\pretolerance=1400 \magnification=\magstep1 %\baselineskip=0.77cm file publ/lzrazky/precip6.tex nic s tym zatial nerobte len si to precitajte a poslite mi mi nejaky draft vasej casti ked to bude v nejakej finalnejsie podobe tak to dam do wordu alebo html, ale ozaj neviem co so vzorcami lebo s tym nemam cas sa prplat vo worde {\bf Precipitation by multivariate RST } \bigskip \noindent HELENA MITASOVA Department of Geography, University of Illinois at Urbana-Champaign, Urbana, IL 61801, U.S.A. \noindent LUBOS MITAS National Center for Supercomputing Applications, University of Illinois at Urbana-Champaign, Urbana, IL 61801, U.S.A. \bigskip \noindent {\bf 2.1 Multivariate spline interpolation} \smallskip Radial basis functions (multiqudrics, splines) belong to the most accurate interpolation and approximation methods for scattered data (REF). Because of its accuracy, flexibility, multivariate formulation, as well as implementation in GIS, we have selected the Regularized Spline with Tension (RST) as one of the methods which may be suitable for interpolation of precipitation in a region with significant variability in terrain. The RST method has been described by Mitas and Mitasova 1993 (an earlier version called Completely Regularized Spline) and Mitasova et al. 1995 (the general d-dimensional formulation with applications for 2D,3D,and 4D data). Therefore, here we briefly recall only the basic principles and focus on new developments and issues specific to interpolation of precipitation data. In general, a spline function $S({\bf x})$ fulfills the condition of minimizing the deviation from the measured points and at the same time its smoothness seminorm $I(S)$ $$ \sum_{j=1}^N |p^{[j]}-S({\bf x}^{[j]})|^2w_j + w_0I(S)={\sl minimum}\, . \eqno(1) $$ \noindent where $p^{[j]}$ are the values measured at discrete points ${\bf x}^{[j]}=(x_1^{[j]},x_2^{[j]}, \dots ,x_d^{[j]}), \; j=1,\dots ,N$ within a certain region of a {\sl d}-dimensional space, $w_j,w_0$ are positive weighting factors and $I(S)$ is the measure of smoothness (smooth seminorm) or roughness penalty. For $w_j/w_0 -> 0$ the function $S({\bf x})$ passes through the data points and works as an interpolation function, for $w_j/w_0 > 0$ the function can deviate from the data points and $S({\bf x})$ is an approximation function. The general solution of the minimization problem given by Eq. (1) can be expressed as a sum of two components (Talmi and Gilat 1977) $$ S({\bf x})=T({\bf x})+\sum_{j=1}^N \lambda_j R({\bf x},{\bf x}^{[j]}) \eqno(2) $$ \noindent where $T({\bf x})$ is a 'trend' function and $R({\bf x},{\bf x}^{[j]})$ is a {\sl radial basis function} with an explicit form depending on the choice of the $I(S)$. The smoothness seminorm $I(S)$ for the RST method has been designed to fulfill the following conditions: \item {a)} $S({\bf x})$ can be derived in an explicit form; \item {b)} $S({\bf x})$ synthesizes in a single function desired properties of several previously known splines (thin plate spline: Duchon 19XX, spline with tension: Franke 1985, Mitas and Mitasova 1988, Hutchinson and xx 1989?, regularized spline: Mitas and Mitasova 1988). \noindent The seminorm which fulfills these requirements includes derivatives of {\sl all orders} with weights rapidly decreasing with the increasing derivative order (Mitas 1993 unpublished manuscript) and for $d=2$ has the following form: $$ I^2(S)=\sum_{\bf \alpha} B_{\bf \alpha} \int \int _{\Omega} \left[ {\partial^{|{\bf \alpha}|}\over \partial x_1^{\alpha_1} \partial x_2^{\alpha_2}} \, S({\bf x})\right]^2 dx_1 dx_2 \eqno(3) $$ \noindent where ${\bf \alpha} =(\alpha_1,\alpha_2)$ is a multiindex ($\alpha_1=0, 1, 2, ...., \, \alpha_2=0, 1, 2, ...)$, with $|{\bf \alpha}|= \alpha_1 + \alpha_2$ and $\{ B_{\bf \alpha}\} $ are nonnegative constants, in our case $$ B_{\bf \alpha}=\cases {\,\quad 0\; , \qquad\qquad |{\bf \alpha}| =0 \cr \quad\cr \quad {|{\bf \alpha}|!\over {\alpha_1! \alpha_2!}} {1. \over {\varphi ^{2|{\bf \alpha}|}\; (|{\bf \alpha}|-1)!}}\; , \quad |{\bf \alpha} |>0 \cr} \eqno (4) $$ \noindent where $\varphi$ is a relative reciprocal weight of particular terms in the sum (generalized tension) which provides the control over the influence of derivatives of certain order on the resulting function. With this particular choice of coefficients, the radial basis functions can be found explicitly and have the following form for d=2,3,4 (Mitas and Mitasova 1999? Mitasova et al 1995) $$ R_2({\bf x},{\bf x}^{[j]}) = R_2(r) = - \left[ {\rm E_1} (\varrho) + \ln\varrho + C_E \right] \, \quad d=2 \eqno(5) $$ $$ R_3({\bf x},{\bf x}^{[j]}) = R _ 3(r)= \sqrt{{\pi \over \varrho}} \;{\rm erf} \left( \sqrt{\varrho} \right) -2 \quad d=3 \eqno(6) $$ $$ R_4({\bf x},{\bf x}^{[j]})=R_4(r)= {1 -e^{-\varrho} \over \varrho } - 1 \quad d=4 \eqno(7) $$ \noindent where $\varrho = (\varphi r/2)^2$, $r^2=\sum_{i=1}^d (x_i-x_i^{[j]})^2$ is the squared distance, $\varphi$ is the generalized tension, $C_E=0.577215...$ is the Euler constant, ${\rm E}_1(.)$ is the exponential integral function and erf(.) is the error function (Abramowitz and Stegun 1964). The coefficients $a_1, \{\lambda _j\}$ are obtained by solving the following system of linear equations $$ a_1+\sum_{j=1}^N \lambda_j \left[R({\bf x}^{[i]},{\bf x}^{[j]}) +\delta_{ji}w_0/w_j \right] =z^{[i]}\, ,\quad i=1,\ldots,N \eqno (8) $$ $$ \sum_{j=1}^N \lambda _j =0 \eqno (9) $$ \noindent where $w=w_0/w_j???$ is the smoothing parameter. The computational demands due to the solution of this system have been (and often still are ) cited as a major limitation for the use of splines (REF) However, the problem has been satisfactorilly resolved by a segmentation procedure implemented in GRASS GIS version of the RST, which is being routinely used to interpolate data sets with hundreds of thousands of points (Mitasova et al. 1995, Mitas and Mitasova 1999, Hargrove, ...REF). \medskip \noindent {\bf Anizotropy } The RST interpolation functions are (a) invariant to a rotation of coordinate space because the basis functions depend only on distance; (b) {\sl not} scale invariant, and a change of scale is equivalent to change in the tension parameter $\varphi$. The second property can be used for implementation of anisotropy by rotating the coordinate system by an angle $\theta$ representing the direction of anisotropy and then rescaling one axis according to the anisotropy magnitude. For $d=2$, the transformation from original coordinates ${\bf x}=(x_1,x_2)$ to new coordinates ${\bf x'}=(x'_1,x'_2)$ is simply $$ x'_1 = x_1 cos(\theta) + x_2 sin(\theta) \eqno (12) $$ $$ x'_2 = -x_1 sin(\theta) + x_2 cos(\theta) \eqno (13) $$ \noindent and the distances are then computed as $$ r^2 = s (x_1'-x_1'^{[j]})^2+ (x_2'-x_2'^{[j]})^2 \eqno (14) $$ \noindent where $s=s_1/s_2$ is a scaling coefficient expressed here as a ratio between the scaling in $x'_1$ and $x'_2$ axis respectively. The rotated and rescaled distances are then used in the solution of the system of equations (8) as well as for computation of interpolated values in grid points. The direction and scaling of the anisotropy can be found using standard geostatistical tools (REF). The generalization of the coordinate transformation for $d=3$ is straightforward. \medskip {\bf Interpolation with influence of additional variable} To interpolate precipitation with the influence of topography we can use an approach similar to the one proposed by Hutchinson and Bischof (1983)(OVER) Given the $N$ values of precipitation $p^{[j]}$ measured at discrete points ${\bf x}^{[j]}=(x_1^{[j]},x_2^{[j]},x_3^{[j]}), \; j=1,\dots ,N$ within a certain region of a {\sl 3}-dimensional space we compute a function $p=F(x_1,x_2)$ (as a grid) representing the spatial distribution of precipitation over the terrain surface $x_3=G(x_1,x_2)$ as $$ p=F(x_1,x_2)=S(x_1, x_2, cG(x_1,x_2)) \eqno (13) $$ \noindent where $c$ is vertical rescaling parameter (vertical tension), and $S(.)$ is the trivariate RST function. Equation (13) can be interpreted as an intersection of volume model of spatial distribution of precipitation (modeled by trivariate RST) with terrain surface. For this approach it is not necessary that the correlation between the elevation and precipitation exists for the entire region - the only condition is that it is sufficiently represented by data and changes smoothly (correlation is only local). \medskip \noindent {\bf Control parameters, their importance, tuning and optimization} One of the advantages of the RST method is its flexibility (adjustability to various types of phenomena) within the single radial basis function (as opposed to the rather complex and often subjective task of selecting the proper semivariogram in kriging). While most applications require only slight adjustment of tension parameters there are applications (such as precipitation in mountanious regions) when a wider set of tunable parameters can improve the predictive accuracy. We investigated the impact of the following parameters on the predictive accuracy of RST: \item {a)} horizontal uniform tension $\varphi$, \item {b)} uniform smoothing $w$, \item {c)} anisotropy: rotation and scale ($\theta,s$), \item {d)} vertical tension $c$ \item {e)} level of detail (smoothing) of 3-rd variable TOTO DALEJ treba napisat podla toho co bude vo vysldekoch - explain what each of the parameters does The parameters can be selected empirically, based on the knowledge of the modelled phenomenon, or automatically by minimization of the predictive error estimated by a cross-validation procedure (Mitasova et al 1995). (as we show later? minimum CV does not necessarily mean the best result? daj do discussion/zaverov - consistency and good results from several measures are important) Discuss normalization issue and role of tension as a rescaling parameter?. \medskip \noindent {\bf Cross-validation} Napiste ako ste to robili - toto je text z nasho IJGIS clanku 1995 We have used an ordinary cross-validation to find both optimal tension and smoothing. If we assume that $\{w_j\}$ are the same for all data points then the smoothing parameter is $w=w_j/w_0$. Let $S_{N,\varphi,w}^{[k]}$ be the smoothing spline with tension using all the data points except the k-$th$. We can take the ability of $S_{N,\varphi,w}^{[k]}$ to predict the missing data point $z_k$ as a measure of predictive capability of the spline with the given tension $\varphi$ and smoothing $w$, which can be represented by a mean error $V_0$ $$ V_0(\varphi,w)= {1 \over N} \sum _{k=1}^N |S_{N,\varphi,w}^{[k]} ({\bf x}^{[k]})-z^{[k]} |. \eqno(15) $$ \noindent Then the optimal values of tension $\varphi_O$ and smoothing $w_O$ are found by minimizing $V_0(\varphi,w)$. The cross-validation error computed for each point can also be used for the analysis of spatial distribution of predictive error and location of areas where this error is high and denser sampling is needed. Maybe include some comments on: many of the geostatistical concepts can be exploited within the spline framework. Try to fit the RST RBF to variogram? -see whether Roling has done that. \bigskip \noindent {\bf References} \smallskip \smallskip\noindent Abramowitz, M., and Stegun, I.A., 1964, {\sl Handbook of Mathematical Functions} (Dover: New York), 297-300, 228-231. \smallskip\noindent Compton, W., and Miller, D., 1994, Extending environmental modeling into fifth dimension: Multiple concurrent phenomena. {\sl Proceedings of the II. Int. Conference on Integration of Environmental Modeling and GIS}, Breckenridge, Colorado, in press. \smallskip\noindent Dietrich, C.R., and Osborne, M.R., 1991, Estimation of covariance parameters in kriging via restricted maximum likelihood. {\sl Mathematical Geology} {\bf 23}, 119-135. \smallskip\noindent Franke, R., 1987, Recent advances in the approximation of surfaces from scattered data. {\sl Topics in Multivariate Approximation}, edited by Chui, Schumaker and Utreras (Academic Press), 79-98. \smallskip\noindent {\sl GRASS4.1 Reference Manual}, 1993, U.S.Army Corps of Engineers, Construction Engineering Research Laboratories, Champaign, Illinois, 422-425. \smallskip\noindent Hutchinson, M.F. ,1991, The Application of thin plate smoothing splines to continent-wide data assimilation. {\sl Data Assimilation Systems}, edited by Jasper, J.D., BMRC Research Report N.27, Bureau of Meteorology, Melbourne, 104-113. \smallskip\noindent Hutchinson, M.F., 1994, Interpolating rainfall means - getting the temporal statistics correct. {\sl Proceedings of the II. Int. Conference on Integration of Environmental Modeling and GIS}, Breckenridge, Colorado, in press. \smallskip\noindent Hutchinson, M.F., and Bischof, R.J., 1983, A new method for estimating the spatial distribution of mean seasonal and annual rainfall applied to the Hunter Valley, New South Wales. {\sl Australian Meteorological magazine} {\bf 31}, 179-184. \smallskip\noindent Hutchinson, M.F. and Gessler, P.E., 1994, Splines - more than just a smooth interpolator. {\sl Geoderma} {\bf 62}, 45-67. \smallskip\noindent Mitas, L. and Mitasova, H., 1988, General variational approach to the interpolation problem. {\sl Computers and Mathematics with Applications}, {\bf 16}, 983-992. \smallskip\noindent Mitas, L. and Mitasova, H., 1994, Multivariate approximation for scattered data. (to be published). \smallskip\noindent Mitasova, H., 1992, 1993, 1994, Surfaces and Modeling. {\sl GRASSClippings}, {\bf 6}/3, {\bf 7}/1,2,3, p. 16-18, 18-19, 18-19, 25-26. \smallskip\noindent Mitasova, H., and Mitas, L., 1993, Interpolation by Regularized Spline with Tension: I. Theory and implementation. {\sl Mathematical Geology }, {\bf 25}, 641-655. \smallskip\noindent Nielson, G.M., Foley, T.A., Hamann, B., and Lane, D., 1991, Visualizing and Modeling Scattered Multivariate Data. {\sl IEEE Computer Graphics and Applications} {\bf 11}, 47-55. \smallskip\noindent Talmi, A., and Gilat, G., 1977, Method for Smooth Approximation of Data. {\sl Journal of Computational Physics} {\bf 23}, 93-123. \smallskip \noindent {\sl Trends in Nitrogen in the Chesapeake bay, (1984-1990)}, 1992, Report CBP/TRS 68/92, Chesapeake Bay Program. \smallskip\noindent Wahba, G., 1990, Spline models for observational data. {\sl CNMS-NSF Regional Conference series in applied mathematics} {\bf 59}, (Philadelphia, Pennsylvania : SIAM). \bye