Finally, the last example illustrates the potential of the ratio method for detecting changes in radar backscatter at the regional scale. We mosaicked together three S A R images collected during a descending pass of ERS-1 that intercepted with the city of Tanana, 200 km west of Fairbanks, AK. One transect was acquired on September 29, 1991 (Rev. # 1072), while the other was acquired 18 days later, on October 17, 1991(Rev. # 1330. The mosaic shown in Fig. 5(a) is 100 km by 400 km in size. The north direction is to the top, with ERS-1 flying from top to bottom, looking to its right. As in the previous cases, radiometric distortions due to topography are exactly eliminated in the ratio image, due to the good registration of the data, and the exact repeat-pass of ERS-1 SAR (a small radiometric modulation of the ratio image was introduced using the amplitude data from one of the SAR transect to help reference areas of change in the radar imagery, and are not due to registration errors, as in all the examples shown in this paper). Air temperatures recorded at Tanana Airport (located in the top portion of the figure at 65O.ON and 15lo.2W) during that period went from f5"C to -9°C. Similar values were reported at Lake Minchumina Airport (located at 63.54"N, 152.16"C, 213-m elevation). These changes in radar backscatter have been interpreted as resulting from a decrease in the dielectric constant of the soil and vegetation with freezing, which results in a decrease in radar backscatter of the soil and vegetation of approximately 3 dB, nearly independent of landscape, except for areas of standing water (where radar backscatter values may either increase or decrease), and areas that are already in frozen conditions such as high mountain tops where subzero airtemperatures are recorded much earlier in the season [18], [19]. Areas colored red in the ratio image shown in Fig.  5(b) correspond to areas where the landscape (Le., the soil and vegetation as a single medium) is frozen, whereas areas colored blue froze earlier (case of the mountain range present (4 (d )  Fig. 4(c) using a maximum a posteriori classifier. Pixels colored yellow correspond to Aao 2 $3 dB, pixels colored red correspond to Aao 5 -3 dB, and pixels colored blue correspond to I AUO < 3dB I. The classification map is overlaid on the S A R amplitude data of Fig. 4(a). The saturation factor for the S A R overlay is between 0.5 and 1.2.
in the upper portion of the image in Fig. 5(a)), or did not (a) (b) freeze (case of the areas in the lower portion of the image in ~i~. s (~) ) because air temperatures were not cold enough yet to yield a complete freezing of the soil and vegetation. The radar backscatter from the Tanana River and Lake Minchumina increased with freezing because of the formation of a rough icelwater interface with freezing.

VI. CONCLUSIONS
Several techniques for change detection have been presented and compared based on their probability of error and on results obtained using repeat-pass ERS-1 SAR data. The method based on the temporal decorrelation of speckle works best with one-look SAR complex amplitude data, but can also be used with multilook intensity data provided that the number of looks is small. Ratioing of the radar intensities works best with multilook intensity data. Areas of change identified by both methods under the same environment do not always agree. For instance, speckle may remain highly correlated although significant changes in radar backscatter are detected. The decorrelation method helps detects changes in the position of the scatterers whereas the ratio method provides information about the magnitude of the observed changes. Both methods combined enhance our characterization of the observed changes. For instance, if we know that the position of the scatterers is unchanged, then any change in radar backscatter will be related to changes in the electrical properties of the surface and not its structure. Indeed, if both position and scattering properties of the scatterers change, the interpretation is more difficult, but the large complementarity of these two techniques justifies that both of them should be used in detailed monitoring studies. The initial results shown in this paper illustrate the great potential of ERS-1 for change detection applications. Much work remains to be done to interpret the observed changes in terms of changes of specific parameters controlling the structure and dielectric properties of the landscape, such as surface roughness, biomass, soil and vegetation moisture conditions, or snow depth and water equivalent to mention a few. Although the radar backscatter from natural surfaces depends on all these parameters, plus topography, it is also clear that under certain environmental conditions and over certain time scales, several of these variables may change slowly with time, if at all, so that temporal changes in radar backscatter could be directly related to changes in only one or two parameters. A n example is the monitoring of freeze-thaw transitions along north-south Alaskan transects using ERS-1 SAR data, independent of topography and vegetation types [19].