Comparison of bias correction methods for climate change projections in the lower Shivaliks of Punjab

The world's agriculture is highly affected by rapid population growth, land use changes due to urbanization, diminishing good quality water resources, and weather extremes occurring due to changing climatic conditions resulting from global warming (Asma et al. 2020). The understanding of the agricultural impacts of climate change may play a pivotal role in developing adaptation and mitigation strategies for tackling the adverse effects of such changes. The historical data analysis for the lower Shivaliks of Punjab, India has shown a significantly increasing trend in temperature (Kaur et al. 2021a). A range of emission scenarios with climate change projections has been a key tool in assessing the impact of future climate change and exploring management strategies to adapt to climate change. Many future climate change impact assessments have been carried out using crop models and hydrological models for specific locations and agricultural regions on a global scale (Rosenweig & Parry 1994).

The futuristic climatic data are available through General Circulation Models (GCMs), at very coarse scales generally more than 200 km in resolution (Feddersen & Andersen 2005). The GCM-generated data are further downscaled using Regional Climate Model (RCM) for the generation of fine-scale data resolutions of a particular area. The RCM data are more reliable compared to GCM data, however, RCM data still require bias correction (BC) when predicted data are compared to observed data for a particular period (Buonomo et al. 2007; Fowler & Blenkinsop 2007). However, it becomes of utmost importance to carry out the BC of the GCM projected data prior to the impact studies, for the correction of the discrepancies between a model's climate and observed historical climate data (Amsal et al. 2019; James et al. 2019; Hawkins et al. 2020; Jaiswal et al. 2021) for fair and reliable prediction and decision-making. There are several BC methods such as the difference method (DM), linear scaling (LS) or statistical methods (Leander & Buishand 2007), quantile mapping (QM) (Boe et al. 2007), and probability density functions (Piani et al. 2010) to correct the biases in GCM–RCM outputs for various impact studies (Charles et al. 2019; Enayati et al. 2021). There is no single BC method that performs best for all regions (Fang et al. 2015); each of these methods has its own merits and demerits, mainly due to spatiotemporal differences in surface and rainfall properties. Various case studies are available depicting the performance of different BC methods (Table 1). Hence, the selection of methods to remove the systematic errors of climate model outputs needs to be based on a comparison of multiple BC methods. Keeping this in view, the present study was undertaken to develop BC functions through statistical comparison of different correction methods, for seasonal projections of temperature and rainfall under different scenarios.

The study was carried out for the Regional Research Station, Ballowal Saunkhri located in the lower Shivaliks sub-mountain zone of Punjab, India which is commonly known as the Kandi region. Ballowal Saunkhri has been situated at 30°07′N, 76°23′E and lies 355 m above mean sea level. The area represents a semi-arid climate with very hot and dry summers from April to June, hot and humid conditions from July to September, cold winters from November to January, and a mild climate during February and March.

There are a series of models for which scenario-wise data are available. An earlier study by Punjab Agricultural University, Ludhiana has reported that CSIRO-Mk3-6-0 was best for the area adjacent to the current location as compared to other models (Kaur et al. 2021b). The daily data for the meteorological parameters like maximum and minimum temperature, rainfall, and solar radiation were downloaded from Marksim Weather Generator using the output of the CSIRO-MK3-6-0 model for the Ballowal Saunkhri location. The model-predicted data are available from 2010 to 2095 separately for four climate change scenarios (RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5). The historical daily weather data (temperature, rainfall, and solar radiation) recorded at the Agrometeorological Observatory of Regional Research Station, Ballowal Saunkhri during 2010–2020 was used for BC. The correction functions were estimated by using baseline and observed data from 2010 to 2015, whereas data from 2016 to 2020 were used for validation of corrected model data.

During the performance evaluation, different statistical parameters including coefficient of residual mass (CRM), normalized root mean square error (NRMSE), mean absolute error (MAE), mean bias error (MBE), mean absolute percentage error (MAPE), Wilmot d-index, percent bias index (PBIAS), and Nash–Sutcliffe efficiency (NSE) were used to check the performance of BC methods (Figure 1(a) and 1(b)). The CRM shows the difference in observed and predicted data relative to the observed data. Its zero, negative, and positive values indicate a perfect fit, over-, and under-prediction, respectively. NRMSE provides a measure (%) of the relative difference between predicted versus observed data. The simulation is considered excellent, good, fair, and poor if NRMSE is <10, 10–20, 20–30, and >30%, respectively (Jamieson et al. 1991). The MAE is an absolute measure of bias that varies between 0 to +∞. An MAE value close to 0 indicates an unbiased prediction, i.e. a low MAE indicates a good fit between two variables. The MBE provides information of over- and under-prediction by the model. Positive MBE indicates over-prediction, negative values indicate under-prediction, and a zero indicates equal distribution between negative and positive values (Singh et al. 2013). The MAPE is the mean or average of the absolute percentage errors of forecasts. Willmott (1981) proposed an index of agreement (d) as a standardized measure of the degree of model prediction error which varies between 0 and 1. The PBIAS, on the other hand, calculates the relative volume difference between modeled and observed volume. A negative value indicates under-prediction, whereas a positive value indicates over-prediction. The NSE is calculated as one minus the ratio of the error variance of the modeled time-series over the variance of the observed time-series.

GCMs provide a credible picture of global climate change, yet most of the crucial phenomena are considerably misreported. For instance, some of the key processes controlling climate systems are biased along with high uncertainties in the representation of changes in these phenomena. These biases then affect the representation of regional climate conditions. The biases in the climate model simulations and observed data need to be corrected in order to predict possible changes in the future. Although the spatial resolutions and skills of GCMs/RCMs have obviously increased in recent years, it is still not enough for their direct application in impact studies at local or site scales. Under such a case, BC methods provide a solution, which combines information from the local observations and simulations leading to smaller biases and higher-resolution predictions/projections. The correction functions developed for the maximum temperature, minimum temperature, rainfall, and solar radiation are given in Tables 2–4. Twelve different equations have been developed for each weather parameter separately as values of these weather variables differ for every month of the year. The study area, which is located in the subtropical zone, has experienced higher monthly fluctuations in weather characteristics. This area has both hot and cold weather extremes, as well as dry and wet spells. The goal of employing the equations independently for each month in the research region was to precisely assess and adjust the model-generated forecasted data. Hence, the futuristic corrected model data were obtained by using these equations for each month. All these methods are used to fetch model and observed data closer but out of these methods the difference method gives closer values of model data relative to the observed data, so a difference method is selected.

BC methods are post-processing tools for numerical modeling, which aim to improve the model agreement with the observations. Overall, data bias-corrected with either of the methods exhibited substantial improvements in all the statistical parameters compared to raw model data, but improvements were more pronounced for the difference method (Table 5). In the case of maximum temperature, the uncorrected model data gave NRMSE values of 14.1, 14.0, 13.8, and 13.6% under RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5, respectively, which were reduced to 6.47, 6.48, 6.37, and 6.51%, respectively, by the difference method (Table 5). However, other methods showed NRMSE values above 8%. Likewise, the above 23% NRMSE values for uncorrected model minimum temperature under different RCP scenarios depict poor simulation results. But the BC by the difference method reduced it to 8.2, 8.1, 8.3, and 8.1% under RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5, respectively, indicating excellent simulation results by this method (Table 6). All other BC methods lagged behind the difference method due to their higher NRMSE values (more than 11%). Similar results were observed for the uncorrected model rainfall data having high NRMSE values which were reduced more by the difference method. In the case of solar radiation, the NRMSE values (115.3, 115.2, 114.5, and 115.9% under RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5 scenarios, respectively) in the uncorrected model data were reduced by the difference method (16.6, 20.9, 21.9, and 21.3 under RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5 scenarios, respectively) after the BC (Table 7).

Similarly, uncorrected model data for maximum temperature, minimum temperature, rainfall, and solar radiation represented higher values of MAE, MBE, MAPE, and PBIAS, but low CRM and NSE values, depicting more errors in the uncorrected model data (Tables 5–7). All these statistical parameters were improved after applying different correction techniques for all the climatic variables. However, the difference method consistently and uniquely adjusted the mean and standard deviation correctly in comparison with other methods. The difference method followed by VS and LB methods gave good fits between observed and simulated temperature as indicated by various statistical parameters (Table 5). The difference method and VS corrected the biases of simulated temperature very well, but the better statistics of the difference method over the VS encouraged us to adopt it for further correction and projections of model data beyond 2020. Hence, this method was very suitable for removing bias from GCM/RCM at this location. Furthermore, in this study, which mainly focused on the lower Shivalik region of Punjab, the difference method is the best-performing one for removing the RCM/GCM bias compared to the other methods.

The distribution of the annual temperature over the lower Shivalik region indicates that considerable biases exist between uncorrected model outputs which ultimately advise against the direct application of model projections in climate change studies over this region. From the BC for weather parameters by different correction methods, the difference method has been found to be better as compared to other methods in terms of statistical results. Therefore, this method was selected as the method of correction for the future data under all the RCP scenarios. The corrected model data for 2016–2020 were validated by its comparison with the observed data for that period. The results of the comparison between the monthly averages of actual, model, and corrected data for maximum temperature, minimum temperature, rainfall, and solar radiation under RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5 scenarios have been shown as follows.

As discussed earlier, the difference method was selected for the correction of futuristic data (2021–2095) under different RCP scenarios (Table 8). The annual and seasonal perusal of projected maximum temperature, minimum temperature, rainfall, and solar radiations under different RCP scenarios for 2021–2095 showed that the maximum temperature has been expected to increase by 1.3–8.3 °C, 1.2–2.7 °C, 1.7–3.3 °C, 1.0–2.7 °C, and 1.2–2.7 °C and for the annual, kharif, rabi, winter, and monsoon seasons, respectively, during 2021–2095. The minimum temperature (2021–2095) has been expected to increase between 0.7 and 2.6 °C, 0.9 and 2.7 °C, 0.8 and 2.6 °C, 0.6 and 2.5 °C, and 1.0 and 3.0 °C for the annual, kharif, rabi, winter, and monsoon seasons, respectively. The difference in future temperature change for different seasons has been observed over the lower Shivalik region. The projection results denote a lower increase in the maximum and minimum temperature during kharif season as compared to rabi season, i.e. hotter months like July and August are expected to have more increment in maximum and minimum temperature as compared to the colder months such as January, February, and December. You et al. (2014) found a pronounced warming trend for T_mean, T_max, and T_min in winter months with RCP8.5 in China during 2061–2090. Salman et al. (2018) and Almazroui et al. (2020) also revealed higher increases in temperature during winter months but lower in the summer months under future emission scenarios over the Arabian Peninsula. During 2021–2095 rainfall has been expected to increase by 22.4–301 mm, 107–238 mm, 43.7–66.9 mm, 17.3–45.1 mm, and 82.7–168.9 mm for the annual, kharif, rabi, winter, and monsoon seasons, respectively. The solar radiation for the annual, kharif, rabi, winter, and monsoon seasons during 2020–2095 is expected to be increased by 0.6–1.0, 0.8–3.0, 0.9–1.4, 0.4–1.2, and 0.8–1.1 MJ/m²/day, respectively. From the data, it has been clear that more increase in maximum temperature and minimum temperature has been expected under the RCP 8.5 scenario followed by RCP 4.5, RCP 6.0, and RCP 2.6 scenarios, but rainfall and solar radiation has been expected to increase under RCP 8.5 followed by RCP 6.0, RCP 4.5, and RCP 2.6 scenarios. Abaurrea et al. (2018) also revealed that the temperature projections over Spain under RCP6.0 are smaller than those under RCP4.5 during 2031–2050 but a similar growth during 2051–2060. Likewise, a study conducted by Wang et al. (2021) also demonstrated more increase in temperature under the RCP 4.5 scenario over the coastal region of China. Similarly, an increase in the maximum temperature (under RCP 6.0 and RCP 8.5 scenarios) of 1.7–2 °C by the 2030s and 3.3–4.8 °C by the 2080s has been reported in India by Chaturvedi et al. (2018). These temperature changes may be yield limiting for the future which may increase the risk of food security in the region. Hence, different adaptation strategies pertaining to management practices may be identified to sustain the yield of crops in the future.