Human influence on sub-regional surface air temperature change over India

Human activities have been implicated in the observed increase in Global Mean Surface Temperature. Over regional scales where climatic changes determine societal impacts and drive adaptation related decisions, detection and attribution (D&A) of climate change can be challenging due to the greater contribution of internal variability, greater uncertainty in regionally important forcings, greater errors in climate models, and larger observational uncertainty in many regions of the world. We examine the causes of annual and seasonal surface air temperature (TAS) changes over sub-regions (based on a demarcation of homogeneous temperature zones) of India using two observational datasets together with results from a multimodel archive of forced and unforced simulations. Our D&A analysis examines sensitivity of the results to a variety of optimal fingerprint methods and temporal-averaging choices. We can robustly attribute TAS changes over India between 1956–2005 to anthropogenic forcing mostly by greenhouse gases and partially offset by other anthropogenic forcings including aerosols and land use land cover change.


Supplementary information
Human influence on sub-regional surface air temperature change over India Dileepkumar R 1     shown along with IITM observed dataset (solid black line), and best-estimate contributions from ANT (magenta) and NAT (green). Shaded region centered on the observations shows the uncertainty range due to internal variability (two-sigma pentadal variability computed from ! !! ).    represent the "best estimate" and the 5 -95% range for the estimated amplitude of GHG, OA, and NAT forcings respectively for each number of EOFs retained in the truncation. The red and green dotted lines in panel d of I and II represent the 5 -95 % confidence level. The RCC fails in I, and passes in II with maximum EOF number as 7. For truncations ≤ 7, the Cumulative ratio model/Observed variance (solid blue line) ranges around unity shows that the model variability is consistent with observed, but for greater truncation levels it falls outside the 5-95 % confidence band. The scaling factor corresponding to EOF 7 was selected as the signal amplitude. Tables   Table S1. CMIP5 models 4 used in Optimal Fingerprint analysis. Model names, data resolution, number of initial condition ensemble members for each experiment and number of years available for each model's pre-industrial control run also listed. The last column shows the forcing factor accounted for in addition the model runs. Y indicates factor included and N indicates factor not included. Table S2. Area weights for the different homogeneous zones used to compute signal amplitude of "All Regions". Table S3. Observed and attributable trends (°C per period length given as 5-95% ranges) over "All    was selected as the signal amplitude.

Model
Horizontal resolution (lat x lon)    Table S3. Observed and attributable trends (°C per period length given as 5-95% ranges) over "All Region" from the three-signal analysis of TAS change.

Band-pass filter details:
We applied a band-pass filter used in Santer et al., 2011 (ref. 6) to observed and modelled detrended TAS monthly anomalies in order to compare the observed and modelled variabilities in the frequencies of interest. The filtering allows us to focus on variability on 10-year timescales, with half power points at 5 and 20 years.

Processing of data:
The near surface temperature (TAS) from observations and CMIP5 simulations are processed as described below.  (Table S2) data of all seven homogeneous regions (denoted "All Regions"). For the "All Region" case the temporal centering carried out after the concatenation of area weighted datasets . We denote the resulting vector as X ANT (for MME of the historicalAnthro simulation), X NAT (for historicalNat simulation), X GHG (for historicalGHG) etc. The columns of X were given by X ANT & X NAT for two signal case. For the three signal case we used a linear transformation ! for deriving the !′ (a linear combination of GHG, OA, and NAT) from ! (a linear combination of historicalGHG, historical, and historicalNat).

The observation vector
Then the corresponding scaling factors for (!"!, !", !"#) and The two noise covariance matrices C N1 & C N2 were estimated from piControl dataset as follows: Given that the different models have piControl runs of varying lengths, we extracted 500 years (length of the shortest piControl run available) from the end of the piControl run available from each model. We then calculated the spatial average over each homogeneous region as with the individual forcing runs. Then we compute the seasonal averages (Annual, DJF, MAM, JJA, and SON) followed by decadal and pentadal means. We then extract all 10-decade long segments that overlap by all but 1 decade. We get 40 segments from each model's piControl time series of which, we choose the first 20 segments for constructing C N1 and the latter 20 segments for C N2 .
We repeat the above procedure for each model's piControl adding rows to C

Truncation level and Residual Consistency Check (RCC):
Increasing the truncation level (EOF number) will introduce unrealistically low variance. Similarly if choosing small truncation level will give uncertainty estimate are unreliable. Before drawing any further conclusions, therefore, we need to establish the maximum truncation at which the model is reliable. The residual consistency check tests the hypothesis that model simulated internal variability is equal to observed.

Residual consistency check for OLS & TLS method:
The F-test based on ratio of model internal variance to observed residual variance is used. The red dots in Fig. S8 I & II (a, b, c) show the evolution of ! − ! ! ! , where !, is the number of EOFs and ! is the pattern number, and ! ! is the residual (! ! follows ! − ! ! !!!,! distribution, where ! denotes the degrees of freedom ). Therefore ! − ! ! ! follows the