The present-day gridded data used to develop the ANN were from the NCEP-NCAR re-analysis (Kalnay et al., 1996). Climate PD-0332991 molecular weight Explorer (KNMI, 2012) was used to extract area averaged predictor variables from the gridded reanalysis data for 1948–2012 and over a region (17–19 N, −80.0 to −76.0 W) that encompassed Jamaica. The list of predictors investigated is shown in Table 2. Indices representing the North Atlantic Oscillation (NAO) and El Niño South Oscillation (ENSO) were also added as predictors based on their known influence on Jamaican rainfall (Taylor et al., 2002 and CSGM, 2012). Feed-forward ANNs with input, two hidden and output layers were constructed (see Appendix

A). Parameters of the calibrated, verified and corrected ANN were applied to the re-analysis data to derive predictions for the 1, 2, 5 and 10 day precipitation from 1950 to 1991. The respective AMS data was then defined and applied to the gaps in the long duration data. A frequency analysis with parameters with temporal trends was used to estimate the 24-h duration future climate intensities to 2100. Only the 24-h durations were examined because, firstly, the presence of step changes detected in the 2, 5 and 10 day durations are impossible to reliably duplicate into the future. Secondly, since the 24 h durations events were defined primarily from aggregation of original rainfall data, versus being supplemented with infilled data, this limits

the influences of errors in the infilling processes and focuses the analysis on the original check details precipitation data. It should be noted that the non-existence of step changes in the 24-h durations is not the only concern as cyclical and other non-linear

signals can be present and affect the location, scale and shape of the distribution with time (Hall and Tajvidi, 2000). Parameters were defined for the Weibull PDF using Likelihood Method, similar Tideglusib to Cooley (2009), Chavez-Demoulin and Davison (2005), Maraun et al. (2009) and Ramesh (2005) with the linear temporal functions for the present period (1895–2010). Temporal scaling functions for the location (mean), scale (variance) and shape (skewness) parameters were used and follows a similar approach to Withers and Nadarajah (2000). Four variations of the linear temporal functions, using a linear trend in Eq. (3) were used to fit the AMS data: (i) stationary with time; (ii) mean varying; (iii) mean and scale varying; and (iv) mean, scale and skewness varying. This approach allowed for an exploration of the trends in individual statistical parameters that may best explain changes in intensities. At the end of this calibration step there were four models for each station that fit the 1895–2010 AMS data. Temporal extrapolation of the parameters of these models to 2100, using Eq. (3), was then undertaken to estimate the future climate values in the calibrated temporal scaling functions.