The integrated approach to the geodetic determination of crustal deformation Athanasios Dermanis Department of Geodesy and Surveying University of Thessaloniki, Greece 1. Introduction geodesy to what could be justly called integrated geophysics. The importance of the study of crustal deformation needs hardly to be stressed, especially in those parts of the world where tectonic activity is the cause of earthquakes that affect the lives of millions of people. Geodesy, being the science which is concerned with the determination of the shape and gravity field of the earth, can contribute to the assessment of crustal motion in two ways: from the determination of the deformation of the surface of the earth and from the determination of the temporal variation of the gravity field which reflects also the redistribution of internal masses. The basic techniques described here are based in the work of Rossikopoulos (1986), Dermanis and Rossikopoulos (1988) and Dermanis and Rossikopoulos (1991). A review of work in classical integrated geodesy, following the pioneering ideas of T. Krarup (Eeg and Krarup, 1975), can be found in Hein (1986). First applications of relevant concepts in the study of deformation have been presented by Dermanis et al. (1981), Bencini et al. (1982) for horizontal networks, Kanngieser (1983) for leveling networks. Classical geodetic techniques treat such information independently by geometric surveys (horizontal networks and leveling) and gravimetric ones. The standard approaches suffer furthermore from the somewhat arbitrary separation of the horizontal from the vertical part, as well as, from the use of numerous modeling approximations and data processing strategies whose validity is questionable in cases where observations of the highest possible accuracy are analyzed for the determination of physical processes with magnitudes not always well above the level of observational noise. 2. The description of crustal deformation Crustal deformation is usually described by separating the horizontal (planar) from the vertical component of displacement, following the traditional separation of geodetic work in leveling and horizontal networks. In addition to the 2-dimensional "horizontal", crustal deformation can be described in a less fictitious 3-dimensional way, corresponding to the actual deformation of the 3-dimensional earth, by either combining the results from leveling and horizontal networks, or by direct use of 3-dimensional networks. However, the fact that relevant information is limited to points on the earth surface, which is a 2-dimensional medium though not flat or horizontal, makes the study of plane deformation still attractive and perhaps more convenient. In this approach the surface of the earth is mapped on a plane (or a sphere or ellipsoid for larger areas) and the deformation of the imaginary surface produced point-wise by such a "vertical projection" mapping is replacing the actual deformation of the 3-dimensional surface. Integrated geodesy has been introduced for the unified and rigorous adjustment of observations with both geometric and gravimetric information, using clearly defined optimality criteria and precise mathematical models. In recent years integrated geodesy has been considered in its more abstract form as a method for the adjustment of observations depending not only on discrete parameters but also on unknown functions. This generalization allowed its proper application to the study of crustal estimation, utilizing all available geodetic observations depending not only on the (static) gravity field of the earth (classical integrated geodesy) but on a gravity potential which is a function of both time and position, as well as, on network point positions which are functions of time. The observations are, in this context, discrete not only with respect to space but also with respect to time, thus allowing a departure from the classical scheme of comparison between two or more different epochs. With respect to the time component, two possible approaches can be distinguished. The first is the comparison of the change of the shape of the earth crust between two epochs t and t ′ . Material points are identified by their position in one of the two epochs usually the first one t (Lagrangean description). Deformation is then described by means of invariants of the strain tensor, which describes for each point the alteration of shape in a small (infinitesimal) neighborhood of the point. Invariance is an essential property, since deformation is a concept independent of any particular choice of coordinate systems used for the shake of mathematical convenience. The required information is the shape of the deforming medium in the two epochs, which means that the position of every point should be known at both epochs with respect to some arbitrary frame of refer- It is the aim of this work to describe the principles, techniques and some special applications of immediate interest of integrated four-dimensional geodesy. Of course the concept of integration can be expanded beyond the limits of geodesy with the, more or less straightforward, inclusion of non-geodetic observations, such as those from strain-meters, tilt-meters, seismic sounding, etc., leading from integrated 1 ence, i.e. continuous spatial information is needed for this type of description. 3. The concept of integrated geodesy as applied to the study of crustal deformation The second approach refers not to the comparison of any two epochs but to the "rate of change of shape" at any particular epoch t. Deformation is in this case described by invariants of the time derivative of the strain tensor (strainrate tensor). The required information is not only continuous in space as in the previous case, but it should also be continuous in time, i.e. the positions of material points should be known for all epochs at least in a small time-neighborhood of the particular epoch of interest. The positions must refer to a reference frame which is defined arbitrarily for every epoch with the essential restriction of being time-wise smooth, a concept referred to as "framing" (Truesdell, 1977). In classical continuum mechanics there exists always an undeformable environment (the laboratory) which provides the means for the definition of a time independent arbitrary reference frame for the study of the deformation of material bodies. On the contrary geodetic work is confronted with a "sea" of moving points and a reference frame must be defined arbitrarily at every epoch. These epochwise frame definitions constitute a framing and the study of crustal deformation must be addressed to quantities invariant with respect to changes of framings, unlike the case of classical "tensorial" invariants where invariance refers to a change of reference common for all points. This is indeed a critical point not always completely understood from the viewpoint of classical physics. A study of the invariance of such parameters, which are related to the concept of their estimability from framing invariant geodetic observables has been carried out by Dermanis (1981), Dermanis (1985a) and Dermanis and Grafarend (1992). Integrated geodesy has originally been introduced by Krarup (Eeg and Krarup, 1975) as a concept for the integrated analysis of all data related to the gravity field of the earth, even those in traditional "geometric" geodetic observations which relate to gravity through the local direction of the vertical. An essential aspect of the approach was the incorporation of the solution not only for the gravity related parameters (signals) but also for the unknown gravity potential function itself, which in turn allows the determination of the values of any gravity related parameters (new signals) other than those present in the observations. This in fact extended the classical least squares adjustment with a solution to the interpolation problem for the gravity field. This problem was solved by Krarup in his celebrated "contribution" (Krarup, 1969) by the now famous "collocation" method. The resulting adjustment algorithm is sometimes referred to as "least-squares collocation" (Moritz, 1973). There are two independent but equivalent interpretations of the algorithm for integrated geodesy. In the deterministic interpretation the least squares principle for the observation errors is extended to include the square of the norm of the unknown function which is considered to belong to an appropriate class of functions (a Hilbert space with reproducing kernel). The solution is attaining a balance between the requirement for small errors and a smooth interpolated function. In the statistical interpretation the unknown function is modeled as a stochastic process (random function) and the proper model for the adjustment is the linear mixed model with both deterministic (e.g., coordinates) and random parameters (signals). The determination of new signals is achieved by the well-known solution of the prediction problem (determination of outcomes of random variables from the known outcomes of other random variables). Parameters are estimated and signals are predicted in an optimal way by minimizing the mean square error (mean over all possible outcomes) of such estimates or predictions within the class of the linear and unbiased ones. More details on these two interpretations can be found in Krarup (1969), Lauritzen (1973), Dermanis (1976, 1978, 1983), Moritz (1978, 1980), 6DQV Although continuous spatial or continuous spatial-temporal shape information is required, according to the type of deformation description followed, geodetic observations are discrete both in space and time. Even in the two epoch description approach, the possibility to have observations referring to the two epochs only (in fact in two small time periods around these epochs within which deformation can be neglected) can be realized only for horizontal networks and not for the time consuming process of leveling. The contrast between the discrete character of the information provided by geodetic observations and the required continuous information required for deformation description, makes it clear that an appropriate approach to the analysis of geodetic data related to crustal motion detection, must not be confined to the classical adjustment techniques for parameter estimation but should in some way incorporate a solution to the implicit interpolation problem. Interpolation is not only a mathematical necessity for the determination of deformation at the points of observation, but also for obtaining information at all points of a tectonic active area, a matter of obvious geophysical importance. Although the deterministic interpretation is easier to accept (there is hardly anything random about the gravity field of the earth), the stochastic-statistical exposition has gained more popularity, because it does not require the rather deep mathematical foundation of the deterministic one. Integrated geodesy becomes a more general and more powerful technique by the process of abstraction: From a method for the analysis of data related to the unknown gravity potential function, it can be conceived as a method for analyzing any data related to one or even more unknown functions. It is in this framework that it qualifies as a natural tool for the study of crustal deformations. It is for these reasons that the concept of integrated geodesy provides, under proper extension and modification, the most valuable tool for the integrated analysis of all data relevant to crustal motion, not only the purely geometric ones, but also those related to the gravity field of the earth, whose temporal variation is directly connected to crustal motion. The unknown functions in crustal deformation analysis are the gravity potential and the displacements of material points, both being functions of space, as well as, of time. The relevance of the gravity field is not only due to the fact that some geodetic observations are sensitive to it, but mainly to that crustal deformation causes changes of the local gravity vector in two ways: from the motion of the 2 points within the field domain and directly from the change of the field itself due to mass redistribution. is used, where φ1 , φ 2 , …, φ k , are known functions (base functions). 4. Fundamentals of integrated geodesy The (purely) stochastic approach: In this case the mean function E{ f } of f is assumed to be completely known, 4.1. Function modeling approaches i.e. coinciding with the normal function f 0 . f = f0 + fs Of fundamental importance in the integrated geodesy approach is the modeling of the unknown functions, which give rise to the specific signals (in the sense of discrete function-dependent parameters) which appear in the mathematical models for the observable quantities. Every unknown function f may be in general considered to consist of the following parts: f = f 0 + δf = f 0 + m + f s The stochastic part f s has therefore zero mean, E{ f s }= 0 , and a finite covariance function C ( P, Q ) = E{ f s ( P) f s (Q )} (4) which is assumed to be known, where P and Q are any two of the function arguments. (1) An equivalent model is to consider f s as a linear combina- where f 0 is the known (normal) part of the function, δf is the unknown part which in turn consists of two parts, the deterministic part m = E{δf } , which is the "trend" of the unknown part and the zero-mean stochastic part (3) tion of known base functions ψ i f s = b1ψ 1 + b 2ψ 2 ++ biψ i + fs ( E{ f s }= 0 ). It is not necessary to include all the above three parts in the model. It is not even necessary to take into account the dependence of the signals on an unknown function. In the latter case the signals are treated as purely independent deterministic unknown parameters along with the usual geodetic parameters. This is in fact the approach in the three-dimensional network adjustment according to the Bruns' polyhedron concept (Bruns, 1878, Heiskanen and Moritz, 1967). (5) where b1 , b2 , …, bi , … are uncorrelated stochastic parameters with E{bi b j }= 0 for i ≠ j , E{bi2 }=σ i2 (6) which leads to the same model as (4) through the equivalence relation (Dermanis, 1983) C ( P, Q ) = ∑σ i2ψ i ( P)ψ i (Q) . When the dependence of the signals on an unknown function is taken into account, different approaches may arise depending on which of the three parts f 0 , m , f s , are included in the function model (Dermanis & Rossikopoulos, 1988): Note that even an infinite number of terms are allowed in the representation (5) provided that the sum in (7) is finite. The classical method of reduced observations: In this case it is assumed that f = f 0 = known , by taking advantage of an The combined (analytical-stochastic) approach: This is the more general case f = f 0 + m + f s , where the trend m is usu- existing estimate f 0 of f , and the signals become also known parameters. Their effect can be subtracted from the observations to produce "reduced observations". An example is the reduction of theodolite observations from the local vertical to the local normal to the ellipsoid, using existing estimates of the deflections of the vertical. The result is a purely geometric three-dimensional adjustment, or even a two-dimensional one on the ellipsoid by using only horizontal observations further reduced from the actual point to its projection on the reference ellipsoid. ally modeled according to equation (2) and f s has the covariance function (4). f = f 0 + a1φ1 + a 2φ 2 ++ a k φ k + f s , (8) E{ f s }= 0 , E{ f s ( P ) f s (Q)}= C ( P, Q) Let it be noted however that for functions of both space and time, it is possible to follow a different modeling approach with respect to each argument (different treatment with respect to different variables). A typical example is the timeanalytical & space-stochastic approach: a space-time function f (t , P) is treated analytically with respect to time, e.g. by a model of type (2) The analytical (deterministic) approach: In this case f = f 0 + m , where the trend function m = m(a1 ,, a k ) is further modeled to depend on k k deterministic unknown parameters a1 ,, a k . Usually a linear combination model of the form m = a1φ1 + a 2φ 2 + + a k φ k (7) i f (t , P ) = f 0 ( P) + f 1 ( P)(t − t 0 ) ++ f n ( P)(t − t 0 ) n (2) (9) and in turn the coefficient space-functions are modeled independently by one of the approaches described above). Usually the f i (P) , i =1, 2, , n , are considered to be zero3 mean stochastic functions with known covariance functions C i ( P, Q) = E{ f i ( P ) f i (Q)} . The zero term f 0 ( P) = f ( P, t 0 ) can be treated either as a stochastic space-function (case of gravity potential W ( P, t 0 ) ) or as a deterministic function yielding signals which are treated as independent deterministic unknowns (case of coordinate function x( P, t 0 ) . More than one unknown function may be simultaneously present in the model, each one treated by a different modeling approach. achieved using known approximate values x 0 for the parameters x a , and a known approximate function f 0 for f , which gives rise to approximate signals s 0 = s a ( f 0 ) , using the known relation in (10) between f and the signals s a . The relation of the signals to the unknown function is typically a linear one, i.e. s a ( f 0 + δf ) = s a ( f 0 ) + s a (δf ) . In fact the signals arising in geodesy are either the values of the unknown function, or of its derivatives with respect to space and/or time, at a specific point and/or epoch. Setting The question of which model is better to use in a specific problem is a difficult one and cannot certainly be answered a-priori. A trial and error approach will give an insight into the capabilities of specific model choices. Statistical hypothesis testing can provide criteria for choosing between alternative models (see e.g. Dermanis and Rossikopoulos, 1991), although a theoretically complete and universally applicable method is still missing. However some simple rules can give a hint to the proper model choice: A slowly varying physical process can be better modeled by simple analytic models. A process which exhibits an erratic but still smooth variation around zero (or around an already modeled simple trend), can be better described by a stochastic model. In crustal deformation analysis, analytical models are best suited for time-variation and stochastic ones for spatial variation. Even when a stochastic model is used it is still possible to use an analytic model for its non-zero trend, e.g., by allowing some of the parameters of the normal potential to be treated as unknowns for best local fit. The standard normal potential, of either the Somigliana-Pizzetti type (i.e. with reference ellipsoid as equipotential surface) or given by one of the available spherical harmonic models, is designed to achieve a best global fit. xa =x0 +x , s a =s 0 + s , s0 =s a ( f 0 ) , δs = s a (δf ) , (12) b=yb −y 0 , y 0 = y a (x 0 , s 0 ) , (13) equation (10) is transformed into the linear form b = Ax + Gδs + v , (14) where ∂y a ∂x a A= A= , ( x 0 ,s 0 ) ∂y a ∂s a . Further development depends on the specific approach taken for the modeling of the unknown function f . As an example we take the more general case of the combined analytical-stochastic approach according to equation (8). The linear signal-to-function relation s a ( f ) applied to each specific 4.2. Basic form of mathematical model - Signals signal s ia becomes The unknown function f which affect the observed quantities does not appear directly in the mathematical model, but its presence is manifested through discrete unknown parameters which in turn depend on f and are called signals. s ia = s ia ( f ) = s ia ( f 0 + a1φ1 + + a k φ k + f s ) = = s ia ( f 0 ) + a1 s ia (φ1 ) ++ a k s ia (φ k ) + s ia ( f s ) = The mathematical model relating observables y a to un- = s i0 + a1 hi1 ++ a k hik + δs i , known parameters x a and signals s a has the general nonlinear form y a = y a (x a , s a ) , s a =s a ( f ) . (16) with (10) s i0 = s ia ( f 0 ) , The observations y b = y a (x a , s a ) + v , (15) ( x 0 ,s 0 ) hij = s ia (φ j ) , s i = s ia ( f s ) (17) or in matrix notation E{v}= 0 , E{vv T } = C v (11) s a = s 0 + δs = s 0 + Ha + s . are typically affected by zero mean random errors v with known covariance matrix C v . It is further assumed that signals and observational errors are uncorrelated so that C sv = 0 . (18) Replacing δs = Ha + s in (14) we obtain the more general linearized model x b = Ax + GHa + Gs + v =[ A GH ] + Gs + v a 4.3. General linearization scheme (19) where the deterministic parameters (x, a) have been separated from the stochastic signals s and the observational errors v. An essential step for the analysis of the available observations using standard estimation and prediction techniques is the linearization of the mathematical model, which is 4 and/or epoch. This means that in fact the function itself is reconstructed from the available observations thus also solving an interpolation problem. The interpolation approach can be used as an alternative to the above stochastic formulation, especially when the unknown function is known to be of a deterministic rather than a stochastic nature (see e.g. Dermanis, 1987). In this case the least squares principal (21) is replaced by Different linearized models arise when different approaches are followed to the modeling of the unknown function. In the analytical approach, the model becomes b = Ax + GHa + v , and in the purely stochastic one b = Ax + Gs + v . 4.4. The adjustment of the observations for parameter and signal estimation || f s || 2 + v T C −v1 v = min , After the linearization has been performed the mathematical model has the general form b = Ax + Gs + v , where || f s || is the norm of the "disturbing" function f s , assumed to belong to an appropriate function space. If the function admits a representation of the form (5), though with deterministic coefficients bi , it turns out that the norm in (23) has the form of a sum of squares (20) where x contains all possible deterministic parameters (original plus signal trend parameters if so modeled). This model is usually called the mixed linear model in the statistical literature. || f s || 2 = The adjustment of the observations is carried out by applying the least squares principle s T C s−1s + v T C −v1 v = min (23) ∑ pi bi2 (24) i which means that a smoothness condition is imposed on f s , simultaneously with the minimization of the sum of weighted residuals. The connection between the deterministic (interpolation) and the stochastic (prediction) is established by letting p i =1 / σ i2 , σ i2 being the variances of the (21) which leads to best linear unbiased estimates for the deterministic parameters x , a and best linear unbiased predictions for the stochastic ones s , v (best in the sense of minimizing the mean square error, see Dermanis, 1991). coefficients bi according to (6), i.e. in a way similar to that in relating deterministic least squares adjustment with best linear unbiased estimation. For more details on this duality see Dermanis (1976, 1978, 1983, 1987). The covariance matrix C s = E{ss T } of the signals is obtained from C ( P, Q ) by applying the law of covariance propagation to the relations s = s a ( f s ) . The algorithms which can be used for the adjustment of (20) are presented in Appendix A, summarized from Dermanis and Fotiou (1992), see also Koch (1987). 5. 4.5. Prediction of new signals - Interpolation The observations we shall consider here for inclusion in an adjustment according to the integrated approach will be limited to the usual observations in geodetic networks as well those of leveling and gravity surveys. Other types of geometric observations, such as those from strain-meters or tilt-meters, can also be included in a straightforward way. 5.1. Basic models of geodetic observables An essential additional feature of the algorithm is the possibility to predict the outcomes of stochastic signals s ′ others than those s appearing in the model. When s is directly observed s ′ can be predicted using the well known predic−1s , provided that E{s ′} = 0 , E{s} = 0 tion formula sˆ ′ = C s′s C ss and the relevant cross-covariance and covariance matrices C s′s = E{s ′s T } and C ss = E{ss T } are known, except possibly for a common scalar factor which cancels out. When s is not directly observed but appears in the linearized model (20) the (best linear unbiased) prediction turns out to be equivalent to a two-step procedure: First estimates ŝ are obtained (see Appendix C) and the prediction is carried out next according to −1sˆ . sˆ ′ = C s′s C ss Application of integrated geodesy In the more general case an observation y carried out at a point P depends on the position of P , expressed through global cartesian coordinates x , the gravity vector g = gradW at the same point and possibly the position of other (target) points. The linearization of the mathematical model of the form y = y (x, g,) is based on approximate coordinates x 0 and the normal potential U approximating the actual gravity potential W , which gives rise to the normal gravity vector = gradU . Linearization follows the scheme (22) ( ) ∂y ∂y ∂g (x − x 0 ) + y − y x 0 , (x 0 ), = + ∂x ∂g ∂x 0 Prediction is an excellent tool for the direct determination, at any required point and epoch, of exactly those parameters, which are the most appropriate for the description of crustal deformation. ∂y + gradT (x 0 ) + ∂g 0 In particular it is possible to predict signals s ′ , which are the values of the unknown function f at any desired point 5 (25) where T =W −U is the disturbing potential and the relation gradT = g − has been used. The partial derivatives are g = g (g ) = g 12 + g 22 + g 32 evaluated using x 0 and U , in place of x and W , i.e. by replacing g with , gravity g =|g | with normal gravity Λ = Λ (g ) = arctan γ =| | , astronomic longitude Λ and latitude Φ with their normal counterparts λ and φ . Φ = Φ (g ) = arctan Theodolite observations at a point P depend not directly on g , but on the unit vector n in the opposite direction 1 n=− , g g =[cos Φ cos Λ cos Φ sin Λ sin Φ ] T (31) g2 g1 (32) −g3 (32) g 12 + g 22 and they can be linearized according to the simplified general scheme (26) ∂y ∂g ∂y (x − x 0 ) + gradT (x 0 ) y − y ( (x 0 ) )= ∂ ∂ g x 0 ∂g 0 and they are more directly expressed in terms of the pointto-target vector x Q − x P expressed in the local astronomic frame at P The resulting linear models have the following form: = R ( x Q − x P ) = R 1 (90 $ − Φ P ) R 3 (90 $ + Λ P )(x Q − x P ) Two point observations (from x P to x Q ): (34) (27a) α −α 0 = a αT P (x P − x 0P ) + a αTQ (x Q − x Q0 ) + c αT P gradT (x 0P ) (35) or explicitly ζ −ζ 0 = a ζTP (x P − x 0P ) + a ζTQ (x Q − x Q0 ) + c ζTP gradT (x 0P ) (36) 0 0 ξ1 1 ξ = 0 cos(90$ − Φ ) sin(90$ − Φ ) × 2 ξ3 0 − sin(90$ − Φ ) cos(90$ − Φ ) s − s 0 = a TsP (x P − x 0P ) + a TsQ (x Q − x Q0 ) . (27b) Single point observations (at point x ): cos(90 + Λ ) sin(90 + Λ ) 0 ( xQ − xP ) × − sin(90$ + Λ ) cos(90$ + Λ ) 0 ( yQ − yP ) 0 0 1 ( zQ − z P ) $ $ The mathematical models for the azimuth α the zenith distance ζ and the spatial distance s from point P to point Q are ξ α PQ =α (x P , x Q , g P ) = arctan 1 ξ2 ζ PQ =ζ (x P , x Q , g P ) = arctan sPQ = s (x P , xQ ) = (x P − xQ )T (x P − xQ ) = (30) = ( x P − xQ ) + ( y P − y Q ) + ( z P − z Q ) 2 2 (38) Λ − λ 0 = a TΛ (x − x 0 ) + c TΛ gradT (x 0 ) (39) Φ −φ 0 = a TΦ (x − x 0 ) + c TΦ gradT (x 0 ) . (40) Instead of cartesian coordinates x , it is possible to use any other type of curvilinear coordinates, of which more convenient are the geodetic coordinates z =[b λ H ]T . Geodetic longitude λ coincides with the normal longitude since the adopted normal field U is rotationally symmetric. H is the geometric height above the reference ellipsoid and does not relate directly to the dynamic heights computed by standard leveling procedures. (29) ξ3 g − γ 0 = a Tg (x − x 0 ) + c Tg gradT (x 0 ) Expressions for the coefficients in the above relations are also given in Appendix A. (28) ξ 12 + ξ 22 (37) To introduce geodetic coordinates in the linear models it is sufficient to replace 2 ∂x x − x 0 = (z − z 0 ) , ∂z 0 Horizontal angles θ and directions δ are not considered explicitly here, because they are directly related to azimuths, i.e. θ PQR =α PR −α PQ , δ PQ =α PQ −α 0 , α 0 being the un- (41) T ∂x gradT (x 0 ) = ∂z 0 known orientation constant. Observations of gravity g and astronomic observations of longitude Λ and latitude Φ depend only on the gravity vector g =[ g 1 g 2 g 3 ]T and have the form T ∂T ∂T ∂T ∂ ∂b ∂H 0 z where the required derivatives follow from differentiation of the well known relation 6 ( N + H ) cos b cos λ x = x(z ) = ( N + H ) cos b sin λ , [(1− e 2 ) N + H ]sin b N= a H is eliminated from combination with the observed potential W (H ) to obtain the gravity anomaly ∆g = g (x) −γ (x 0 ) : (42) 1− e 2 sin 2 b ∆g = −n T0 gradT − a being the semi-major axis and e 2 the eccentricity of the reference ellipsoid, which yields − ( N + H ) cos b sin λ = ( N + H ) cos b cos λ 0 (43) − ( M + H ) sin b cos λ − ( M + H ) sin b sin λ ( M + H ) cos b 1 n0 =− , γ pseudo-observations, which are already in a simple linear form (44) δx ij − x 0j + x i0 = (x j − x 0j ) − (x i − x 0i ) + v δxij Note that (48) when more than one campaign is involved care should be taken for the different orientation of the reference frames involved (absolute position plays no role in the relative mode). The model for coordinate differences in a satellite frame is δx ijs = R (α , β ,γ )(x j − x i ) , where α , β , γ are the (45) is the unit vector normal to the reference ellipsoid and cos b cos λ N cos b cos λ x = N cos b sin λ + H cos b sin λ = x E + Hm sin b (1− e 2 ) N sin b ∂ , ∂x GPS observations can be included after their independent preprocessing in order to determine and remove systematic effects. The results are estimates of coordinate differences δx ij = x j − x i , between points Pi and P j to be used as with cos b cos λ ∂x = cos b sin λ m= ∂H sin b , (all terms evaluated at x 0 ) which is a more precise form of the usual "spherical approximation" ∆g = − Tr − 2r T . cos b cos λ cos b sin λ sin b M = a (1− e 2 )(1− e 2 sin 2 b) −3/ 2 . n T0 m γ (47) M= ∂x ∂x ∂x ∂x = = ∂z ∂λ ∂b ∂H n T0 Mm T three parameters of rotation relating the earth frame to the satellite one. The linearized form is (46) δx ijs − R 0 (x 0j − x i0 ) = R 0 (x j − x 0j ) − R 0 (x i − x i0 ) + (49) + R α0 (x 0j − x i0 )δα + R 0β (x 0j − x 0i )δβ + R γ0 (x 0j − x i0 )δγ + v δx ij where x E is the projection of x on the reference ellipsoid. Not all the terms in the linearized equations should be retained in the adjustment procedure. The inclusion of a certain parameter in an observation equation is meaningful only when the relevant observation contains sufficient information for the determination of the parameter. where δα =α −α 0 , δβ = β − β 0 , δγ = γ − γ 0 , R α = ∂ R , ∂α ∂ ∂ Rβ = R , R γ = R , and the zero index denotes ∂β ∂γ evaluation with the approximate values α 0 , β 0 , γ 0 . The terms of (x − x 0 ) in single point observations of g , Λ and Φ should be included only when carried out at network points. In isolated points the solution tends to minimize the signals in gradT (x 0 ) even to the point of completely elimi- The signals appearing in the linearized observation equations are the values at specific known points x 0 , of the disturbing potential T and of its gradient gradT , alternatively expressed by the derivatives with respect to geodetic coordinates ∂T , ∂T , ∂T . ∂λ ∂b ∂H nating them by shifting the point position from x 0 to an appropriate point x such that g (x) = (x 0 ) . Elimination of these coordinate terms is equivalent to setting x − x 0 = 0 , which means that it is essentially assumed that x = x 0 is exactly known. 5.2. The treatment of leveling observations There are two ways to treat and model leveling observations. In the classical approach observed height differences ∆h are converted to potential differences ∆W = − g∆h , utilizing observed or independently predicted gravity values, the conversion being a finite approximation to the relation g = − ∂W , where h is height along the plumb line. These ∂h differences are summed along each leveling route to produce an "observed" potential difference ∆W =W B −W A A different linearization approach is based on the concept of telluroid mappings (Grafarend, 1978) which utilize observations at x to obtain the approximate coordinates x 0 . The procedure corresponds to the elimination of the term (x − x 0 ) from a set of observations to obtain a "reduced" observations that depends only on T and gradT . We shall give only the application to observations of gravity where the horizontal position is held fixed ( λ = λ 0 , b = b 0 ) while 7 between the two endpoints A and B . The observation is then related to the disturbing potential by These can be solved for rP , rQ to produce the observed "height" difference ∆W − ∆W 0 = = T (x0B )(x B − x0B ) − T (x0A )(x A − x0A ) + T (x0B ) − T (x0A ) , (50) ∆W 0 =U (x 0B ) −U (x 0A ) ∆hPQ = rP − rQ = T nTL (x L − x P ) n L (x L − xQ ) − = nTL n P nTL nQ 1 nTL xQE nTL x EP 1 = T − T nTL x L + T − n L n P n L nQ n L nQ nTL n P If leveling endpoints are isolated, i.e. not part of a 3-dimensional geodetic network the (x A − x 0A ) and (x B − x 0B ) terms should be related to geometric heights using fixed values λ0 , b 0 , for geodetic longitude and latitude, in which case m A = m A (λ0 , b 0 ) , m B = m B (λ0 , b 0 ) and (55) This model is too complicated to be practical and furthermore contains the position vector of the level x L , which may not be even approximately known. It can be simplified by assuming that n L has the direction of 12 (n P + n Q ) , im- x A − x 0A = ( H A − H A0 )m A , plying that (51) n TL n P = n TL n Q , and leading to the simpler x B − x 0B = ( H B − H B0 )m B , model A more precise model for leveling can be introduced on the basis of the individual height differences produced by setting the level at point x L , and the staff at points x P and ∆h PQ = xQ . If x EP , x QE are the projections on the reference ellipsoid of 1 (n P + n Q ) T ( x Q − x P ) 1+ n TP n Q (56) x P , x Q , respectively, H P , H Q are the geometric heights (above the ellipsoid) and m P , m Q the corresponding unit vectors normal to ellipsoid, it holds that xr x P = x EP + H P m P , r x Q = x QE + H Q m Q (57) n and the model becomes x ∆h PQ = H m 1 (n P + n Q ) T (x QE + H Q m Q − x EP − H P m P ) 1+ n TP n Q (58) xE which has the form ∆h PQ = ∆h PQ (n P , n Q , H P , H Q ) . Its linearized form derived in appendix B is ∆h − ∆h 0 = ahP ( H P − H P0 ) + ahQ ( H Q − H Q0 ) + If the corresponding vertical unit vectors are n L , n P and + cThP gradT (x0P ) + cThQgradT (x0Q ) . n Q , and rP and rQ are the readings on the staff, the position vectors of the reading points are x rP = x P + rP n P , x rQ = x Q + rQ n Q . Perhaps a more reasonable assumption is that g L = 12 (g P + g Q ) but the choice n L = 12 (n P + n Q ) made (52) here has the advantage that eliminates x L from the model. A similar linearization scheme may be found in Milbert (1988). Taking into account the fact that the lines of sight from x L to the reading points x rP , x rQ should be horizontal, i.e., perpendicular to vertical n L , we obtain the orthogonality conditions + rP n P − x L ) = 0 (53) n TL (x rQ − x L ) = n TL (x Q + rQ n Q − x L ) = 0 (54) n TL (x rP − x L ) = n TL (x P (59) 5.3. Dynamic modeling The passing from the static to dynamic modeling can be more conveniently performed after the linearization, by noting that position vectors x(t ) , geometric heights H (t ) , are functions of time t in a deformable earth, while the 8 distance d PQ when a planar approximation is used within a disturbing potential T (x, t ) and its gradient gradT (x, t ) are functions of both space and time. By setting small area. The gravity related signals T (x 0 , t ) and gradT (x 0 , t ) depend on the modeling of the time varying disturbing potential function T (x, t ) . This can be treated in a stochastic way for both the time and the space argument, with the introduction of a space-time covariance function x P (t ) = x P (t 0 ) + δ t x P (t ) , (60) H P (t ) = H P (t 0 ) + δ t H P (t ) , where t 0 is a chosen reference epoch, position information is separated into two parts: A deterministic unknown part x P (t 0 ) , H P (t 0 ) and a displacement signal δ t x P (t ) , CT ( P, Q, t , t ′) = C (ψ PQ , | t ′ − t |) . δ t H P (t ) . The three-dimensional displacements or the height displacements have the character of signals because they depend on corresponding displacement functions A simpler and perhaps more appropriate choice is to follow the same procedure as for the displacements and set T (x, t ) = T (x, t 0 ) + T (x)(t − t 0 ) + δ t x( P, t ) = δ t x(x P (t 0 ), t ) , (61) δ t H ( P, t ) = δ t H (x P (t 0 ), t ) , (66) (67) where T (x, t 0 ) and T (x) are treated as independent stochastic functions with known covariance functions C T ( P, Q) and C T ( P, Q) . where points are identified by their coordinates at the reference epoch t 0 . Summarizing, the signals present in the linearized models are of three types: "displacement signals" d related to the Further model development depends on how the displacement functions are modeled. Since crustal motion is a slow process (apart from the occurrence of earthquakes), it can be sufficiently modeled in an analytical way with respect to time, e.g., functions x (P ) or H (x) , "gravity at reference epoch" signals g related to the disturbing potential T (x, t 0 ) and "gravity variation" signals q related to the rate of the dis- δ t x(x, t ) = δ t x0 (x, t ) + δ t x1 (x, t )(t − t0 ) + turbing potential T (x) . When all signals are treated as stochastic parameters the mixed linear model (20) takes the form + δ t x 2 (x, t )(t − t 0 ) 2 + = = x (x)(t − t 0 ) + x(x)(t − t 0 ) 2 + b = Ax + Gs + v = Ax + G d d + G g g + G q q + v , (62) where the first time-independent term has been eliminated since it is already contained in x = x(t 0 ) . Usually a piecewise linear trend between earthquake events is sufficient and only the x (x) function is preserved. This function is characterized by a strong similarity between neighboring points (not separated by faults) since crustal motion is known to be a spatially smooth process. This similarity in proximity can be best described by a spatially correlated stochastic model for the function x (x) . We assume zero mean E{x (x)}= 0 , and a known matrix of covariance functions between the components of x , C x ( P, Q ) = E{x ( P )x (Q) T } . G = [G d G g G q ] , δt 0 ( x)(t − t )+ H 5.4. Adjustment strategies The linearized observations equations for the study of crustal deformation are formulated from the static linearized models by introducing time dependence, replacing observables y with observed values y b and adding the observa- (63) 0 )2 + (68) Although other modeling possibilities are possible we will restrict ourselves to the above ones for the shake of demonstration. tional errors v y : A completely similar approach is taken for the vertical displacements H (x, t ) = H (x)(t − t d s = g q α b − α 0 = a αT P (x P − x 0P ) + (t − t 0 )a αT P x (x 0P ) + (64) + a αTQ (x Q − x 0Q ) + (t − t 0 )a αT Q x (x Q0 ) + E{H ( P)}= 0 , C H ( P, Q ) = E{H ( P ) H (Q )} . (65) + c αT P gradT (x 0P ) + (t − t 0 )c αTP gradT (x 0P ) + vα The covariance functions C x and C H are assumed to be stationary and isotropic, i.e. to depend only on the spherical distance ψ PQ between the two points, or their horizontal ζ b − ζ 0 = a ζTP (x P − x 0P ) + (t − t 0 )a ζTP x (x 0P ) + + a ζTQ (x Q − x Q0 ) + (t − t 0 )a ζTQ x (x Q0 ) + 9 (69) + c ζTP gradT (x 0P ) + (t − t 0 )c ζTP gradT (x 0P ) + vζ This means that coordinates of isolated observation points cannot be determined like the ones in a geodetic network. The same holds for heights at points not belonging in a leveling network. The signals do not impose a determinability problem, in theory at least, but a very crucial point from the practical point of view is whether their predicted values are obtained from information contained mainly in the observations or in their adopted covariance functions. A related problem is the separability of various different signals ( d , g , q ) present in the same observations, namely whether the separation is a result of the experimental design (i.e. the form of the matrices G d , G g , G q ) or of their adopted (70) s b − s 0 = a TsP (x P − x 0P ) + (t − t 0 )a TsP x (x 0P ) + + a TsQ (x Q − x Q0 ) + (t − t 0 )a TsQ x (x Q0 ) + v s (71) g b − g 0 = a Tg (x − x 0 ) + (t − t 0 )a Tg x (x 0 ) + + c Tg gradT (x 0 ) + (t − t 0 )c Tg gradT (x 0 ) + v g (72a) covariance functions. The choice of covariance functions is a critical problem, because they cannot be determined by sampling from directly observed signal values (available observations depend simultaneously on all of them). or when a telluroid mapping x → x 0 with W (x) = U (x 0 ) is involved ∆g b = −n T0 gradT − (t − t 0 )n T0 gradT − − n T0 Mm T n T0 Mm T − − + vg ( t t ) 0 n T0 m γ n T0 m γ The computational burden should be also taken into account. The simultaneous treatment of all available observations leads to extensive computations involving the inversion of large matrices. Sequential treatment of the data and deviations from the rigorous sequential solution are sometimes adopted for computational convenience in the hope that the derived solution is close to the theoretically optimal one. (72b) Λb − λ 0 = a TΛ (x − x 0 ) + (t − t 0 )a TΛ x (x 0 ) + + c TΛ gradT (x 0 ) + (t − t 0 )c TΛ gradT (x 0 ) + v λ (73) Two different sets of observations play a different role in the data analysis. Observations at isolated but densely distributed points in the studied area contribute to the determination by interpolation of the functions giving rise to the signals. Observations at network points contribute to the determination of position and position variation. This leads to a question with respect to the displacement signals. These can be treated as independent deterministic signals at network points, ignoring their dependence on a smooth displacement function, but their interpolation outside the network or their treatment at isolated points calls for the use of a stochastic model. The simultaneous use of two different models for the same displacement signals is very unsatisfactory (not to say unacceptable) from a theoretical point of view but it has been implemented in practice, even indirectly by the use of independent coordinate unknowns at the different epochs of repeated observational campaigns. Φ b − φ 0 = a TΦ (x − x 0 ) + (t − t 0 )a TΦ x (x 0 ) + + c TΦ gradT (x 0 ) + (t − t 0 )c TΦ gradT (x 0 ) + vφ (74) ∆W b − ∆W 0 = T (x 0B )(x B − x 0B ) + (t − t 0 ) T (x 0B )x (x 0B ) − − T (x 0A )(x A − x 0A ) + (t − t 0 ) T (x 0A )x (x 0A ) − (75) + T (x 0B ) − T (x 0A ) + (t − t 0 )T (x 0B ) − (t − t 0 )T (x 0A ) + vW or ∆h b − ∆h 0 = a hP ( H P − H P0 ) + (t − t 0 )a hP H (x 0P ) + + a hQ ( H Q − H Q0 ) + (t − t 0 )a hQ H (x 0Q ) + Coming to the signal separability problem, we note that displacements d and gravity variations q are physically interrelated and they cannot be effectively separated since they have similar effects on observations: It is impossible to say whether the observed change is due to the motion of a point within an invariant gravity field, or to the variation of the gravity field at a point which does not move at all, or to a combination of the two effect. The actual modeling of the interrelation between the two requires knowledge of the internal masses and (in principle) global data coverage. An approach that has been suggested is to consider an invariant gravity field with respect not to an earth-fixed global reference frame but with respect to a local one moving along with the network and the surrounding area (Reilly, 1981). + c ThP gradT (x 0P ) + (t − t 0 )c ThP gradT (x 0P ) + + c ThQ gradT (x Q0 ) + (t − t 0 )c ThQ gradT (x Q0 ) + v h . (76) In all the above relations, coordinates (x P , x Q , x) refer to the reference epoch t 0 . The derivation of the linearized observation equations (68) does not lead to a straightforward adjustment, even if the problem of choosing the proper models for the functions T , T , x or H has been effectively solved. Here we will assume a purely stochastic model for these functions. We shall point out and discuss the main problems involved. Keeping all the above problems in mind we shall give a rigorous sequential approach to the adjustment of observations at a network (horizontal, three-dimensional or vertical) in combination with observations in a dense network of isolated points in the sane area. The coordinates of isolated points are assumed known with sufficient accuracy so that Deterministic parameters should be included only when their expected influence on the observations is significant (above the noise level) and they could be determined by standard least squares if the signals were exactly known. 10 they do not appear as unknowns. The same can be done for the displacement signals at isolated points, though this approach is not followed here in the demonstrated algorithm. ~ b1 = A 1 x + G 1δs 1 + v 1 , The structure of the observation equations (98) can be more easily seen when they are separated into two sets, one b 1 0 δs1 = s1 − sˆ 10 ~ δQ 11 = Q 11 − Q11 . for network points and one b 2 for isolated points: STEP 4: Solution for the reduced model : b 1 = A 1 x + G 1d d 1 + G 1g g 1 + G 1q q 1 + v 1 = ~ T M 11 = G 1δQ11G 1T + P1−1 = M 11 − M 12 M −221 M 12 d 1 1 1 1 = A 1 x + [G d G g G q ]g 1 + v 1 = A 1 x + G 1s 1 + v 1 q 1 (82) (83) where (77) M 11 = G 1Q11G 1T + P1−1 , ( ~ −1 xˆ = A 1T M 11 A1 b 2 = G 2d d 2 + G 2g g 2 + G 2q q 2 + v 2 = ) −1 (78) M 12 = G 1Q 12 G T2 , (84) ~ −1 ~ A 1T M 11 b1 ~ δsˆ 1 = δQ 11G 1T M −1 (b1 d 2 = [G d2 G 2g G q2 ]g 2 + v 2 = G 2 s 2 + v 2 q 2 ~ b1 = b1 − G 1sˆ 10 , (85) − A 1 xˆ ) . STEP 5: Prediction of residual signal for s 2 (contribution ~ from b1 : where d , g and q are the vectors of displacement, gravity and gravity variation signals respectively. T T − Q 22 G T2 (G 2 Q 22 G T2 + P2−1 )G 2 Q12 Q δs 2δs1 = Q 12 (86) The two sets contain no common unknowns or signals, but they cannot be treated independently due to the existing non zero cross covariance matrices C d1d 2 , C g1g 2 , C q1q 2 . It is ~ −1 ~ −1 δsˆ 2 = Q δs 2δs1 δQ 11 δsˆ 1 = Q δs 2δs1 G 1T M 11 (b1 − A 1 xˆ ) . (87) this correlation which allows information about the underlying functions x , T , T to pass from the isolated to the network points. The observation equations have the form of the mixed linear model b = Ax + Gs + v , s ~ σ 2 Q , STEP 6: Reconstruction of model signals: sˆ 1 = sˆ 10 + δsˆ 1 , (88) STEP 7: Prediction of new signals: v ~ σ 2 Q v = σ 2 P −1 , with solution presented in appendix C. Taking into account the special form of the model (77), (78) which is of the form b 1 A 1 G 1 = x + b 2 0 0 sˆ 2 = sˆ 02 + δsˆ 2 . −1 T δsˆ 1 = sˆ ′ = Q s′s 2 Q −221 sˆ 02 + (Q s′s1 − Q 22 G T2 Q 12 )δQ 11 (89) = Q s′s 2 G T2 M −221 b 2 + 0 s1 v1 + , G 2 s 2 v 2 ~ T + (Q s′s1 − Q 22 G T2 Q12 )G T M −1 (b1 − A 1 xˆ ) . (79) s 1 Q 11 ~ T s 2 Q 12 Q 12 , Q 22 v 1 P1−1 v ~ 2 0 0 P2−1 The above solution is a "rigorous sequential solution" (Dermanis, 1986) identical with the non-sequential solution from the algorithms described in appendix C. the solution can be described in a sequential scheme as follows: An alternative to the optimal common adjustment that has been commonly followed is to arrive at a suboptimal solution in a two step procedure: In the first step the b 2 obser- STEP 1: Prediction of s 2 from b 2 only: vations are used to obtain estimates of the signals s 2 (in fact only g 2 , q 2 because d 2 cannot be detected in isolated M 22 = G 2 Q 22 G T2 + P2−1 , sˆ 02 = Q 22 G T2 M −221 b 2 . points and d 2 = 0 is assumed), as well as to predict the (80) "new" signals s 1 (i.e. g 1 , q 1 ), at the network points, thus STEP 2: Prediction of s 1 from ŝ 02 : ignoring the availability of the b 1 observations which also contain information on s 1 . In the second step the signals s 1 , are fixed to their predicted values from the previous step, their effect is subtracted from the observed values b 1 , and the adjustment is carried out on the basis of the reduced observations sˆ 10 = Q 12 Q −221 sˆ 02 = Q 12 G T2 M −221 b 2 , (81) 0 T Q11 = Q sˆ 0sˆ 0 = Q 12 G T2 M −221 G 2 Q12 . 1 1 STEP 3: Reduction of the model for b 1 to the new form 11 b 1R = b 1 − G 1g g 1 − G 1q q 1 = A 1 x + G 1d d 1 + v 1 (90) Since horizontal position cannot be determined from such observations we will assume in eq. (91) that This approach is computationally attractive but it suffers from two defects: (a) it leads to suboptimal solutions (i.e. with larger mean square errors) (b) the models used in the two steps for the signals (stochastic in the first, deterministic in the second) are inconsistent. x − x 0 = m( H − H 0 ) = m∆H , (92) in which case the observation equation for gravity becomes In conclusion, the answers to problems related to modeling, adjustment strategy and adoption of computationally attractive compromises, cannot be answered by theory, but mainly from repeated practical experience obtained from trial and error. g b − g 0 = a g ∆H + (t − t 0 )a g H (x 0 ) + + c Tg gradT (x 0 ) + (t − t 0 )c Tg gradT (x 0 ) + v g (93) with 5.5. An example: Combination of leveling and gravity a g = a Tg m = c Tg Geometric uplift of the earth surface can be separated into two parts. One part corresponds to the variation of orthometric height, i.e. to motion with respect to the geoid, while the other refers to the variation of the geoid undulation, i.e. to the motion of the geoid with respect to the reference ellipsoid. The corresponding geodetic techniques for the determination of these two components of vertical displacement are, respectively, repeated leveling and repeated gravity observations. Strictly speaking, dynamic leveling provides potential values W , which are converted to "heights" when divided by a constant mean value γ 0 of gravity in the area. These (dynamic) heights are only an approximation to orthometric heights and heights above the geoid along the ellipsoidal normal. ∂ ∂ m = −n T0 m = −n T0 Mm , ∂x ∂x (94) ∂ , M= ∂x where the values of a Tg and c Tg from appendix A have been used. Eq. (93) is appropriate for gravity observations at the leveling points. For gravity observations at isolated points within the same area, the heights cannot be determined and H = H 0 must be set to obtain ∆H = 0 and g b − γ 0 = (t − t 0 )a g H + c Tg gradT + (t − t 0 )c Tg gradT + v g . Repeated gravity observations can be used to obtain the "velocity" of gravity g , or of gravity anomalies ∆g , which (95) can be integrated to obtain velocities ζ of the geoid undulations ζ . This procedure corresponds to a solution of a geodynamic boundary value problem, presented by Sanso and Dermanis (1981) and Heck (1981). For leveling the observation equation is eq. (76): ∆h b − ∆h 0 = a hP ∆H P + (t − t 0 )a hP H P + Apart from the above classical approach to the combination of independent leveling and gravimetric work, an integrated approach is also possible. A first step is the use of gravity data in the leveling area for the prediction of gravity values necessary for the conversion of primarily observed "height differences" ∆h into the geopotential differences ∆W = g∆h , which will enter in the adjustment of the leveling network. However in high precision leveling aiming to the study of crustal motion it is standard practice to make gravity observations along the leveling route, instead of relying on predicted gravity, or even worst, in theoretical normal gravity values γ in place of g . Further information on the above approach can be found in Heck and Mälzer (1983, 1986). + a hQ ∆H Q + (t − t 0 )a hQ H Q + + c ThP gradTP + (t − t 0 )c ThP gradTP + + c ThQ gradTQ + (t − t 0 )c ThQ gradTQ + v h . (96) The parameters of the above equations can be distinguished to deterministic parameters x , displacement signals d , gravity signals g , and gravity variation signals q : H P ∆H P x= , d = H Q , ∆H Q H A purely integrated approach will be outlined here, based on the precise models of leveling and gravity observations, which make use of geometric heights H above the reference ellipsoid as parameters. The observation equation for gravity is eq. (72a): g b − g 0 = a Tg (x − x 0 ) + (t − t 0 )a Tg x (x 0 ) + + c Tg gradT (x 0 ) + (t − t 0 )c Tg gradT (x 0 ) + v g x = mH (91) 12 gradT P g = gradTQ , gradT gradT P q = gradTQ gradT (97) We can differentiate between observations at leveling points (subscript 1) and observations in isolated points (subscript 2) H P d1 = , H Q gradT P g1 = , gradTQ gradT P q1 = , gradTQ g 2 = gradT , d 2 = H , c T hP 1 T G g = c gP 0 (98) q 2 = gradT , to obtain two sets of observation equations b 1 = A 1 x + G 1d d 1 + G 1g g 1 + G 1q q 1 + v 1 = b2 = G d2 d 2 + G 2g g 2 (100) + G q2 q 2 + v 2 = G 2s 2 + v 2 G 2g (101) 0 ∆h b − ∆h 0 b 0 b 1 = g P − γ P A 1 = 0 b 0 0 gQ −γ Q a hP a hQ a gP 0 0 a gQ c TgQ = (t g − t 0 )c Tg = c Tg (t h − t 0 )c ThQ 0 (t gQ − t 0 )c TgQ (104) , . (105) 0 0 , 0 Appendix A: Linearization of basic geodetic observables In this appendix the coefficients a , c of the linearized observation equations (35)-(40) will be given adopted from Dermanis (1985b) and Dermanis and Rossikopoulos (1988). For any observable y = y (x, , g (x) ) , the linearized observation equations for the observed value y b have the form b 2 = g b − γ 0 G d2 = (t g − t 0 )a Tg 0 G 2g The sequential adjustment algorithm described in section (5.4) can now be applied. where (t − t )a 0 h hP G 1d = (t gP − t 0 )a gP 0 , (t − t )c T 0 hP h G 1q = (t gP − t 0 )c TgP 0 (99) = A 1 x + G 1s 1 + v 1 c ThQ y b − y (x 0 , , (x 0 ) ) = a Ty (x − x 0 ) + + c Ty gradT (x 0 ) . (102) (A1) (t h − t 0 )a hQ 0 (t gQ − t 0 )a gQ , The coefficients rows a Ty , c Ty , for the various observables take the following values: c αT ξ = 22 d − sin φξ 2 − cos φξ 3 − sin φξ 1 0 0 cos φξ 1 ξ1 d2 sin λ γ cos φ 1 × sin φ cos λ γ cos φ cos λ (103) ξ a αT P = 22 d 13 − ξ1 d 2 − cos λ γ cos φ 1 sin φ sin λ γ − cos φ sin λ 0 −ξ3 ξ2 1 − cos φ γ − sin φ 0 0 0 × 0 0 0 ∂ 0 0 (−R 0 ) + c αT (x P ) , ∂x 0 (A2) R0 = R ( λ (x 0P ), φ (x 0P ) ξ a αT Q = 22 d cT = 1 23 ds − ξ1 d 2 − ds 2 sin λ γ cos φ sin φ cos λ × γ cos φ cos λ a TQ = 1 23 ds c TΛ sin λ = γ cos φ 2 3 ds cos λ 0 γ cos φ sin φ sin λ cos φ − γ γ − cos φ sin λ − sin φ 0 − cos λ − γ cos φ sin φ cos λ c TΦ = γ a TΦ = c TΦ 0 0 = [− cos φ cos λ a Tg = ∆g = g − γ 0 = a Tg m (A6) ∂ 0 (x ) ∂x + c Tg gradT (x 0 ) = γ (x 0 )n T0 (x 0 )m T (x 0 ) − n T0 (x 0 )gradT (x 0 ) 1 1 + n TP n Q (n P + n Q ) T (x Q − x P ) = 1 1 + n TP n Q (n P + n Q ) T (x QE + H Q m Q − x EP − H P m P ) is carried out by fixing x EP (b P , λ P ) , x QE (bQ , λ Q ) and (A12) m P (b P , λ P ) , m Q (bQ , λ Q ) , i.e. by assuming that the horizontal position of leveling points is known. The linearization follows the scheme (A13) ∂∆h ∆h − ∆h = ∂H P 0 A different type of linearization for g utilizes knowledge of W (x) and the telluroid mapping x → x 0 ( H → H 0 ) defined by W (x) = U (x 0 ) . The term (x − x 0 ) in the linearized single point observation (38) for g is related to the disturbing potential T , by noting that cos b 0 cos λ 0 x − x 0 = ( H − H 0 ) cos b 0 sin λ 0 = ( H − H 0 )m sin b 0 γ (x 0 )n T0 (x 0 )m (B1) (A11) − sin φ ]0 = m. The linearization of the leveling model (58) which in terms of gravity has the form = − cos φ sin λ γ (x 0 ) Appendix B: Linearization of leveling observations (A10) −n T0 T (x 0 ) n T0 (x 0 )M (x 0 )m ∆h PQ = cos φ γ 0 T (x 0 ) (A17) (A8) − γ (x 0 )n T0 (x 0 )m m≈ With this value of (x − x 0 ) the linearized relation for g becomes (A5) (A9) sin φ sin λ γ T (x 0 ) (A16) (A7) ∂ 0 (x ) ∂x c Tg x − x 0 = ( H − H 0 )m = =− d R0 s2 0 (A15) which in view of W (x) = U (x 0 ) and = −γn 0 ( n 0 being the unit vector in the normal vertical direction), leads to the relation − ∂ 0 d (−R 0 ) + c T (x P ) 2 ∂x s 0 ∂U 0 (x )(x − x 0 ) + T (x 0 ) = ∂x = U ( x 0 ) + ( H − H 0 ) T ( x 0 )m + T ( x 0 ) 0 0 − 3 0 × 2 0 0 ∂ 0 (x ) ∂x a TΛ = c TΛ c Tg 2 W ( x) = U ( x) + T ( x) ≈ U ( x 0 ) + (A4) sin 3 2 − cos 3 3 d − sin 31 s 2 0 cos 31 − 2 2 3 ds ) 0 R 0 , 0 2 3 a TP = 1 23 ds (A3) * ∂∆ h ( H P − H P0 ) + ∂H Q 0 * ( H Q − H Q0 ) + 0 ∂∆ h ∂∆h gradT (x 0Q ) . gradT (x 0P ) + + ∂ ∂ g g P 0 Q 0 (B2) The derivatives with respect to H P and H Q are implicit ones and must be calculated from the explicit ones taking into account the dependence n = n(g ) = − 1 g g , and (A14) g = g (H ) . The explicit derivatives are easily found to be where m 0 = m is supposed to be known, and 14 ∂∆h 1 = (n P + n Q ) T m P , ∂H P 1 + n TP n Q ∆h 0 = 1 T 1 + n 0P n Q0 (n 0P + n Q0 ) T (x Q0 − x 0P ) (B11) (B3) ∂∆h 1 =− (n P + n Q ) T m Q T ∂H Q 1+ n PnQ ∂n Q ∂n P 1 = (n P n TP − I ) , ∂g P gP ∂g Q = n 0P = n 0 (x 0P ) , 1 (n Q n TQ − I ) gQ c ThP = (B4) c ThQ = (B5) ∂H Q = ∂g Q ∂x Q ∂x Q ∂H Q = ∂g Q ∂x Q 1 (B12) T T γ (x 0P )(1 + n 0P n Q0 ) (x Q0 − x 0P − ∆h 0 n Q0 ) T (n 0P n 0P − I ) (B13) ∂g P ∂g ∂x P ∂g = P = P mP , ∂H P ∂x P ∂H P ∂x P ∂g Q n 0Q = n 0 (x Q0 ) 1 T T γ (x Q0 )(1 + n 0P n Q0 ) 0 0 − ∆h 0 n 0P ) T (n Q0 n Q − I) (x 0P − x Q (B14) mQ a hP = (n 0P + n Q0 ) T m P and the implicit ones are T 1 + n 0P n 0Q + c ThP M (x 0P )m P , (B15) ∂∆h ∂H P T ∂ ∂ ∂U = M= , ∂x ∂x ∂x * ∂∆h ∂∆h ∂n P ∂g P = + , ∂H P ∂n P ∂g P ∂H P M ij = ∂ U ∂x i ∂x j 2 (B6) ∂∆h ∂H Q * ∂n Q ∂g Q = ∂∆h + ∂∆h ∂H Q ∂n Q ∂g Q ∂H Q a hQ = − The derivatives with respect to g can be obtained using the chain rule ∂∆h ∂∆h ∂n P = , ∂g P ∂n P ∂g P ∂∆h ∂∆h ∂n Q = ∂g Q ∂n Q ∂g Q (n 0P + n Q0 ) T m Q T 1 + n 0P n Q0 + c ThQ M (x Q0 )m Q . Appendix C: Adjustment algorithms for the mixed linear model. (B7) Given the mixed linear model of the form b = A x+G s+ v n×1 where ∂∆h 1 =− (x Q − x P ) T (n P + n Q )n TQ + 2 T ∂n P g P (1 + n P n Q ) + = 1 g P (1 + n TP n Q ) 1 g P (1 + n TP n Q ) n× q q×1 n×1 (C1) s ~ (v s , C s ) , (C2) C sv = E{(s − v s ) v T } = 0 Cs = σ 2Qs , (B8) C v = σ 2Q v (C3) the adjustment for the estimation of the deterministic parameters x and the prediction of the random parameters s and v can be performed using alternative algorithms which of course give identical optimal results. The choice of algorithm in each specific application depends on numerical considerations. One tries in general to implement the inversion of matrices with smaller dimensions. with a similar expression for Q . It follows after some rather lengthy computations that ∆h − ∆h 0 = a hP ( H P − H P0 ) + a hQ ( H Q − H Q0 ) + + c ThP gradT (x 0P ) + c ThQ gradT (x Q0 ) n× m m×1 v ~ (0, C v ) , (x Q − x P ) T = ( x Q − x P − ∆h n Q ) T (B16) Algorithm A: (B9) M = GQ s G T + Q v (C4) Q xˆ = ( A T M −1 A) −1 (C5) xˆ = Q xˆ A T M −1b (C6) where: x 0P = x EP + H P0 m P , x Q0 = x QE + H Q0 m Q , (B10) 15 sˆ = Q s G T M −1 (b − Axˆ ) (C7) vˆ = Q v M −1 (b − Axˆ ) ˆ , C ˆ , C ˆ of the covariance matrices C , Estimates C xˆ xˆ xˆ sˆ sˆ C xˆ sˆ , C sˆ , can be obtained by multiplying the "cofactor" (C8) matrices Q xˆ , Q xˆ sˆ , Q sˆ , respectively, with the estimate Q xˆ sˆ = −Q xˆ ( A T M −1G )Q s (C9) σˆ 2 = −1 Q sˆ = Q s − Q s (G M G )Q s + T + Q s (G T M −1 A)Q xˆ ( A T M −1G )Q s Q vˆ = Q v M −1Q v − Q v M −1 AQ xˆ A T M −1Q v vˆ T Q −v1 vˆ + sˆ T Q s−1sˆ . n−m (C27) (C10) When the covariance matrices C s , C v , are assumed to be (C11) completely known (no common unknown factor σ 2 ) the above algorithms still hold by replacing all cofactor matrices Q with the corresponding covariance matrices C . However the estimate σ̂ 2 should still be obtained and tested N G = G T Q −v1G + Q s−1 (C12) −1 T Q xˆ = [ A T Q −v1 A − G T Q −v1 AN G A Q −v1G ] −1 (C13) −1 T xˆ = Q xˆ A T Q −v1 (b − GN G G Q −v1b) (C14) statistically for σ 2 = 1 , in order to evaluate the overall validity of the model. Prediction of new signals s ′ with E{s ′} = 0 not contained in the model can be performed when the cross-covariance matrix C s′s = E{s ′s T } is known. It is also assumed that the new signals are uncorrelated to the observational errors, i.e., C s′v = E{s ′v T } = 0 . The prediction is given by −1 T sˆ = N G G Q −v1 (b − Axˆ ) (C15) sˆ ′ = C s′s C s−1sˆ vˆ = b − Axˆ − Gsˆ (C16) The prediction errors e = sˆ ′ − s ′ have covariance matrix −1 Q xˆ sˆ = −Q xˆ A T Q −v1GN G (C17) C e = C s′ − C s′s G T C −v1Q vˆ C −v1GC Ts′s −1 −1 T −1 + NG Q sˆ = N G G Q −v1 AQ xˆ A T Q −v1GN G (C18) where C s′ is the covariance matrix of s ′ . Algorithm B: −1 T Q vˆ = [I − ( A − Q xˆ + GQ sˆxˆ A T Q −v1 ][Q v − GN G G ]× × [I − ( A − Q xˆ + GQ sˆxˆ A T Q −v1 ]T (C28) (C29) References (C19) Bencini, P., A. Dermanis, E. Livieratos and D. Rossikopoulos (1982): Crustal Deformation at the Friuli Area from Discrete and Continuous Geodetic Prediction Techniques. Bollettino di Geodesia e Scienze Affini, XLI, 2, 137-148. Bruns, H. (1978): Die Figur der Erde. Publ. Preuss. Geod. Inst., Berlin. Dermanis, A. (1976): Probabilistic and Deterministic Aspects of Linear Estimation in Geodesy. Report No. 244, Department of Geodetic Science, The Ohio State University. Dermanis, A. (1978): Adjustment of Geodetic Networks in the Presence of Signals. Proceedings, International School of Advanced Geodesy, 2nd Course: "Spacetime Geodesy, Differential Geodesy and Geodesy in the Large". 'Ettore Majorana' Centre for Scientific Culture, Erice, Sicily, May 1978. Bollettino di Geodesia e Scienze Affini, 38, 4, pp. 513-539. Dermanis, A. (1981): Geodetic Estimability of Crustal Deformation Parameters. Quaterniones Geodaesiae, vol. 1, no. 2, 159-169. Dermanis, A. (1983): Theory and Applications of Collocation in Surveying. In: M. Unguendoli (ed.): "Techniche moderne di analisi dei dati geodetici con particolare riguardo alla collocazione", pp. 37-67, Editice CLUEB, Bologna. Algorithm C: Same as above with the following alternative relations Q sˆ = [G T Q −v1G − G T Q −v1 A( A T Q −v1 A) −1 A T Q −v1G + Q s−1 ] −1 (C20) sˆ = Q sˆ [G T Q −v1b − G T Q −v1 A( A T Q −v1 A) −1 A T Q −v1b] (C21) xˆ = ( A T Q −v1 A) −1 A T Q −v1 (b − Gsˆ ) (C22) vˆ = b − Axˆ − Gsˆ (C23) Q xˆ sˆ = −( A T Q −v1 A) −1 A T Q −v1GQ sˆ (C24) Q xˆ = ( A T Q −v1 A) −1 + + ( A T Q −v1 A) −1 A T Q −v1GQ sˆ G T Q −v1 A( A T Q −v1 A) −1 (C25) Q vˆ = R − RGQ sˆ G T Q −v1 R , (C26) R = [I − A( A T Q −v1 A) −1 A T Q −v1 ] 16 Grafarend, E.W. (1978): The definition of the telluroid. Bulletin Géodésique, 52, 25-37. Heck, B. (1981): Combination of Leveling and Gravity Data for Detecting Real Crustal Movements. Int. Symp. on Geodetic Networks and Computations, Munich 1981, DGK, Reihe B, Heft Nr. 258/VII, 20-30. Heck, B. and H. Mälzer (1983): Determination of Vertical Recent Crustal Movements by Leveling and Gravity Data. Tectonophysics, 97, 251-264. Heck, B. and H. Mälzer (1986): On Some Problems Connected with the Determination of Recent Vertical Crustal Movements from Repeated Leveling and Gravity Measurements. Tectonophysics, 130, 299305. Hein, G. (1986): Integrated Geodesy - State-of-the-Art 1986 Reference Text. In: H. Sünkel (ed.), Mathematical and Numerical Techniques in Physical Geodesy, Springer Verlag, 505-548. Heiskanen and Moritz (1967): Physical Geodesy. Freeman Publ. Co., San Francisco. Kanngieser, E. (1983): Application of Least-Squares Collocation to Gravity and Height Variations Associated with a Recent Rifting Process. Tectonophysics, 97, 265-277. Koch, K.-R. (1987): Parameter Estimation and Hypothesis Testing in Linear Models. Springer Verlag. Krarup, T. (1969): A Contribution to the Mathematical Foundation of Physical Geodesy. Danish Geodetic Institute Publ. No. 44, Copenhagen. Lauritzen, S. (1973): The Probabilistic Background of Some Statistical Methods in Physical Geodesy. Danish Geodetic Institute, Publ. No. 48, Copenhagen. Milbert, D. (1988): Treatment of Geodetic Leveling in the Integrated Geodesy Approach. The Ohio State University, Department of Geodetic Science and Surveying, Report No. 396, Columbus, Ohio. Moritz, H. (1973): Least Squares Collocation. Deutsche Geodätische Kommission, Reihe A, Heft Nr. 75. Moritz, H. (1978): The Operational Approach to Physical Geodesy. OSU Department of Geodetic Science, Report No.77. Moritz, H. (1980): Advanced Physical Geodesy. Wichmann Verlag, Karlsruhe. Reilly, W. (1981): Complete Determination of Local Crustal Deformation from Geodetic Observations. Tectonophysics, 71, 111-123. Rossikopoulos, D. (1986): Integrated Control Networks. Ph.D. Dissertation, Dept. of Geodesy and Surveying, University of Thessaloniki (in Greek). Sansó, F. (1986): Statistical Methods in Physical Geodesy. In: SŸnkel H. (ed.) Mathematical and Numerical Techniques in Physical Geodesy, Springer, 49-155. Sansó, F. (1978): The Minimum Mean Square Estimation Error Principle in Physical Geodesy (Stochastic and Non-Stochastic Interpretation. Proceedings 7th Symposium on Mathematical Geodesy, Assissi, 1978. Sansó, F. and A. Dermanis (1981): A geodynamic boundary value problem. 8th Hotine Symposium on Mathematical Geodesy, September 1981, Como, Italy. Bollettino di Geodesia e Scienze Affini, XLI, 1982, 1, 6587. Truesdell, C. (1977): A First Course in Rational Continuum Mechanics, vol. 1. Academic. Dermanis, A. (1985a): The Role of Frame Definitions in the Geodetic Determination of Crustal Deformation Parameters. Bulletin Géodésique, 59, 247-274. Dermanis, A. (1985b): Optimization Problems in Geodetic Network with Signals. 3rd Course of the International School of Advanced Geodesy "Optimization and Design of Geodetic Networks", 'Ettore Majorana' Centre for Scientific Culture, Erice, Sicily, April 25 May 10, 1984. In: E. Grafarend and F. Sanso (eds.): "Optimization and Design of Geodetic Networks", Springer Verlag, pp. 221-256. Dermanis, A. (1986): Integrated Models and Combined Adjustment Versus Sequential and Single Solutions. In: H. Ebner, D. Fritsch, G.W. Hein (eds): Minutes of the Joint Workshop on "Combined Adjustment of Heterogeneous Geodetic and Photogrammetric Data", International Association of Geodesy Special Study Groups 1.73 & 4.60, International Society of Photogrammetry and Remote Sensing Working Group III/1. Universität der Bundeswehr, Munich, September, 22-24, 1986. Dermanis, A. (1987): Geodetic Applications of Interpolation and Prediction. International School of Geodesy "A. Marussi". 4th Course: "Applied and Basic Geodesy: Present and Future Trends". Ettore Majorana Centre for Scientific Culture, Erice-Sicily, 15-25 June 1987. Eratosthenes, 22, 229-262. Dermanis, A. (1991): A Unified Approach to Linear Estimation and Prediction, pp. 65. Presented at the 20th General Assembly of the International Union of Geodesy and Geophysics, Vienna, August 1990. Dermanis, A., E. Livieratos, D. Rossikopoulos, D. Vlachos (1981): Geodetic Prediction of Crustal Deformations at the Seismic Area of Volvi. Proceedings International Symposium "Geodetic Networks and Computations", Munich, 1981. Veröffentlichungen Deutsche Geodätische Kommission, Reihe B, Nr. 258/V, 234-248. Dermanis, A. and D. Rossikopoulos (1988): Modelling Alternatives in Four-Dimensional Geodesy. Proceedings of the International Symposium "Instrumentation, Theory and Analysis for Integrated Geodesy", Sopron, Hungary, May 16-20, 1988, Vol. 2, 115-145. Dermanis, A. and D. Rossikopoulos (1991): Statistical Inference in Integrated Geodesy. Presented at the 20th General Assembly of the International Union of Geodesy and Geophysics, Vienna, August 1990. Dermanis, A. and A. Fotiou (1992): Adjustment Methods and Applications (in Greek). Editions Ziti, Thessaloniki. Dermanis, A. and E.W Grafarend (1992): The Finite Element Approach to the Geodetic Computation of Two- and Three-dimensional Deformation Parameters: A Study of Frame Invariance and Parameter Estimability. Presented at the International Conference "Cartography-Geodesy", 5th Centenary of the Americas: 1492-1992, Maracaibo, Venezuela, 24 November - 3 December, 1992. Eeg, J. and T. Krarup (1975): Integrated Geodesy. In: Brosowski, B. and E.Martensen (eds.), Methoden und Verfahren der Mathematischen Physik, Band 13, Mathematical Geodesy, Part II, Bibliographisches Inst., Mannhein. 17 Acknowledgement This work was prepared during the author's stay at the Department of Earth Sciences, Ibaraki University. Part of the paper has been contained in lectures presented to the Department of Earth Sciences, Ibaraki University and a lecture at the Department of Geophysics, University of Kyoto. The author would like to thank his host Prof. Yoichiro Fujii for motivating discussions and his valuable comments. Financial support of the author's stay in Japan by the Japanisches-Deutsches Zentrum Berlin is gratefully acknowledged. 18