A full scale, non stationary approach for the kriging of large spatio(-temporal) datasets

Transcription

1 A full scale, non stationary approach for the kriging of large spatio(-temporal) datasets Thomas Romary, Nicolas Desassis & Francky Fouedjio Mines ParisTech Centre de Géosciences, Equipe Géostatistique Réseau RESSTE 11 mai 2015 T. Romary et al. (Mines ParisTech) RESSTE 1 / 25

2 Introduction Introduction Z(x, t) = µ(x, t) + S(x, t) second order spatio-(temporal) random field Inference Possible non-stationarity in the covariance structure Prediction complexity O(n 3 ), storage O(n 2 ) T. Romary et al. (Mines ParisTech) RESSTE 2 / 25

3 Introduction Notations n size of the dataset z = {z(x 1 ),..., z(x n )} : data C covariance fonction C covariance matrix associated to the data f deterministic covariates T. Romary et al. (Mines ParisTech) RESSTE 3 / 25

4 Introduction Universal kriging system Matrix form We have f 00 f 0 =. f 0L, F = Kriging system and variance f f 1L..... f n0... f nl Z = Λ t z, µ = µ 1. µ L ( ) ( ) C F Λ F t 0 µ = ( C0 f 0 ) and σ 2 UK = C(0) Λ t C 0 µ t f 0 T. Romary et al. (Mines ParisTech) RESSTE 4 / 25

5 Introduction Universal kriging system Z = C 0C 1 z + (f 0 C 0C 1 F)(F C 1 F) 1 F C 1 z Idea C = A + P BP t where A is sparse P is a n p matrix with p n Sherman-Woodbury-Morrison formula C 1 = A 1 A 1 P (B 1 + P t A 1 P ) 1 P t A 1 T. Romary et al. (Mines ParisTech) RESSTE 5 / 25

6 Introduction Universal kriging system Z = C 0C 1 z + (f 0 C 0C 1 F)(F C 1 F) 1 F C 1 z Idea C = A + P BP t where A is sparse P is a n p matrix with p n Sherman-Woodbury-Morrison formula C 1 = A 1 A 1 P (B 1 + P t A 1 P ) 1 P t A 1 T. Romary et al. (Mines ParisTech) RESSTE 5 / 25

7 Introduction Spatio-temporal framework Satellite / Remote sensing data {Z(x, t) : x X R d, t = 1,..., T }, d = 2, 3 Spatio-temporal mixed effect model (Katzfuss and Cressie [2011]) where µ mean function Z(x, t) = µ(x, t) + S(x, t) + S second order random function p ξ i (t)p i (x) Ξ t = (ξ t 1,..., ξt p) centered, square-integrable random variables P i (x) basis functions i=1 T. Romary et al. (Mines ParisTech) RESSTE 6 / 25

8 Introduction Spatio-temporal framework Temporal Structure µ(x, t) = f t (x) β t Ξ t = H t Ξ t 1 + u t, such that var(ξ t Ξ t 1 ) = B t S(x, t) = σ(x, t)ε xt, where ε is a Gaussian white noise Inference by EM algorithm, see Katzfuss and Cressie [2011] Prediction by kriging, filtering, smoothing T. Romary et al. (Mines ParisTech) RESSTE 7 / 25

9 Introduction Spatial-only framework Assume E(Z(x)) = 0 where Z(x) = S(x) + p ξ i P i (x) + ε(x) i=1 cov(z(x), Z(y)) = C(x, y) = D(x, y) + P (x) t BP (y) + σ 2 xδ xy cov(s(x), S(y)) = D(x, y) = C NS (x, y)c CS (x, y) C NS is a non stationary covariance function (Fouedjio et al. [2014]) C CS is a compactly supported covariance function (Furrer et al. [2006], Gneiting [2002]), with fixed range parameter T. Romary et al. (Mines ParisTech) RESSTE 8 / 25

10 Inference Inference Several steps 1 Adjust D 2 Select p basis function from a dictionary by minimizing 3 Estimate B z P β λ β 1 T. Romary et al. (Mines ParisTech) RESSTE 9 / 25

11 Inference D term Compute and fit a local kernel variogram estimator at a set of knots Interpolate the parameters over the domain to get D(x, y) = C NS (x, y)c CS (x, y) see Fouedjio et al. [2014] for more details T. Romary et al. (Mines ParisTech) RESSTE 10 / 25

12 Inference Selection of basis functions Dictionary Multi resolution basis made of columns of compactly supported covariance matrices (including anisotropic ones) Selection by LASSO (glmnet) λ is set by cross-validation z P β λ β 1 T. Romary et al. (Mines ParisTech) RESSTE 11 / 25

13 Inference Estimation of B 1 Compute a non-parametric estimator of the covariance at a set of knots Ĉ K (x, y) = 1 K ij (z(x i ))(z(x j )) Σ Kxy where K ij = K h (x x i, y y j ) and Σ Kxy = i,j K ij 2 Compute P = QR 3 Compute B = R 1 Q T (ĈK A θ)q(r 1 ) T i,j T. Romary et al. (Mines ParisTech) RESSTE 12 / 25

14 Example Example Synthetic data on [0, 1] 2, Gaussian RF, centered with covariance where C(x, y) = 0.3sph( x y, 0.1) + P (x) t BP (y) + 0.1δ xy P 50 basis functions taken at random among 736 B drawn from a Wishart distribution Comparison between full model, small scale only, basis functions only, true model range of the small scale term = 0.05 T. Romary et al. (Mines ParisTech) RESSTE 13 / 25

15 Example Reference realization T. Romary et al. (Mines ParisTech) RESSTE 14 / 25

16 Example Reference realization sample (5000) T. Romary et al. (Mines ParisTech) RESSTE 15 / 25

17 Example Building the dictionary Ingredients models: cubic, Wendland 2 ranges: 0.5, 0.3 angles: 0, π/6, π/3, π/2, 2π/3, 5π/6 736 basis functions in total T. Romary et al. (Mines ParisTech) RESSTE 16 / 25

18 Example Selection T. Romary et al. (Mines ParisTech) RESSTE 17 / 25

19 Example Kriging map: full model T. Romary et al. (Mines ParisTech) RESSTE 18 / 25

20 Example Kriging map: full model T. Romary et al. (Mines ParisTech) RESSTE 19 / 25

21 Example Kriging map: short scale only T. Romary et al. (Mines ParisTech) RESSTE 20 / 25

22 Example Kriging map: basis only T. Romary et al. (Mines ParisTech) RESSTE 21 / 25

23 Example Kriging map: true model T. Romary et al. (Mines ParisTech) RESSTE 22 / 25

24 Example Results Model MSE MSE (true) Full Taper only Basis only (36) True model T. Romary et al. (Mines ParisTech) RESSTE 23 / 25

25 Example Discussion Work in progress Lots of tuning: knots (basis functions, variogram estimation, covariance estimation) parameters of the basis functions bandwidth parameters Prediction variances T. Romary et al. (Mines ParisTech) RESSTE 24 / 25

26 Example References S. Banerjee, A. E. Gelfand, A. O. Finley, and H. Sang. Gaussian predictive process models for large spatial data sets. Journal of the Royal Statistical Society. Series B, 70(4): , N. Cressie and G. Johannesson. Fixed rank kriging for very large spatial data sets. Journal of the Royal Statistical Society. Series B, 70: , Francky Fouedjio, Nicolas Desassis, and Jacques Rivoirard. A generalized convolution model and estimation for non-stationary random functions. arxiv preprint arxiv: , R. Furrer, M. G. Genton, and D. W. Nychka. Covariance tapering for interpolation of large spatial datasets. Journal of Computational and Graphical Statistics, 15(3): , T. Gneiting. Compactly supported correlation functions. Journal of Multivariate Analysis, 83(2): , Matthias Katzfuss and Noel Cressie. Spatio-temporal smoothing and em estimation for massive remote-sensing data sets. Journal of Time Series Analysis, 32(4): , H. Sang and J.Z. Huang. A full scale approximation of covariance functions for large spatial data sets. Journal of the Royal Statistical Society. Series B, 74(1): , M. L. Stein. A modeling approach for large spatial datasets. Journal of the Korean Statistical Society, 37:3 10, T. Romary et al. (Mines ParisTech) RESSTE 25 / 25

27 Sparse approximation: Covariance Tapering Principle C tap (x, y) = C(x, y)c θ (x, y) where C θ (x, y) is a compactly supported kernel = C tap (x, y) generates a sparse matrix Complexity O(n), storage O(n) R. Furrer, M. G. Genton, and D. W. Nychka. Covariance tapering for interpolation of large spatial datasets. Journal of Computational and Graphical Statistics, 15(3): , 2006 T. Romary et al. (Mines ParisTech) RESSTE 26 / 25

28 Low rank approximation Principle C LR (x, y) = σ 2 δ xy + P (x) B P (y) where σ 2 is the variance of the nugget, P is a vector of p basis functions, B R p p with p n Complexity O(n p 2 ), storage O(n p) N. Cressie and G. Johannesson. Fixed rank kriging for very large spatial data sets. Journal of the Royal Statistical Society. Series B, 70: , 2008 S. Banerjee, A. E. Gelfand, A. O. Finley, and H. Sang. Gaussian predictive process models for large spatial data sets. Journal of the Royal Statistical Society. Series B, 70(4): , 2008 T. Romary et al. (Mines ParisTech) RESSTE 27 / 25

29 Combination Principle C(x, y) C tap (x, y) + P (x) B P (y) Inversion by the Sherman-Morrison-Woodbury formula Complexity O(n p 2 ), storage O(n p) M. L. Stein. A modeling approach for large spatial datasets. Journal of the Korean Statistical Society, 37:3 10, 2008 H. Sang and J.Z. Huang. A full scale approximation of covariance functions for large spatial data sets. Journal of the Royal Statistical Society. Series B, 74(1): , 2012 T. Romary et al. (Mines ParisTech) RESSTE 28 / 25