Temperature grid dataset for climate monitoring based on homogeneous time series in Switzerland

Transcription

1 Federal Department of Home Affairs FDHA Federal Office of Meteorology and Climatology MeteoSwiss Temperature grid dataset for climate monitoring based on homogeneous time series in Switzerland F. A. Isotta, M. Begert and C. Frei 30th October 2015

2 Content Introduction: motivation, method Results and evaluation Conclusion and outlook 2

3 Content Introduction: motivation, method Results and evaluation Conclusion and outlook 3

4 Introduction - Motivation Develop new datasets for monthly temperature and precipitation suitable for climate monitoring (regularly updated) (1864-) (-now) and (-now) Only with homogenized station data Continuous measurements (no gaps) Constant station density (same stations every time step) 4

5 Introduction - Motivation The amount of stations fulfilling all requirements is low Temperature Precipitation T, , 57 stations T, , 28 stations T, , 18 stations P, , 336 stations P, , 39 stations P, , 14 stations 5

6 Introduction - Method + High-resolution component , ~85 stations 2-km grid dataset Long-term component stations (continuous) 6

7 RSOI - Overview Reduced Space Optimal Interpolation (Kaplan et al., 1997; Schmidli et al. 2001, 2002; Schiemann et al., 2010; Masson et al., 2015) PCA + High-resolution component CALIBRATION PERIOD ( ) (anomalies, data transformation if needed) PC loadings Dimensionality reduction (truncation) Station data (sparse network) CALIBRATION PERIOD ( ) (transformed if needed) Station data (sparse network) RECONSTRUCTION PERIOD ( ) (transformed if needed) + OI Reconstruction (sparse network) RECONSTRUCTION PERIOD ( ) (back-transformed if needed) 7

8 Content Introduction: motivation, method Results and evaluation Conclusion and outlook 8

9 RSOI Results and evaluation Calibration period: Reconstruction period: , , Dimensionality reduction (truncation): 12 Evaluation: Tests with changing calibration (length and period), truncation, data quality, stations amount Use of crossvalidation (leave-one-out): x i,rrrrrrrr, x i,ooo Mean absolute error (MAE) MMM = 1 n x i,rrrrrrrr x i,ooo Mean-Squared Error Skill Score (MSESS) i=1 1= perfect reconstruction, 0=no skill MSSSS = 1 n x 2 i=1 i,rrrrrrrr x i,ooo 2 n x Variance i=1 i,ooo x i,ooo Trend n 9

10 PCA PC loading 1 91% PC loading 2 5% PC loading 3-1% PC loading 4 1% 10

11 Reconstruction examples (anomalies ) Tanomalies Tanomalies 11

12 Reconstrucion examples (anomalies ) Direct interpolation Reconstruction July 1976 ALL MSESS 57 stations 28 stations stations MSESS 28 stations MSESS 28 stations

13 Reconstrucion examples (anomalies ) Direct interpolation Reconstruction October 1962 ALL MSESS 57 stations 28 stations stations MSESS 28 stations MSESS 28 stations

14 Mean absolute error (degc) n MMM = 1 n x i,rrrrrrrr x i,ooo i=1 t Grid # stat ALL DJF MAM JJA SON # Reconstr Reconstr Direct grid ~ Reconstr Reconstr stat. 18 stat. 14

15 Mean absolute error (degc) n MMM = 1 n x i,rrrrrrrr x i,ooo i=1 t Grid # stat ALL DJF MAM JJA SON # Reconstr Reconstr Direct grid ~ Reconstr Reconstr stat. 18 stat. 15

16 Mean absolute error (degc) n MMM = 1 n x i,rrrrrrrr x i,ooo i=1 t Grid # stat ALL DJF MAM JJA SON # Reconstr Reconstr Direct grid ~ Reconstr Reconstr stat. 18 stat. 16

17 Mean absolute error (degc) n MMM = 1 n x i,rrrrrrrr x i,ooo i=1 t Grid # stat ALL DJF MAM JJA SON # Reconstr Reconstr Direct grid ~ Reconstr Reconstr stat. 18 stat. 17

18 Mean absolute error (degc) n MMM = 1 n x i,rrrrrrrr x i,ooo i=1 t Grid # stat ALL DJF MAM JJA SON # Reconstr Reconstr Direct grid ~ Reconstr Reconstr stat. 18 stat. 18

19 Skill: MSESS 1901/ Explained temporal variance Most of the stations have MSESS > , 57 stations , 28 stations 19

20 Skill: MSESS Explained spatial variance Expl. variance calibration period Explained variance fraction (28 stations) MSESS ALL DJF MAM JJA SON Median q q

21 Variance Variance in time, 28 stations Reconstructed Observed 21

22 Trend Theil-Sen trend estimate (degc/y) Stippling: statistically not significant (0.05, Mann-Kendall) Non homogenized data Homogeneous data (28 stations) Direct interpolation Homogeneous data (57 stations) 22

23 Trend 1864/ Theil-Sen trend estimate (degc/y) Stippling: statistically not significant (0.05, Mann-Kendall)

24 Standard error Standard deviation of difference reconstruction vs. direct interpolation 24

25 Content Introduction: motivation, method Results and evaluation Conclusion and outlook 25

26 Conclusion and outlook RSOI method RSOI is an attractive method to benefit of short-term high-resolution information to reconstruct longer time scales with less observations available. Method suitable for complex terrain where variations are spatially anchored. Successful reconstruction of time series and spatial distribution of temperature The discrepancies between observations and reconstruction are relatively moderate (MAE 0.25) Reconstruction improves long-term consistency Outlook Additional analysis (compare with HISTALP, ) Develop a regularly updated climate monitoring product at MeteoSwiss Apply same method for precipitation fields (station homogenization ongoing) Potential for application in the entire Alpine Region 26

27 27

28 RSOI Details Optimal interpolation Find scores a ti (reconstruction period t i ) minimizing the cost function S Gauge measurement (sparse network k) Matrix with L eigenvalues of the covariance matrix in the diagonal (from PCA) S(a ti ) = (H a ti x o t i ) T R 1 (H a ti - x o t i ) + a ti T C 1 a ti = e ti r = Ho e ti t + e ti o Covariance Matrix of e ti r (estimated from calibration period) Truncation error (dimension reduction) Error due to difference between gauge measurement and grid-cell value assigned (H o ) Guarantee balance between regions with different station density and lowers weight of highly correlated gauges. Disfavour high scores for high-order PC loadings 28

29 RSOI Details x tc = E. a n +e tc t o x tc = H o o. x tc + e tc = H o. E. a tc +H o t o r. e tc + e tc = H. a tc + e tc Gauge measurement (sparse network k) Find scores a ti (reconstruction period t i ) minimizing the cost function S Gauge measurement (sparse network k) calibration period t c Matrix with L eigenvalues of the covariance matrix in the diagonal (from PCA) S(a ti ) = (H a ti x ti o ) T R 1 (H a ti x ti o ) + a ti T C 1 a ti = e ti r = Ho e ti t + e ti o Covariance Matrix of e ti r (estimated from calibration period) 29

30 Trend (degc/y) Reconstructed Reconstruction 57 stations Reconstruction 28 stations Inhomogeneous data Observed 30