UNDERSTANDING THE CONNECTION BETWEEN THE HYDRO- CLIMATIC EXTREMES AND DIARRHEAL DISEASES OVER BENGAL DELTA: THE VULNERABILITY ASSESSMENT OF PAST, PRESENT AND FUTURE

Living in the age of space exploration and nanotechnology, the significant portion of human population still have threatened by diarrheal diseases throughout the globe. Being a major contributor of the global mortality, the diarrheal diseases account for an estimated 3.1% of the total burden of disease in terms of Disability-Adjusted Life Year (DALY) where cholera and rotavirus diarrhea comprise more than two-thirds of the diarrheal morbidity in developing countries of South Asia. Alongside with many more challenges like climate change or civil war, the capability to resolve the diarrheal disease burden in developing countries remains questionable. As the primary reasons for the disease transmission in epidemic scale are due to the exposure of contaminating pathogens via unsafe drinking water sources, lack of sanitation, deficient hygiene, insufficient drainage infrastructures and poor access to health care, ensuring clean water sources and improved sanitation may seem to untangle the problem. However, it will take a longer time to achieve such improvement by the developing countries as many of them are already missed the Millennium Development Goals (WHO/UNICEF 2015). Moreover, ongoing global climatic change also leads the disease vulnerability in much degrading state (Woodward et al. 2014). In this context, the Bengal delta of South Asia, exhibits the highest population density of globe and is one of the most vulnerable region of the world in terms of both climate change and diarrheal diseases (Bowen and Friel 2012). Therefore, the challenges to tackle the vulnerability of diarrheal disease under ongoing global warming is paramount in this region.

Previous studies found that the diarrhoeal diseases like cholera and rotavirus are significantly influenced by environmental factors in the developing counties of Asia and Sub Saharan Africa. The outbreaks of the diseases can occur in the wake of climatic extremes like heavy rainfall, flooding, cyclone, drought, extreme temperature and El-Nino (Bradley et al. 1996;Corwin et al. 1996;Patz et al. 2000;Vanasco et al. 2001;Chhotray et al. 2002;Kalashnikov et al. 2002;Qadri et al. 2005;Yang et al. 2005;Harmeling 2012).
However, most studies have explored the influence on disease transmission for particular climatic extremes or related natural disasters, but the integration of multiple variables along with disease cases is infrequently done. Moreover, a deterministic quantification of the diseases epidemic based on the hydro-climatic factors is absent in existing literatures.
In terms of diarrheal disease epidemic as well as climate vulnerability, the Bengal delta is frequently considered as one of the high-risk region of the globe (IPCC, 2014). The policy makers of the region not only need to tackle the burden of diarrheal disease but also need to understand the future impact of these diseases under ongoing climate change. However, not only the future assessment of the disease is challenging but also, the meaningfully quantification of climatic extremes under future climate change scenarios require robust assessment due to the absence of such kind of studies. Therefore, the objective is this dissertation is to develop the deterministic models that can project the future risk of diarrheal diseases, primarily rotavirus and cholera, driven by hydro-climatic extremes over the climate vulnerable region of Bengal delta. In order to achieve this objective, I developed a bias-correction method for the high-resolution regional climate models, generated an observed data set over the Bengal delta, formulated a deterministic epidemic model for rotavirus that accounts intra-annual variability, proposed a spatial risk model of rotavirus and cholera and projected the future of the diarrheal disease for 21 st century. The work has been described in the following three manuscripts, as per the Graduate School Manual guidelines: The objective of this work was to explore the climate and its extremes over a monsoon dominated country like Bangladesh by following the latest RCP (Representative Concentration Pathways) emission scenarios, considering fine scale regional physics, incorporating the uncertainties range, and also by conducting advance bias correction methods to accomplish most reliable future projections. In this relation, the article aimed to investigate (1) the future probabilistic climate of Bengal delta, using five regional climate model projections driven by GCM results, (2) to develop a new spatial gridded observed data that represents historical climate extremes set and (3) to implement the latest QM (Quantile Mapping) bias correction methods over multi-model RCM outputs.
In this manuscript, we investigated the role of climatic extremes on one of the prevalent diarrheal disease, rotavirus. The study aimed (1) to determine the effect of climatic extremes on the rotavirus epidemic over Bangladesh, both in spatially and temporal scale, (2) to evaluate the rotavirus patterns over the cities of South Asia to understand the relation of the virus to regional hydro climatic processes and (3) to implemented a deterministic multivariate modeling for risk assessment and integrating near real-time satellite products (with GPM for rainfall and MODIS for temperature).
The objective of this manuscript was to project the future the diarrheal disease risk based on the epidemic models driven by the bias-corrected regional climate models. To implement the long-term development medical initiatives under ongoing climate change, the policy makers requires comprehensive and meaningfully estimate of the future vulnerability of the diseases. Thus, the manuscript aimed (1) to develop some spatial multivariate models of the rotavirus and cholera epidemic over Bengal delta, (2) to assess the effect of relative humidity on rotavirus cycle, and (3) to project the probable future risk during the rising phase for both the diseases in the early, mid and late 21 st Century.
In conclusion, the diarrheal diseases are a recurrent burden in the developing world.
Though there are many factors such as population dynamic, poor water sanitation and hygiene can be responsible for diarrhreal outbreak in the region, the climate drivers still can plays a significant role in the diseases epidemic thus essential to pre-epidemic management. As this study proposed a risk based methodology rather than prevalence or incidence based method, the method can overlook the influence of the population infectivity the disease and can be utilize to detect the influence of climatic change. This will allow the relevant stakeholders to improve the decision-making process. The novel approach and result of this dissertation can be utilized as a guideline for long-term diseases preparedness or vaccination strategy for Bangladesh. High-resolution regional model results will also provide valuable insight to the disease burden estimation which can be implemented in sub-district level with appropriate stakeholder. The findings of this study will be shared with ICDDRB (International Centre for Diarrhoeal Disease Research, Bangladesh) and Bill & Melinda Gates Foundation for further improvement of the vaccination and surveillance strategy over the region.   Table 1. Description of selected regional climate models over Bangladesh……………145 Table 2: The results from the multivariate analysis, temporal analysis and spatial analysis for Rising phase of rotavirus (without considering relative humidity)…………………146 Table 3: The results from the multivariate analysis, temporal analysis and spatial analysis for Rising phase of rotavirus……………………………………………………………147 Table 4: The results from the multivariate analysis, temporal analysis and spatial analysis xv for the rising phase of fall cholera…………………………………………………….148 In the era of global warning, the insight of future climate and their changing extremes is critical for climate-vulnerable regions of the world. In this study, we have conducted a robust assessment of Regional Climate Model (RCM) results in a monsoon-dominated region within the new Coupled Model Intercomparison Project Phase 5 (CMIP5) and the latest Representative Concentration Pathways (RCP) scenarios. We have applied an advanced bias correction approach to five RCM simulations in order to project future climate and associated extremes over Bangladesh, a critically climate-vulnerable country with a complex monsoon system. We have also generated a new gridded product that performed better in capturing observed climatic extremes than existing products. The biascorrection approach provided a notable improvement in capturing the precipitation extremes as well as mean climate. The majority of projected multi-model RCMs indicate an increase of rainfall, where one model shows contrary results during the 2080s (2071-2100) era. The multi-model mean shows that nighttime temperatures will increase much faster than daytime temperatures and the average annual temperatures are projected to be as hot as present-day summer temperatures. The expected increase of precipitation and temperature over the hilly areas are higher compared to other parts of the country. Overall, the projected extremities of future rainfall are more variable than temperature. According to the majority of the models, the number of the heavy rainy days will increase in future years. The severity of summer-day temperatures will be alarming, especially over hilly regions, where winters are relatively warm. The projected rise of both precipitation and temperature extremes over the intense rainfall-prone northeastern region of the country creates a possibility of devastating flash floods with harmful impacts on agriculture.

Introduction
Observations show that the global land and ocean temperature has risen by 0.85 °C over the period of 1880 to 2012, and the warming trend has accelerated in the last 60 years (IPCC, 2013). Rising global temperatures have been accompanied by changes in the mean state of the climate as well as their associated extreme events. As climate extremes and weather events have significant impacts on the socio-economic stability and sustainability of any region, the information about their probabilistic future as well as existing understanding has received wide attention in the scientific community (Hartmann et al., 2013;IPCC 2007;IPCC 2013 (Revadekar et al. 2011;Rao et al. 2014;Freychet et al. 2015). To have a better understanding of socio-economic and hydrologic impacts, the perception of climate and its associated extremes over both upstream and downstream of the basin are important. Bangladesh, situated at the downstream of the basin, is at the front line of climate change, at-risk due to its flood-prone flat topography, overcrowded population and challenging socio-economic condition (Mirza et al. 2003;Rajib et al. 2011;Schiermeier 2011;Dastagir 2015). In this study, the country has been selected as a case study area of the downstream parts of the GBM basin to analyze the changes of the imminent climate and its associated extremes.
Located on the low-lying deltaic floodplains of the GBM basin, Bangladesh is already experiencing adverse impacts of global warming, disasters related to potential climatic changes and associated mean sea level rise (Mirza et al. 2003;Rajib et al. 2011;Schiermeier 2011;Dastagir 2015). As a consequence of observed increasing trends in the number of wet days, the region is likely to experience more seasonal flooding (Shahid 2010 Monsoon-dominated micro scale climate processes play a strong role in precipitation and related extremes over Bangladesh, understanding which are essential to evaluate and explain the condition of future extremes. Therefore, the assessment of the climatic processes at a regional scale is required to derive consistent and reliable projections of probable future climate. In this context, high-resolution (0.25°×0.25° or 0.5°×0.5° resolution) projections are imperative for climate evaluation at a regional level, where results from Global Climate Models (GCMs) are sparsely gridded (typically more than 1°×1° resolution) (Dankers et al. 2007;Bhaskaran et al. 2012). The GCMs downscaling are a widely used method for regional climate studies. Although statistical downscaling is a computationally inexpensive tool for simulating climate projections, dynamic downscaling is proven to be more representative of fine scale physical processes (Paul et al. 2008;Hong and Kanamitsu 2014;Kang et al. 2014;Lee et al. 2014;Lee and Hong 2014). Dynamically downscaled data generated by regional climate models (RCMs) such as PRECIS and RegCM have been used to project the future climates of Bangladesh (Rajib et al. 2011;Rajib and Rahman 2012;Hasan et al. 2013;Hussain et al. 2013;Murshed et al. 2013;Nowreen et al. 2014 (Raneesh and Thampi 2013;Shashikanth et al. 2014;Apurv et al. 2015). Climate projection using QM based bias correction in RCM results will thus be an advancement to evaluate the South Asian monsoon climate and its extremes in forthcoming years. Therefore, this study has conducted a quantile based bias correction approach to evaluate future climate and its extremities in available CMIP5 level RCM Projections over a monsoonal South Asian country; in this case, Bangladesh.
The aim of this study is thus to explore the climate and its extremes over a monsoon dominated country like Bangladesh by following the latest RCP emission scenarios, considering fine scale regional physics, incorporating the uncertainties range, and also by conducting advance bias correction methods to accomplish most reliable future projections. In this relation, the article has attempted to analyze the future probabilistic climate of the country, by using five regional climate model projections driven by four GCM results; i.e., EC-EARTH, CNRM-CM5, CCSM4, MPI-ESM-LR. All of the projections utilized three different RCP scenarios: historical, RCP 4.5 and RCP 8.5, to capture the whole range of future uncertainties. To remove the systematic biases from multi-model RCM outputs, latest QM bias correction methods are applied to a newly generated observed gridded data product. Developed by comparing six available observed datasets, the data product presents extreme events in a spatial gridded form.
The remainder of the study is organized as follows: in Section 2, a description of the observed and model data is presented. Method of bias-correction and description of selected extremes are also provided in the same section. The starting part of Section 3 explains the performance of past climate extremes considering different sets of observed data as well as the performance of the six RCM projections. The results are analyzed and a detailed discussion of the study and the potential implications are concluded in later part of Section 3.

Observed data
The assessment of climate and associated extremes by incorporating climate models, requires evenly spaced data network with long term, reliable time series. Land-based station data has more reliable extreme information, especially in the monsoon dominated regions due to erratic occurrences of rainfall (Singh 2015). However, such datasets have lack of spatial coverage. On the other hand, gridded data products tend to provide spatially rich climate information although it loses some accuracy in terms of magnitudes of daily extremes (Yatagai et al. 2007). Therefore, conjugating the land-based observed data with the best performed gridded product has the potential to provide most accurate information of climatic extremes (Khandu et al. 2015;Song et al. 2015 For both minimum and maximum temperature, the only exception from the grid generation process of precipitation was the absence of a distance threshold after ordinary kriging. In that case, only BMD land-based data was used for temperature gridding. The refined gridded products were used as a basis for further climate analysis and referred as the observed data throughout the rest of the article.

Model data
Climate data derived from the five available RCM outputs have been selected for this study.
The datasets were made available through COordinated Regional Climate Downscaling Experiment (CORDEX), a program that brought forth a collective effort to regional climate projections globally (Giorgi et al. 2009). The CORDEX aims to advance and coordinate the science and application of regional climate downscaling through global partnerships.
The project defined some specific domains around the globe and invited communities to conduct regional downscaling in those designated domain. Through the project's data portal, several RCM results became available over South Asia (CORDEX, 2015). As domain selection could be sensitive in a regional modeling study (Bhaskaran et al. 2012 However, as solution to those limitations, they proposed detail representation of bias and unbiased products as 'short' term and multi-model projections as 'mid-term' solutions. In this study, we have adopted both of these solutions to cover the limitations of the biascorrection method.

Bias correction
In this study, we have conducted bias correction on high-resolution RCM results to evaluate better climate projection. Previously published literature has shown that GCMs have limitations in the representation of the mean monsoon climate; thus their results need further refinement for regional studies (Dankers et al. 2007;Bhaskaran et al. 2012). In case of the bias-correction of GCMs, the coarsely simulated model results may impose a common projected trend over the finer-scale regional projections. If there are several observation stations located within the same GCM grid cell, after bias-correction, the projected climate of those stations will follow the same projected trend of that particular grid cell. However, in the regional scale, the observed trends may vary within the stations due to local geophysical characteristics such as orographic effects and land use/cover patterns (Wood et al. 2004;Jang and Kavvas 2014;Bieniek et al. 2016). On the other hand, dynamical resolved RCM projections retain these characteristics and provide more regional information of the projected climate within the same GCM cell. As a result, the bias- were constructed for each individual day of a year with a 30 days' window. Observed cumulative empirical distribution was also generated using no-parametric kernel approach.
Comparing two ECDF, a relation of respective biases was established for each grid cell. Therefore, we excluded the results of CISRO-CCSM4 model in our subsequent analysis.

Performance of observed gridded product
Prior to the generation of observed grid, we compared suitable existing gridded data, ERA-Interim and APHRODITE, with gaged values. The assessment was presented in Table 2.
Annual values of the precipitation from both gridded products deviate from observed values, where for ERA-Interim, they exceed 400mm in both decades. At monthly scales however, during monsoon season, significant disagreement has been observed in ERA-Interim data. This finding agrees with those of Rahman et al., (2012b), which showed that ERA-40, a previous version of the ERA-Interim product poorly captured the monsoon rainfall, especially in the Sylhet region (North-eastern part of Bangladesh). Such dry biases of ERA-Interim were also observed over north-eastern Bangladesh by Ménégoz et al., (2013). In case of APHRODITE data, its shows considerable difference in compare to BMD data in some of the daily climatic extremes like 50mm rainy days, 90th percentile of rainfall and wet days (Table 2). However, from the overall gridded product assessment, APHRODITE data has proven to be better than the ERA-Interim data product. Therefore, for further improvement of daily extremes, APHRODITE data were combined with BMD gauged data to formulate a new gridded data product.
The performance of precipitation and its extremes for the newly generated dataset also are also showed in Table 2. The gridded dataset not only increase the accuracy of the mean at yearly and monthly temporal scale but also provides improved values of the daily rainfall extremes over Bangladesh. Notable improvement of monthly rainfall is also observed in the new dataset during the monsoon season, which is ranging from month of July to September.

Performance of the models before bias-correction
Model requires comparative analogy in respect to the gridded product. The performance assessment of each model with the observed data is shown in Figs. 1a, b, c, at a monthly temporal scale.
The annual rainfall cycle in South Asia is dominated by the monsoon season during the months of June, July, August and September, when almost 70% of the rainfall occurs.
Among the five regional models, only the RCA4-EC-EARTH model overestimates the monsoon rainfall over the region. The simulations obtained by the CISRO RCMs show very similar temporal biases of precipitation, which makes it evident that the boundary data from GCMs plays a dominant role in those RCMs for simulating regional precipitation over the region. This argument is also supported by Ghimire et al., (2015), where similar annual patterns are also found in the far north of the country over the Himalayan area. The study revealed overestimation of rainfall over the Himalayan region, whereas, in our case monsoon rainfall is underestimated. This discrepancy might be due to the inaccuracy of the orographic rainfall at the high altitudes produce by the regional model. For REMO-MPI, there is a large dry bias during monsoon season. Similar findings have also been reported by Jacob et al., (2012). . In general, the temperatures derive from the models show better performance than precipitation in terms of monthly climatic biases. Minimum temperature produced by RCA4-EC-EARTH model, has cold biases during winter and post monsoon seasons. During the months of March and April, all of the RCMs show some biases in maximum temperature, but the rest of the annual cycle are consistent with observation.
The choice of convective schemes across the regional climate models could play a crucial role in reproducing monsoon rainfall, which eventually affect the dynamic downscaling process (Prein et al. 2015). In this study, we have utilized the results of three regional models ( Table 2) that have their own convective schemes. The cloud formation scheme of REMO was adopted from the MPI global model and it is based on the approach from Sundqvist (1978). In case of CCAM from CSIRO, the atmospheric climate model utilizes a conformal cubic grid. It includes CSIRO's mass cumulus convection scheme that incorporates downdrafts and the evaporation of rainfall (McGregor 2005). For the RCA of SMHI, the regional model resolves convective processes with an entraining and detraining plume model using the Kain-Fritsch (KF) scheme (Kain and Fritsch 1993;Kain 2004 After the assessment of mean climate, we also explored the efficiency of the models in generating selected extremes. Spatial and temporal evaluations of the precipitation indices have been shown in Fig. 2, 3, 4. The spatial analysis reveals dry biases in Rx50 for the REMO-MPI model results, especially in the northwestern region. All three CSIRO models have been able to capture Rx50 quite well and show little spatial variability among them. RCA4-EC-EARTH simulation data shows number of Rx50 lesser in the Sylhet region (north-eastern part), whereas in the east-central part Rx50 is much higher than observed climate. In context of lower end of rainfall distribution, Rx1 are well captured by all three CSIRO simulations (Fig. 3). However, the CSIRO models estimate higher number Rx1 over the northwestern parts, while REMO-MPI and RCA4-EC-EARTH show lower Rx1 over the north and the northeastern areas compare to observed precipitation. In terms of heavy rainfall (Rx10), except RCA4-EC-EARTH simulation, all other RCM simulations suggest lower values compare to observation, but they preserved the existing spatial pattern during the baseline period (not shown in the figures). Spatial average of Rx10 and Rx50 from the models present some dry temporal biases than respective observed extremes ( Fig.   4). As these model biases rapidly change between individual years for a particular model, average of all model results can give us a better picture of model performance over the country. After observing a constant dry biases in both indicators, correction of these biases become essential for further extreme analysis.

Performance of the models after bias-correction
Bias corrected data results along with raw model data are shown in the Taylor diagram values are found over the northeastern region (Fig. 3). The temporal analysis reveals that bias-correction of precipitation actually made the model results much wetter than before.
Similar comments about these wet biases from the quartile mapping were also found in previous studies (Wilcke et al. 2013 However, as Bangladesh is a monsoon dominated region, the pattern of normal to extreme precipitation have significant impacts on its agro-based economy and socio-economic outcomes (Ahmed 2003;Islam et al. 2005;Ahmed 2006). Deriving accurate information of precipitation variability from the climate models are essential for studying the climatic impacts for future decision making of the country (Shahid 2011). In this context, the daily bias corrections of RCM rainfall allowed significant improvement by providing a more realistic capture of high intensive rainfall events. Moreover, the corrected model results also exhibit improved accuracy of the mean climate over the region. For temperature, the daily bias correction method reduces the disagreement of models results drastically, where corrected model datasets are highly correlated (temporal correlations > 0.9) with observed data (Fig. 1). Therefore, spatial analyses of these datasets have not been shown in this study.

Projected temporal state of PR, TMAX and TMIN
The mean annual PR over Bangladesh under RCP 4.5 and RCP 8.5 are shown in Fig. 5a indication of an appalling future, where the number of winter days will be much lower than the current state. It is also noteworthy that TMIN shows a much faster increase than TMAX (about 1°C higher by the end of the 21stcentury) under both scenarios. Thus, it signals a reduced variation between day temperature-TMAX and night temperature-TMIN over the region. of rainfall under both RCP 4.5 and RCP 8.5 scenarios. In this projection, a faster decrease of rainfall is observed in the northern parts than southern parts for both pathways. It is notable that, under RCP 8.5, estimated decrease of 900mm annual rainfall is observed in Sylhet region by the end of the century.

Projected spatial changes of PR and average temperature
As TMAX and TMIN changes are very similar to each other, spatial changes of average temperature by averaging TMAX and TMIN are presented in the study (Fig. 7). The Fig. shows the spatial distribution of mean annual changes in average temperature in the 2050s (2041-2070) and 2080s (2071-2100) under RCP 4.5 and RCP 8.5 scenarios, relative to baseline. The southwestern coastal zone of the country projected to be the most affected with definite increase in temperature. Interestingly, all projections strongly agreed with such observations, strengthening the CMIP3 climate model results and the argument that the southwestern part of Bangladesh is the most vulnerable due to its socio-economic condition and population density (Ali and Islam 2014;Dastagir 2015). Vulnerability due to such increase of overall temperature ranges will result in disastrous outcomes for this region by the end of the 21st century. During 2050s, projected temperature estimated to increase by 0.75°C-1.75°C under RCP 4.5 and by 1°C-2.5°C under RCP 8.5 in the southern parts of the country. Although RCP 4.5 scenarios are known to be less ambitious and CO2 controlled emission scenarios, most of the climate models still suggest at least a 2°C increase of temperature during 2080s over most part of the country. On the other hand, the RCP 8.5 scenarios provided an alarming projection of up to 4.5°C within 2100 over the whole country. It should be noted that CCSM4 model result shows much lesser increments of temperature among CSIRO RCMs.
In general, north-western part of Bangladesh experiences lower rainfall compared to the hilly regions located in the eastern parts of the country. Analyzing model projections and their bias correction, it is observed that RCA4 and REMO show an increase of temperature over low rainfall zones. Contrary to that, CSIRO models conclude higher increase of temperature in the wet hilly regions. The contrasting characteristics of the projected results reaffirm that the relationship of rainfall depends on complex interaction of ocean, land and atmosphere, and not just temperature characteristics. Thus, it again emphasized the importance of dynamic over statistical downscaling considering regional scales (5km to 50km) for a clear understanding of the future climate.

Projected changes of climatic extremes
In this study, we have tried to portrait the effect of bias-correction on future extremes.
Therefore, we demonstrated the difference between bias-corrected and uncorrected climate in a spatial analysis of the extreme indices. The spatial changes of two precipitation (Rx10 and Rx50) and two temperature (TXx, TNn) indices, both bias-corrected and uncorrected versions, are presented in Fig. 8, 9. The uncorrected changes of Rx10 and Rx50 differs from bias-corrected values of same variables, both in terms of pattern and magnitude. The pattern of uncorrected Rx10 deviates drastically from 2050s to 2080s in both scenarios, where changes in the pattern between the time-slices were gradual in the bias-corrected results. In terms of magnitude, uncorrected Rx10 shows smaller changes under RCP 4.5, but more increase on the southern parts of country during 2080s of RCP 8.5 in compare to bias-correction.
The changing variability is also much higher in the uncorrected than the bias-corrected In cases of the temperature extremes, the uncorrected and bias-corrected pattern is more agreeable than the precipitation extremes, as expected, due to fact that the models in hindcast simulation better captured temperatures. However, the variability between uncorrected and bias-corrected changes are still persistent, where uncorrected extremes shows hotter climate with a higher uncertainty range. Both figures for extremes suggest that the uncorrected changes give much wider uncertainties and exhibit erratic pattern of changes in future projections.
In addition, for uncorrected projections, the policy makers need to keep in mind that the changes that are represented in the Figures are not changes from past climate, rather the changes are from the individual model results of the baseline run. This can cause a great ambiguity in conferring the results in the prospective decision-making process. In this context, the uncorrected climate projections can give idea of the wider range of uncertainties, where bias-corrected model results can provide a higher level of confidence.
Utilizing the projected changes in extremes by the bias-correction techniques, the decision maker might have a smaller uncertainty range but they will have a realization of changes from the actual observed climate. As the focus of this study is to covey assertive information of extremes for decision makers, the bias corrected results should be more relevant in this context. Therefore, we have described the bias-corrected results as the projected result in the following sections.
Projected mean of precipitation extremes (Rx10 and Rx50) suggest an increase of rainfall all over the country, where northeastern part indicate higher increment rate than the rest of the country under both RCP scenarios. The region is mostly hilly and important for industrial tea plantation (Islam and Miah 2003). As almost all models conclude the increase of extreme rainfall at end of the 21st century, such changes may have significant impact on the tea plantations of the area. These areas are also prone to high amount of intense rainfall events and flash flooding. Thus, increase of heavy rainfall events will eventually extend the risk of flash flooding to an alarming level. In addition, heavy rainfall events (Rx10) will increase in much faster rate than extremely heavy rainfall events (Rx50) all over the country. The projected Rx10 of model mean refers an increase of at least 80 or more of the heavy rainfall days within a decade (8 days per year) from current climate. Majority of model result also suggest that the country will experience at least four more days of extremely heavy rainfall annually from present time to the end of the century. By looking at variance of Rx50 between model projections, it can be said that all five models show more confidence in projecting extremes over the western parts than other parts of the country. In conjunction with precipitation and temperature increases, it is found that the eastern part of the country will experience more changes in terms of both mean climate and extremes than western parts of the country. Heavy and extremely heavy rainfalls in majority models show a significant increase over the hilly regions (northeastern and southeastern part) of the county in both time slices. Interestingly, temperature extremities will also increase over the same regions.
Probability distribution (PD) of the two precipitation extremes (Rx10, Rx50) and two temperature extremes (TXx & TNn) are presented in Fig. 10a shows a reduction of its extremity in future years. However, such results can be uncertain due to the inability of GCMs to produce such extremes (Barros et al. 2014). It is evident that, increase of TNn during 2080s is much higher than increase of TXx during the same period. Such incremental shift in PD can lead to a less variable diurnal temperature range over Bangladesh.
In this study, we have utilized five combinations of RCMs and GCMs that gave the projections of future probable climate and its extremes over Bangladesh. A larger set of RCMs and GCMs can cover a wider uncertainty range, and provides us more information of the projected future. However, we have a limited choice of GCMs and RCMs available for this study due to unavailability of projections and computational constrains over the CORDEX domain (explained in the Methods section). This limitation could potentially improve in the future as additional CMIP5 RCM projections become available.
The replication of non-stationarity by climate models is important for climate projections.
To explore the confidence of the projections, Salvi et al. (2016) demonstrated a strategy to check the assumption of stationarity for statistical downscaling methods and also recommended to test the framework on dynamical downscaling as well as on GCM results.
This framework could potentially be very useful for the bias-corrected and uncorrected RCM results over Bangladesh. However, it was not covered in this study, and thus provides an opportunity for future improvement.

Conclusion
To examine future changes in climate and climatic extremes in a monsoon-dominated region, we have analyzed the results from the five available regional climate models from IPCC AR5 over Bangladesh for the present  and future . For a climatic extremes analysis, we have generated a new gridded rainfall data product over the country to address information gap on extremes, insufficient observed grid-data at a daily scale, and uncorrected high bias values of future projections over the region. Moreover, the most recent CMIP5 Regional Climate Models with higher accuracy under emission trajectories have been evaluated. The summarized findings can be presented as follows: After bias correction with newly generated observed data product, the patterns of extreme climate events are preserved between the model and observations over Bangladesh. The comparison of Taylor diagram also validated the performance of bias correction between observed and model data. Under the RCP 4.5 and RCP 8.5 scenarios, rainfall increase has been observed significantly and with high confidence over the eastern hilly regions, especially in the northern parts. A possible reduction of rainfall will be more prominent in the northern zones than southern zones of the country. Maximum and minimum temperature changes are more incremental in the southwestern parts compare to other parts.
The model result under RCP 4.5 scenario shows large uncertainly in rainfall and much steady rise of temperature in the middle of the 21st century. Under RCP 8.5, the projected increase in rainfall events is observed over the areas where temperature will increase faster and vice-versa. Due to the energy difference between RCP 4.5 and RCP 8.5, the rainfall projections of RCP 8.5 shows more uncertainty than RCP 4.5 at 2100. The projected model results have also indicated the changes in the frequency of extreme precipitation and temperature events. The extremities of rainfall tend to be more variable than temperature extremes and the number of heavy rainfall days will be much higher in future years. The results imply that there would be much higher heavy rainfall events over the northeastern hilly regions than other parts of the country. Alarmingly, the temperature extremity also tends to have a drastic increase over the same regions.
Although bias correction of RCM provides a useful basis for the impact studies, considerable uncertainties remain in GCM, RCM and the bias correction method itself.
Despite these uncertainties, bias-corrected projections at the appropriate spatial-temporal scales are the most reliable tools for understanding hydro-climatic impacts. In this study, an initial investigation of the hydroclimatic extremes has been performed with an appropriate daily scale bias correction method with a new gridded climate dataset over a monsoon region, the Bengal Delta region of South Asia. Further analyses of monthly or seasonal extremities in precipitation and temperature should be pursued in future years.              and bias uncorrected (bottom) regional climate projections have been compared.    (4) 26 (4) 20 (1) 24 (8) 22 (1) 30 (1)   Rotavirus is the most common cause of diarrheal disease among children under five.
Especially in South Asia, rotavirus remains the leading cause of mortality in children due to diarrhea. As climatic extremes and safe water availability significantly influence diarrheal disease impacts in human populations, hydroclimatic information can be a potential tool for disease preparedness. In this study, we conducted a multivariate temporal and spatial assessment of thirty-four (34)  cycle. The proposed model shows lag components, which allowed us to forecast the disease outbreaks one to two-months in advance. The satellite data-driven forecasts also effectively captured the increased vulnerability of dry-cold regions of the country, compared to the wet-warm regions.

Introduction
Living in the age of satellites and nanotechnology, a significant fraction of the global human population is still threatened by diarrheal diseases. Spectroradiometer (MODIS) land surface data product can provide daily temperature data at 1-km spatial resolution (Pagano & Durham, 1993). These datasets, not only improve data acquisition intervals compared to station data, but also provide more spatial information in a near-real-time basis.
With establishment of the links between diarrheal diseases and new generation earth data, including satellite observations, there is a great potential to develop models for disease prediction at higher spatial and temporal resolutions. Such systems are especially crucial in developing countries, where the population faces a massive burden of rotavirus related mortality and morbidity each year. Bangladesh, a South Asian country with an emerging economy, still suffers a heavy toll every year due to rotavirus. In this study, we have explored the effect of climatic extremes on the rotavirus infection cycle in Bangladesh both spatially and temporally. We have evaluated rotavirus patterns over several cities inside the country and across South Asia to understand the larger context in relation to regional hydroclimatic processes. We also implemented a deterministic multivariate modeling for risk assessment and integrating near real-time satellite products (with GPM for rainfall and MODIS for temperature).

Study Area:
A robust epidemiologic assessment of rotavirus diarrheal outbreak with climate requires a sufficiently long time series and good spatial coverage of disease data.

Weather Data:
We obtained daily maximum (TMax) and minimum temperatures ( For detecting spatial variability, we utilized two types of satellites data products in this study. The Global Precipitation Measurement (GPM) data were used as the source of the satellite precipitation, collected from March 2015 to December 2015. The GPM mission is an international network of satellites that provides the next-generation global observations of rain and snow (Hou et al., 2014). We also utilized an additional satellite-derived rainfall dataset from the Tropical Rainfall Measuring Mission (TRMM) for validation purposes.
Among the various products that are available, we used the TRMM3b42v7 version with a spatial resolution of 0.25 degree x 0.25 degree and a temporal resolution of 3-hour. A global Land Surface Temperature (LST) data product was acquired from the Moderate Resolution Imaging Spectroradiometer (MODIS)-Aqua satellite (MYD11A1.005 version) for both day and night temperatures at a 1-km spatial resolution.

Method
Our study approach can be separated into three sections: temporal assessment, spatial analysis, and multi-variate modeling and validation with satellite data.
A robust analysis of the hydro-climatic influence on the transmission cycle of a disease requires specific climate realizations. For example, the mean or maximum state of a monthly temperature may not directly influence a disease outbreak, but a specific temperature range or consecutive rainfall events can trigger an epidemic. Therefore, for a comprehensive examination of environmental drivers on rotavirus diarrhea, we selected 36 climate indices based on various properties of weather events (Table 1). We either applied or adopted the climate indices from the Expert Team on Climate Change Detection and Indices (ETCCDI) (WMO, 2007). These indices were used in various climate studies to analyze the extremity of the climatic phenomenon (Alexander, 2015;Hasan, Islam, and Akanda 2017;Keggenhoff et al., 2015). The selections of the indices in those studies were conducted based on particular objectives of individual studies. In this case, we selected the indices that are most relevant to rotavirus transmission dynamics.
Therefore, we selected 3°C as threshold interval to classify 9°C temperature range for developing TxijGE and TnijGE indices. As the minimum monthly DTR of Bangladesh is 6°C, we selected half of that (which is 3°C) to capture the temperature effect in both day and night (Islam & Hasan, 2012). Any threshold interval lower than 3°C will result in redundant indices. On the other hand, any threshold interval higher than 3°C will plausibly miss the variation of temperature that can influence rotavirus. The duration of hot or cold days based on a particular threshold were described by the rest of temperature indices (i.e. Tn10, Tx90, etc.).
In case of rainfall, intensity and amount were characterized with SDII and PRECIPTOT.
The magnitude of rainfall was described with Rx1and Rx5 indices. The durations of various kinds of storms were classified using the rest of the precipitation indices. However, among all the indices, many are season specific and have interdependency among them.
On this ground, we categorized the indices into two seasons; October to April as the dry winter season and July to September as the wet monsoon season. The indices that have 60% or more zero values were dropped and eventually we concluded with 22 and 28 indices among 36 indices for winter and monsoon seasons, respectively. For example, we did not select Tn1618GE for the monsoon season. As days with minimum temperature range of 16 to 18 degree will be zero for monsoon months, any correlation value between rotavirus and Tn1618 will result in misleading information. Therefore, some indices were dropped from the pool of 36 indices, when we conducted the season specific analysis. All the indices for temporal and spatial analysis are generated from BMD observed data, where the validation analysis of the indices is generated with daily satellite data.
Evaluating spatial risk of a disease can be modeled with existing stochastic methods like the Bayesian approach (Cheng & Berry, 2013), Monte Carlo simulations (Prosser et al., 2016) or Susceptible-Infectious-Recovered (SIR) (Grassly & Fraser, 2008) models. While the stochastic methods are useful to capture probable spatial patterns of diseases transmission, the complexity of the methods sometimes miss the deterministic influence of a particular driver. As the goal of our paper is to evaluate the influences of climate indices on rotavirus diarrhea, we utilized a deterministic model to formulate the risk of the disease and avoided the population effect. In the process to eliminate the influence of population, we standardized and scaled the disease cases for each of the selected cities and combined the disease cases into a single series of the same time frame (January 2013 to June 2015) to conduct spatial analysis. The standardization method was adopted from Jagai et al.
(2012), where we considered our scaled values as z-scores of rotavirus risk. As a result of removing the effect of population, the analysis thus represents the severity of disease cycle rather than actual cases of diseases. Any value that exceeds one (1)

Seasonal characteristics of Rotavirus in South Asia:
In this section, we discuss the general spatio-temporal pattern of rotavirus outbreaks seen in South Asian cities. Annual rotavirus cycles over South Asia are presented in Figure 2 In this analysis, we calculated temporal correlation only over Dhaka and not the other cities due to lack of data availability (the disease data of other cities starts from 2012). Among the precipitation indicators over the city, RR1 was found to be one of the influential indicators on rotavirus. The correlation analysis suggests (Figure 3(a)) that a decrease in RR1 in September affects the winter rotavirus cycle especially for the month of November.
The secondary outbreak during the July, August and September is affected by the number of days with rainfall events of 70mm or more (RR70) (Figure 3(b)). However, both the rotavirus cases and RR70 were higher during the 2007 floods over the city.

Univariate correlation between climate indices and rotavirus
To assess the effect of individual climate variables and indices on rotavirus transmission, we conducted univariate analysis considering moving average and lag of related variables.
The correlations for the winter and monsoon seasons are presented in Figure 4.
During the winter season, rotavirus outbreak in Dhaka shows a strong negative lag relation (1-month) with the selected rainfall-related indices (Figure 4(a)). In case of other cities ( Figure 4(b)), the same indices show significant but lower correlation values. Unlike Dhaka, the correlations of indices in other cities do not exhibit any substantial lag effect.
Thus, we can say that the low duration of rainfall events seems to be an influential driver for the season, where the effects come with a delay (1-month) over Dhaka compared to other places. The temperature indices related to the colder spells strongly impact the winter epidemics in both spatial and temporal analysis. However, the spatial correlations are weaker than the temporal values in both types of indices, probable due to the varying rainfall patterns between the six locations. The temperature indices that display the strongest correlation (0.5 or more) are Tmax, Tmin, Tn1621GE (number of nights with temperature between 16°C to 21°C) and Tn1921GE (number of nights with temperature between 19°C to 21°C). All these indicators confirm the effect of colder temperatures on the rotavirus cycle similar to Atchison et al. (2010).
During the monsoon season, the temporal investigation of rotavirus indicates significant correlation with all rainfall indices where such relationships are absent in the spatial assessment (Figure 4(c-d)). The outcome is expected, as the secondary monsoon outbreak and its impacts are most profound in Dhaka among the six selected cities of Bangladesh ( Figure 4(a)). Tn2225GE significantly correlates with 2-month lag rotavirus outbreak, which is the strongest relationship among the indices. The relationship suggests that a night temperature range of 22°C to 25°C has a potent role in the monsoon cycle of rotavirus over Dhaka.
From Figure 2 and 4, it is evident that the winter cycle of rotavirus is more prominent than the monsoon cycle over the study region and is strongly influenced by climatic factors.
Thus, we focused the investigation on the winter epidemic for the rest of the study. For a detailed understanding of the winter cycle, we characterized it into three phases; rising, peak and falling phases. The rotavirus outbreak starts to appear during the months of October and November, thus can be classified as the 'rising' phase. As the cycle, typically reaches its 'peak' during the months of December, January and February, we considered it as the 'peak' phase. From February to April, the cycle enters in its recession phase, therefore, this phase was defined as the 'falling' phase. Based on the three phases, we conducted two levels of correlation analysis as described previously between rotavirus cases and climate indices. As temperature and precipitation indices have dependency among them, many indices show similar correlation in particular phases. Therefore, to make a concise judgment, we presented only the most significant correlation for each phase of the epidemic cycle in Table 2.
The rising phase of rotavirus cycle has significant influence by the night temperature as The synthesis of the analyses revealed that Tn1621GE and RR1 are commonly correlated during the rising and falling phases, both temporally and spatially ( Table 2). The longest time series for Dhaka cases also disclose the significant relationship of Tn1621GE at the winter peak. On that account, we can say that a specified night temperature range with dry weather is a prominent force to the spread of the disease during the winter.
The assessment between three selected phases of the rotavirus winter cycle confers the effect of climate more strongly in the rising and falling phases rather than peak phase.
Therefore, to achieve more clarity, we conducted a moving average analysis of one, two The month-wise correlation analysis for the individual cities would be statistically insignificant, as a common data period between the six cities are only available for approximately 3 years (for a seasonal analysis, it will generate 3 points in three years). In this case, we considered two of the most influential variables of the winter cycle; Tn1621GE and RR1, and compared them with rotavirus proportions of these cities in Rajshahi, the same cycle shows a two-month lag relation instead of one. Moreover, the rotavirus peak also follows distinct patterns with RR1 or rainy days. In case of Barisal and Sylhet, the peak of rotavirus occurs during the driest month (or lowest RR1) without showing any lag. Over Rajshahi, this relationship extends for a two-months lag. This variation in lag for both indices explains why there is no significant relationship found during the peak phase (Table 2) in the spatial analysis.

Multivariate assessment
From the univariate analysis, we identified the RR1 and Tn1621GE as the most influencing variables on the winter rotavirus cycle. Using these climatic indices, we developed a multivariate regression model for evaluating the winter cycle. As the indices poses different correlation values in explaining the transmission process in different phases, we conducted three separate multivariate models for the three phases of the cycle and combined them into a single model. As we explored the spearman rank correlation values, we also incorporated non-linear relationship between the indices and rotavirus cases. For checking the distribution of the response (response here is z-score of rotavirus) variable of the model, we conducted Shapiro-Wilk (Shapiro & Wilk, 1972) and Kolmogrorov-Sminov (Massey, 1951) Figure 6 shows the spatial prevalence of observed and model estimated rotavirus over Bangladesh. For October, the eastern parts of the country largely agree with the observed disease incidences, where magnitude slightly deviates. In case of November, the observed patterns are well captured by the model; however, magnitude deviates over the Barisal and Rajshahi regions. We also presented the potential of using TRMM satellite with MODIS datasets (Figure 7) to predict disease risk over the focus region. Figure 7 shows the October and November outbreaks from model and observed data during 2014. The TRMM derived disease map is able to capture the pattern better than GPM derived product. However, it should be noted that 2014 winter data are also utilized in model formulation, thus it cannot be considered as a validation result.

Discussion
Our initial assessment infers that the rotavirus cycle is strongly influenced by the dry and cold winter season climate in Dhaka. In Great Britain, Atchison et al. (2010) explored the temperature dependence of rotavirus and conferred that above the 5°C threshold, an increase of the average temperature decreased the infection rate of the disease. A similar understanding was also found in Australia (D'Souza et al., 2008), where rotavirus diarrhea admissions are associated with lower temperatures and lower humidity. Although these two studies were conducted in different climatic zones altogether, we believe that the dearth of overall number of studies linking rotavirus with climatic indices, these findings are still important evidences towards the influence of temperature on rotavirus incidence.
In South Asia, Jagai et al. (2012) also showed that the reduction in annual temperature and precipitation increases the level of infections of rotavirus, supporting our findings.
As our assessment separated the timeframe into two seasonal cycles, the correlation from winter cycle over all six selected cities strengthens the findings of previous studies.
However, we also found significant positive association of rotavirus infections during monsoon over Dhaka. Dhaka is a densely populated city with a high number of informal settlements, or slums, with poor water and sanitation conditions (Akanda & Hossain, 2012). As rotavirus pathogens can be transmitted through the fecal oral route, high precipitation events can create waterlogging and eventually connects to the pathogen transmission pathways (Dennehy, 2000). Thus, Dhaka experienced an additional monsoon outbreak compared to other cities and the outbreak may be influenced by heavy rainfall events. Such phenomena also clarify why the monsoon indicators showed insignificant relationships with rotavirus in other cities. Dhaka typically observes the annual highest rotavirus incidence during January, but some exceptions were observed during March 2009 and July 2004 (Figure 2(b)). The 2004 flood event was one of the most devastating floods in the last decade in Bangladesh (Schwartz et al., 2006). Dhaka including cholera, rotavirus, and dysentery (Harris et al., 2008, Cash et al., 2014).
Our study also provides some detailed assessment of the winter rotavirus cycle. We found that the rising phase of rotavirus is negatively correlated with SU or Tx2532GE, which represents the amount of warm days in month. As the virus prefers low temperature environments, the lower number of warm days eventually helps to initiate the spread of the disease. Previous studies indicated that the virus can be active in the environment for up to 4 weeks or one month without a host body (Levy et al., 2009). Therefore, reduction of warm days may increase the rotavirus sensitivity and the effect can be delayed up to one month. Our findings also suggest that the beginning of the winter cycle (October-November) is highly correlated with RR1 and Tn1621GE, both spatially and temporally.
Average night temperature during September-October are 25°C. As Tn1621GE represents the night temperature of 16°C to 21°C, some nights in September start to experience temperatures below 21°C. Therefore, the index can be reflected as colder nights of that month. In a laboratory test, rotavirus was found to be active for several days in 4°C and 20°C temperatures without human contact (Moe & Shirley, 1987). In aerosol, the virus is also infectious in low temperatures (Moe & Harper, 1983). Therefore, higher values of Tn1621GE, which act as cold nights during September-October, may promote the infectivity of rotavirus up to a 4-week delay.
On the other hand, the RR1 index represents the number of wet days in a month rather than magnitude or intensity of rainfall events. As rotavirus transmission can be driven with air, reduction of rainfall may raise the propensity of aerial transport (Ansari et al., 1991) of contaminated fecal matter. Therefore, RR1 can be considered a barrier to air-borne transport of rotavirus. Consequentially, the joint effect of RR1 and Tn1621GE triggers the one month delayed outbreak during the rising phase of the winter cycle. During the peak month of rotavirus in December, RR1 becomes nearly zero over Dhaka, thus allowing aerial transport of the virus to its highest potential. In this phase, the correlation with Tn1621GE shifts from positive to negative. During the month of December, the average nighttime temperature also drops below 21°C. Such a drop of night temperature, transforms the Tn1621GE index to a representative of a warm night, as temperatures can be higher than 21°C during this month. As Atchison et al. (2010)  In other cities of Bangladesh, the timing of the cycles did not match in the same way, thus correlation values decreased. In spatial cases, the rising and falling phase still showed a significant correlation with RR1 and Tn1621GE, but values of the correlation coefficient are lower than the values of the temporal analysis. During September, Tn1621GE acts as an indicator of cold night. In Sylhet and Barisal, as the increase of cold and dry nights coincide, rotavirus infection experiences a sharp rise, thus no lag relationship is observed.
However, in places like Dhaka and Mymensingh, where dryness comes early but temperature suitability comes in a delayed manner, the places experience a one-month delay in an outbreak. If these two phenomena have a much wider gap, it can result in up to a two-month delay, which was observed in Rajshahi. Therefore, our findings suggest that the timing of coldness and dryness can locally affect the spread of a rotavirus epidemic.
This finding increases the potential of using a high-resolution satellite data product in forecasting the local onset of the outbreaks. It is difficult to draw a generality from only three or four years of rotavirus observations; upon availability of more surveillance data, such analysis can be explored in more detail in future.
From the multivariate analysis, we are also able to confirm our hypothesis through the model selection process. All components of Equation 1 significantly influence corresponding prevalence values of the rotavirus cycle and confirm the role of environmental factors on the whole rotavirus transmission cycle. The forecasted prevalence matched some spatial areas of observed values during November but not in October. As we conducted a detailed analysis of the climate extremes that are able to explain about 44% variance, such discrepancy was expected in spatial mapping. Due to the lack of sufficient spatial disease and climate data, the spatial signature was not captured properly, thus the accuracy of the model suffers. Moreover, factors like population dynamics and social behavior, or environmental factors such as flood and soil moisture can be important in improving modeling accuracy. In addition to that, the accuracy of satellite datasets can also be a plausible reason for the less than satisfactory performance of the spatial mapping.
However, the satellite products such as GPM, TRMM and MODIS not only give near realtime information, but also great spatial coverage, and have great potential to improve the resolution of the risk maps for such infectious diseases.

Conclusions
In this study, we have analyzed the relationship of various climate variables and indices with rotavirus outbreaks in South Asia, formulated outbreak models and proposed a forecast mechanism. In the validation process, we have utilized satellite-derived climate products, which have the capacity to provide climatic information within a 24-hour latency period after the acquisition of data. To quantify the disease outbreaks, we used a spatial   The cities with green dots were selected for the spatial analysis.           The future risk of diarrheal disease over Bengal delta based on climatic driven epidemic models: a case study with bias-corrected regional climate model results. continuing, however, due to limited resources and lack of institutional capacities, the 123 complete removal of these diseases are not likely to happen in the near future Chowdary et al. 2009). In this context, understanding the current and future influence of environment to the disease epidemic is essential for operational initiatives and policymakers.

Department of Civil and Environmental
The seasonal occurrence of diarrheal diseases like rotavirus and cholera confirms the influence of climate and environment in their transmission cycles over the South Asia (Leckebusch and Abdussalam 2015;Prasetyo et al. 2015;Ali et al. 2016). The rotavirus diarrhea is generally prominent during the winter and cholera during the post-monsoon season (Hasan et al., 2018;Akanda et al., 2009). Though the spread of the diseases can be associated with the lack of safe drinking water, inadequate sanitation, and poor hygiene, the hydroclimatic and related environmental factors are found to be strongly co-related in the events of the diarrheal outbreak. However, such outbreaks not only depend on the mean climatic state of a season but it can also be triggered by extreme climatic events like heavy rainfall, drought, and floods (Gurarie and Seto 2009; Remais, Liang, and Spear 2008;Bandyopadhyay, Kanji, and Wang 2012;Jutla et al. 2015;. Recent studies attempted to evaluate the relationship between the mean state of climate with the disease epidemics, but very few studies considered climatic extremes to investigate the diseases (Hasan et. al, 2018). Thus, a robust assessment of the relationship between the diarrheal disease and climatic extremes warrants much attention not only to reduce the present-day's outbreak but also to initiate future prevention strategies to face the effects of ongoing climate change. 124 The climate models are widely accepted tools for projecting probable future climate around the globe (Wright et al. 2015). Global Circulation Models (GCM) are used to project largescale global phenomenon under different climate scenarios, such as RCP scenarios (Wang et al. 2014). The GCMs usually produce coarse spatial resolution products, thus regional climate models (RCM) were introduced to generate high-resolution climate projections through downscaling techniques (Bhaskaran et al. 1996). Though the RCMs were proven to be more accurate and considered a wider range of physical parameters than GCMs, they still comprise some biases, especially in the small study areas. For the purpose of impact assessment, the bias-correction methods were later, introduced to further improve the RCM results in the regional studies (Bennett et al. 2014;Murakami et al. 2014;Macadam et al. 2016). However, adjustment of projected mean climate can be done using various methods, but the meaningful projections of climatic extremes were still challenging and handful methods to use it in the impact studies were introduced recently (Tian et al. 2007;Srivastava et al. 2015;Macadam et al. 2016). In this context, the impact study like the risk assessment of infectious diarrheal disease by climatic extremes not only needs some reliable inter-annual epidemic models, but also requires some meaningful climate extremes to drive the models. In existing literature, the impact studies that incorporated the disease risk with climatic extremes with climate extremes are rare (Teutschbein and Seibert 2012).
To project climate extremes for the impact study, there are sets of climate scenarios proposed by IPCC assessment reports (Stocker et al. 2010;Hartmann et al. 2013). The latest scenarios, that were published in IPCC 5th assessment reports, known as the RCP scenarios are considered to be the most up-to-date scenarios for climate change studies 125 (Kim et al. 2013 The manuscript was arranged in the following order: The description of obtained data and method were described in the methodology section.
The details of the climatic extremes and disease risk were explained in the method section. 126 The results section described the model validation of two diseases in their rising or outbreak triggering phases. The integration of climate data and disease risk were explained in the same section. The future directions were explained and discussed in the conclusion section.
Date and Methodology:

Study Area:
The Bengal Delta region of South Asia is considered to be the ancient place of origin, or native homeland, of cholera, the deadliest among the diarrheal diseases (Hu et al. 2016).
The region still experiences cholera outbreak each year during summer and post-monsoon seasons (Akanda et al. 2009;. On the other hand, rotavirus diarrhea, the most common type of diarrhea also occurs every year in the same region. The region mostly comprises the area of Bangladesh, a country with 160 million population and the most densely populated country in the world. The country has a tropical monsoon climate and most threatened by ongoing global warming, perhaps more than any other place of the world. In one end, the country experiences two major diarrheal disease outbreaks and on the other hand, it also has one of the highest vulnerable people on the earth due to the intersections of high population density, poverty, and effects of climate change. Therefore, in this study. For the consistency of the data, we converted and combined all three sources of data to z-score. The detail advantages of z-score in epidemic modeling were discussed in the later parts of the methodology section.

Weather Data:
We have utilized two types of weather data, one for validation and other for future projections.
For the validation of the spatial diseases models, we used the long-term observed time series from existing ground stations of Bangladesh. We obtained daily maximum (TMax) and minimum temperatures (TMin), and precipitation ( Homogeneity and quality control tests were conducted to ensure the removal of outliers. The tests were carried out using the RHtestsV4 software package which was developed by the joint CCl/CLIVAR/JCOMM Expert Team (ET) on Climate Change Detection and Indices (ETCCDI) (X. L. Wang & Feng, 2013). 129 For projecting the disease risk, we gathered RCM model-simulated data for the latest climate change scenarios. Climate data derived from the five available RCM outputs is selected for this study. The datasets were made available through COordinated Regional Climate Downscaling Experiment (CORDEX), a program that brought forth a collective effort to regional climate projections globally (Giorgi et al. 2009). The CORDEX aims to advance and coordinate the science and application of regional climate downscaling through global partnerships. The project defined some specific domains around the globe and invited communities to conduct regional downscaling in those designated domains.
Through the project's data portal, several RCM results became available over South Asia (CORDEX, 2015). As domain selection could be sensitive in a regional modeling study (Bhaskaran et al. 2012

Z-score:
To represent the diarrheal disease outbreak, we have incorporated the z-score metric instead of prevalence or incidence rate of a disease (Jagai et al. 2009). Both prevalence number of incidence rate is population dependent. From all the disease cases, we first converted them to z-score to avoid population effects. To remove population effect, normalization of data could be another option (Kao, 2009).In the normalization process, the data are kept to a fixed range, typically (0-1). However, in disease outbreak analysis, outbreak can vary in magnitude and do not follow the normal distribution. Thus, we utilized z-score to represent the disease risk in this study. The negative scores were scaled up to positive values for meaningful outbreak representation and to implicate log transformation for multivariate modeling. However, the relative magnitude can be biased for larger skewed values; thus we assumed the -3.5 as the minimum global state of the outbreak.
Hence, z-score was transformed to modified z-score by adding 3.5. Equation 1 shows the calculation of z-score for this analysis. As a mean state of the outbreak can be expressed as 0 in dual sign series, after transformation of the minimum value, we can assume that one 131 can be the epidemic threshold. In this study, we have attempted to quantify only the effect of climate on the diseases by any other influencing factors. Therefore, any element of the population will not represent the actual impact of the disease, and adjusted z-score would be a great way to overcome such problem. Similarly, as RCP scenarios are population independent, we have considered them in this study.
Where, X is disease cases or prevalence per month.

Seasons of the outbreaks
Earlier research showed that the annual epidemics of rotavirus and cholera over the Bengal delta occur during the winter and post-monsoon season respectively (Hasan et.al 2018, Akanda et. al, 2008. As the trigger of the outbreak correlated with climatic drivers, the rising phase of both diseases were analyzed in this study. The rising phase of rotavirus and cholera epidemics are the November-December and August-September months, respectively. It should be notated that in case of Cholera, the data we used are mostly from 1998 to 2003 period. During the period, the dominant cycle of the cholera outbreak was at post-monsoon season. Thus, we have considered the post-monsoon cycle as primary outbreak cycle.

Model development
By integrating the climate extremes and z-score value of the rotavirus diarrhea and cholera, we developed two spatial models for the rising phase of the disease outbreak. We have conducted multivariate analysis with 46 temperature and precipitation extremes taken from 132 Hasan et al., (2018) for both diseases. Additionally, as rotavirus outbreak occurs during winter where rainfall is low compared to other times of the year, we also incorporated relative humidity indices to develop the additional model for the rising phase of rotavirus.
However, relative humidity is unavailable in the RCMs results; hence, we avoided the improved RH model to conduct the future projections of the analysis.
We have selected multivariate analysis as our primary method to formulate the spatial models. However, the correlation values of multivariate analysis for top selected models are close for the different combination of the indices. Thus, we need to extend our analysis to temporal and spatial correlation analysis for individual variables to understand the influential role of each variable. Finally, we have combined the values of the temporal and spatial correlation by equation 2, to obtain a model score.
= * 1 + * 1 (2) = * 0.5 + * 0.5 Where, Corpr , Cortmp , Cortpr and Corsp is the correlation of precipitation index, temperature index, temporal analysis and spatial analysis, respectively. We adopted the model score from unified score proposed by Sikder et al, (2016) where they utilized several other performance matrix. In our model score, we provided equal weightage to the correlation of temporal variables. For example, for two variables, precipitation and temperature, we provided 0.5 weight to each correlation factor. For spatial analysis, we add an equal weight as temporal analysis thus provided a weightage of one. Therefore, the range of our model score would be 0 to 2, where 0 means the worse model and 2 means the best model. Based on the model score, we selected our final model from the pool of ten models for each 133 disease.

Bias-corrected extremes:
In this study, we have conducted bias correction on high-resolution RCM results to evaluate robust climate projections (Bennett et al. 2014). Detail of the bias-correction method was available in Hasan et al., (2017). The projections of bias-corrected extremes were further analyzed in the developed disease models under moderate RCP 4.5 and the strongest RCP 8.5 scenarios with respect to baseline climate. The changes of the z-score values from the baseline climates were presented as the future state of the diseases.

Result and Discussion:
Hasan et al., 2018 showed that the rainfall and temperature extremes could influence the rotavirus outbreak during winter. The study utilized z-scores to establish the relationship between rainfall and temperature extremes. However, the study was conducted with 1month moving average values where the crucial indicators from the relative humidity values were absent in the study. Moreover, studies found that the relative humidity could be a potential indicator of the rotavirus outbreak. Therefore, the role of relative humidity on rotavirus needed to be explored to understand the disease epidemics more clearly. On the other hand, the relationship between climatic extremes and z-score of cholera was never done in previous studies.
Using z-score and observed meteorological information, we conducted multivariate analysis for the rising phase of both diseases. For rotavirus, we have conducted two types 134 of multivariate analysis, with or without RH. The R2 values from the multivariate analysis were presented in Table 2, Table 3 and Table 4.
The top-ten best models for the rotavirus diarrhea that exclude the relative humidity indices show a short range of R2 values (the range is 0.03) in the multivariate analysis. The temperature variable, Tn16gTx30l (concurrent climate with two months moving average) is standard in all the top models, confirms its significant role in the rotavirus propagation. found the same influence of RR1 during the same phase of the diseases. Therefore, the model can be considered as the best model for two-variable multi-variate analysis.
We also developed rotavirus diarrhea models by considering relative humidity indices as another influential factor and the R2 values were shown in Table 3. Within top 10 models, 135 four types of temperature indices showed stronger relationship but all of them associated with the daytime temperature of the previous months. It should be noted that the models without RH show day and night time temperature range as the primary driver in the model, but that relationship is for no-lag 2-month moving average, where with RH, it is a 1-month lag with 1-month realization. For the rainfall indices, the variation between variables are quite high, but all represent small sudden rainfall amount. For the humidity indicator, the tops models are found mostly correlated with RH minimum values. Now from the combined score, we found that the CR5-Tx2932GE-RHmin performed best in the multivariate analysis. The correlation of CR5 and RHmin also confirm the significant relationship with the rotavirus outbreak phase. From an accuracy point of view, (R2 values) the model with RH performs better than the models without RH. However, the relative humidity variable is unavailable in the selected climate projections. Therefore, we selected the model without RH as our primary model for the future risk analysis.
The top-10 models of the cholera epidemic present R2 values less than 0.5, where the difference between the top and bottom model is around (0.01) ( Table 4). With such short difference, any model among the tens can be considered as the best model. Regarding temperature indices, the 2-month moving average of Tx2632GE with two-month lag shows significant influence in all top-ten models. It represents that day temperature between 26 degrees to 32 degrees is a potential trigger for the post-monsoon cholera outbreak. Akanda et al (2009) found that the increased temperature in water can accelerate the growth of the bacterial host of cholera within the water; thus Tx2632GE supports such phenomenon. On the other hand, an index like RR70 represents the high amount of rainfall, which leads to water overflow resulting in a connected fecal-oral route for the bacterial cycle. Moreover, 136 such water networks take times to influence the outbreak and could come to the effect after two months of such situation. Therefore, two month lag of RR70 can be considered as an appropriate indicator that can trigger post-monsoon cholera outbreak (Ryan et al. 1996).
From the combined score, (Table 4), it is evident that model no 9 is the best model among the top ten and it consists RR70 and Tx2632GE index as the driving variables. Figure 1 and Figure 2 represents the spatial distribution of model and observed rotavirus over Bangladesh. The spatial pattern of the model showed agreement with observed data over the central and southwestern part of the country. However, even with relative humidity indicator, the model result in the western part of the country deviates from observe. This bias can occur not only from unknown uncertainties but also from poor coverage of observed dataset. We used six available disease station data to generate the spatial maps where the coverage in the western part is scarce compare to eastern part of the country.
Thus, observed map might missing information which reflected in the developed models result.
The bias-corrected climate extremes were driven to the best cholera model for the threetime frames were presented in Figure 4. The change of risk of cholera was presented spatially over the Bengal delta. Under the RCP4.5 scenario, the change of cholera outbreak was minimum at early 21st century. However, with the progression of time, the outbreak increased toward the end of the 21st century. From the results, it is also found that the southern parts of the country will experience a higher rate of cholera than other parts of the country. With the deteriorating salinity problem, such change will create danger zone on 137 the southern locality.
In case of rotavirus, the spatial analysis suggests that rotavirus will decrease gradually toward the end of the century ( Figure 5). Due to the decrease in RR1 and increase in winter temperature, the rotavirus will have less wetness in the soil and lower amount of cold in the winter. This will create a hindrance in the disease propagation, thus the rate of the disease will decrease in the future years.
The time series for both of the diarrheal diseases over Bangladesh were presented in Figure   6. The mean change of risks for both selected scenarios were plotted in the Figure. Similar to spatial analysis, the Cholera risk showed a gradual increase in future years during postmonsoon season. The heighted risk are more certain in RCP8.5 scenarios than 4.5 scenarios. On the other hand, the risk of rotavirus overall decrease in 21 st century. The RCP8.5 shows much safer future than RCP4.5 for the disease.

Conclusion:
In summary, in this study, we conducted multivariate, temporal and spatial analyses to quantify the risk of diarrheal disease outbreaks in Bengal Delta using climatic extremes.
We have utilized five bias-corrected RCM results to project the disease risk for the 21st Century. From the study, the following conclusion can be made: For rotavirus, the multi-model analysis shows satisfactory performance using RR1 and Tn16gTx30l as driving variables. The 1-day monthly rainfall with 16°C to 30°C 138 temperature range, plays a critical role in triggering the outbreak. The observed spatial pattern suggests that the central region, Dhaka is the most vulnerable region among the country. The model able to capture the Dhaka outbreak, but it also over-estimated other regions of the country. The inclusion of relative humidity indices into the model, increase the performance of rotavirus outbreak prediction significantly, especially in the northeastern part of the region. However, as we do not have projected humidity data from RCM, we utilized the rainfall and temperature-driven model to project the disease risk. For the case of Cholera, the best model is driven by RR70 and Tx2632GE, which indicated the wet warming post-monsoon. In case of cholera, the projected disease risk map suggests that the southern part of the country will experience more risk of the disease in the future years. On the other hand, rotavirus outbreak is expected to decrease according to five selected RCM projections.
The projected disease risk can be utilized to conduct epidemic management and to improve vaccination strategy. A high-risk area can be given higher importance for the introduction of rotavirus vaccination. The decision maker and stakeholder can introduce new intervention strategy to improve the disease preparedness. In this study, we have introduced      Statistical Downscaling (ESD), applied over a limited area and driven by GCMs can provide information on much smaller scales supporting more detailed impact and adaptation assessment and planning, which is vital in many vulnerable regions of the world.
Global Climate Models (GCM) can provide us with projections of how the climate of the earth may change in the future. These results are the main motivation for the international community to take decisions on climate change mitigation. However, the impacts of a changing climate, and the adaptation strategies required to deal with them, will occur on more regional and national scales. This is where Regional Climate Downscaling (RCD) has an important role to play by providing projections with much greater detail and more accurate representation of localised extreme events.
Regional climate downscaling (RCD) techniques, including both dynamical and statistical approaches, are being increasingly used to provide higher-resolution climate information than is available directly from contemporary global climate models. The techniques available, their applications, and the community using them are broad and varied, and it is a growing area. It is important however that these techniques, and the results they produce, be applied appropriately and that their strengths and weaknesses are understood. This requires a better evaluation and quantification of the performance of the different techniques for application to specific problems. Building on experience gained in the global modelling community, a coordinated, international effort to objectively assess and intercompare various RCD techniques will provide a means to evaluate their performance, to illustrate benefits and shortcomings of different approaches, and to provide a more solid scientific basis for impact assessments and other uses of downscaled climate information.
The WCRP views regional downscaling as both an important research topic and an opportunity to engage a broader community of climate scientists in its activities. The Coordinated Regional Climate Downscaling Experiment (CORDEX) has served as a catalyst to achieve this goal.
As demonstrated at the second International Conference on Regional Climate -CORDEX 2013 held on 4-7 November in Brussels, Belgium, and co-sponsored by WCRP, the European Commission and IPCC, the CORDEX concept had gained maturity and was showing strong buy-in from the science community and VIA practitioners. To meet stakeholders' expectations the conference outcomes were followed-up to improve the experimental framework so as to improve the CORDEX framework.
At the third International Conference on Regional Climate -CORDEX 2016 held on 17-20 May in Stockholm, Sweden, and co-sponsored by WCRP, SMHI, Bolin Centre, FORMAS, ECRA, ESA, EUMETSAT, and APN, it was shown that CORDEX has contributed vastly to the development and production of regional climate data and information.