1 Introduction

The Indian Summer Monsoon (ISM) is mainly associated with synoptic-scale systems, which bring copious rainfall over the east and west coast of India. Climatologically, the Bay of Bengal basin is more active in terms of genesis of the monsoon synoptic scale system, contributing about 45% of the seasonal rainfall over the Indian region. The disruption of everyday life and loss of vital national assets due to extreme weather events over the South Asian regions are on the rise. These synoptic-scale systems (MD and DD) generally originate over the Bay of Bengal and propagate towards coastal India along with their intensity changes (Hunt et al. 2016). The numerical weather prediction (NWP) model has been recognized as the front-runner in the operational weather forecasting of heavy rainfall events; therefore, it is imperative to examine the performance of these operational models in order to improve their skills. In the latest decade, the regional NWP models have successfully resolved sub-grid scale processes using various parameterization schemes and determining isolated location-specific heavy rainfall events across the globe (Dudhia 2014). However, the exacerbation of the input and model-based uncertainties at finer spatial resolutions was strongly influenced by the region, time of the study, forecast time scale, spatial resolution, and the evaluation metrics (Prein et al. 2015). In addition, the scarcity of high-resolution observations over the Indian region and its surroundings is a significant constraint to predict local severe weather events with appropriate temporal, spatial, and intensity accuracy (Kumar et al. 2014). However, despite these constraints, there is an urgent need to improve the skills of localized rainfall events due to the rapid increase in the frequency and intensity of localized heavy rainfall events over the Indian region in recent years (Baisya et al. 2017).

These studies provide an insight into the capability and limitations of NWP models to predict these such drastic high impact events (heavy rainfall, cloud bursts, thunderstorms) causing heavy fatalities and damage to property, particularly during ISM seasons. The primary restriction found in previous studies is the inability of NWP models to resolve convection initiation at the right time and place (Cintineo et al. 2014; Sun et al. 2014). Further, dynamics and thermodynamics play an important role in predicting ISM rainfall over a point location with adequate lead time (Sisodiya and Pattnaik 2021). Interactions of different grid-scale processes and a proper representation of the same in NWP models is still a challenge and most sought after by the model developing community and operational forecaster, and it needs a multipronged long term approach to address the same effectively (Roy et al. 2007).

Further, intercomparisons of results across studies are complicated because different studies use different metrics of variability, and few have systematically addressed the comparisons across spatial and temporal scales. A deterministic forecast skill does indicate the forecast sensitivity to its initial condition and parameterization approximation errors (Das et al. 2008). Considering the errors in the propagation of these large-scale systems, the displacement error can lead to higher false prediction and missing the accurate deterministic forecast at the correct place. Thus, ensemble prediction is becoming increasingly important and popular on all timescales from the short (one or three days ahead) to the seasonal range and beyond (Ancell 2013; Palmer 2019; Bouallegue et al. 2020). An ensemble prediction system is an NWP system in which several predicted scenarios from the same model represent the uncertainty in a weather forecast and the most likely outcome. One of the main goals of the ensemble forecast is to provide the probabilistic estimates of the weather events, which requires its resolution to be sufficiently high to resolve the small-scale features of these events. Thus, an ensemble forecast provides a detailed picture of the range of possible future weather states consistent with our knowledge of the current state. However, very limited studies have been carried out using this technique to predict localized rainfall events over the Indian region during ISM (Ashrit et al. 2020). In this context, no studies have been carried out so far to examine the region-specific forecast skills of the operational forecasts of IMD. These synergetic exercises are essential for the development of better models and techniques for the Indian region. This study is an attempt in that direction.

The hyper-local forecast is heavily dependent on the parameterization schemes to its resolve dynamics and thermodynamics (cloud microphysics, PBL, and convection) and inaccurate representation of these fine-scale processes in global models leading to large errors in quantifying the amount, intensity, and location of rainfall and typically miss heavy rainfall events or falsely predict rain (Bechtold et al. 2004; Holloway et al. 2012). More recently, Vogel et al. (2018) suggested that parameterizing such processes is the potential cause of low skill in the individual forecast than ensemble prediction systems. In this study, a comprehensive analysis of the performances of PBL and CM parameterization schemes (deterministic and ensemble approach) has been investigated with thrust on rainfall forecast for MD and DD moving inland over the eastern coast of India using the WRF model. Further, a comparison of the model is carried out with IMD operational forecast outputs for 3 MD cases to examine its skills and the role of parameterization to determine the fidelity of the model in predicting the rainfall at the local scale.

The objective of this study is to evaluate the multiple physics ensemble prediction systems for improving rainfall forecast at the local scale (over Odisha) for different rainfall intensities (i.e., light rain, moderate rain, heavy rain, and very heavy rain) with lead time up to 72 hrs (day-3). Further, this study will assess the prediction skills of individual deterministic and IMD operational WRF model compared to forecast generated through multiple physics ensemble frameworks. In addition, different major rain-bearing synoptic-scale systems (i.e., MD, DD, and post-MDs) contributing to the rainfall over the study region are comprehensively investigated to test the robustness of the ensemble multiple physics techniques. The manuscript is segregated into five sections, including data and methodology (section 2), results and discussions (section 3), and summary (section 4).

2 Data and methodology

2.1 Model configuration

Weather Research and Forecasting (WRF-ARW 4.0) model is used to carry out all the simulations. The WRF is a compressible, non-hydrostatic model designed for both research and operational applications (Skamarock et al. 2019). The model is configured with two interactive nested domains (9 and 3 km, figure 1a). All simulations are carried out for 72 hrs (up to day-3) with hourly model forecast output. All numerical experiments using the WRF model have identical configurations except for twenty-five combinations of parameterization schemes (figure 1c). The Kain Fritsch scheme (Kain 2004) was used for cumulus convection for the outer domain (9 km), and the inner domain is explicitly resolved. Table 1 illustrates a detailed configuration of the model setup. This study focuses on evaluating model forecast for Odisha and hyper-local scale over smart city Bhubaneswar (figure 1a).

Figure 1
figure 1

(a) Nested domain for WRF simulations, (b) IMD tracks of all the cases, and (c) combination of PBL and CMP schemes used in the simulations.

Full size image
Table 1 WRF configuration.
Full size table

2.2 Datasets

A total of 21 cases are considered in this study (10 MD, 8 DD, and 3 post-MDs) (table 2). Total 525 simulations (21 cases × 25 simulations for each case) are carried out for this study, and the results are presented. The cases are chosen based on their propagation path, which has caused heavy to very heavy rainfall over Odisha while progressing towards the Indian mainland (figure 1b). European Centre for Medium-Range Weather Forecasting Re-analaysis5 (ERA5) data at six hourly intervals are used as model initial and boundary conditions (Hersbach et al. 2020). IMD observed rainfall data are also used to validate model results at a daily temporal scale and further used for computing various verification skill scores (Pai et al. 2014). Further, global precipitation measurement (GPM) half hourly IMERG rainfall (Huffman et al. 2014) is used to analyze the intensity, duration, and timing of rain events (Huffman et al. 2014). Further, Odisha Block level daily rainfall data for Khurda district (consisting of 10 blocks) is used to validate the model at hyperlocal stations for Bhubaneswar smart city and neighbourhood. Finally, IMD operational WRF model forecast (up to day-3) at 9 km resolution was also used for comparison.

Table 2 List of the cases taken for simulation.
Full size table

2.3 Multi-physics ensemble approach

There is 25 combination of deterministic models simulated for 21 cases (table 3). Thereafter, three multi-physics ensembles outputs (E1, E2, and E3) are designed to investigate its performance along with the optimal combination of deterministic model forecasts. The ensemble E1 and E2 are obtained from YSU and MRF PBL parameterization schemes (best-performing PBL), respectively, along with all five microphysics parameterizations. Along with these two best PBL schemes, all MP schemes combinations are incorporated to create E1 and E2. This optimized combination has been designed due to the strong influence of MP on rainfall forecast. Apart from that, the third ensemble (E3) is created using the top-performing five deterministic models based on their rainfall forecast skills. Details about the top five deterministic and three ensemble members are provided in table 4. All the results discussed in the following section are from the inner domain (3 km) for 525 simulations. A dichotomous forecast approach is used to evaluate the model forecast. This approach is based on binary outcomes as one (1) if the event occurs and zero (0) if the event does not occur. Detailed descriptions about different verification skill scores used in this study are presented in table 4.

Table 3 Various combinations of deterministic model members.
Full size table
Table 4 Verification skill scores.
Full size table

3 Results and discussions

3.1 Rainfall

The validation results of model forecast output with the IMD gridded rainfall product for all the simulations are discussed in this section. Figure 2(a–o) shows the box and whisker plot for verification skill score of daily accumulated rainfall of model simulations up to day-3. The box and whisker plot shows the variability in skill score by showing percentile values at 5, 25, 50 (median), 75, and 95th percentile values for all the 525 cases. The rainfall verification skill scores are computed based on IMD daily rainfall classification for light rain (LR), moderate rain (MR), heavy rain (HR), and very heavy rain (VHR) with the corresponding daily accumulated rainfall criteria of 7.5, 35.5, 64.5, and 125.5 mm, respectively. Here the aim is not only to validate the rainfall skills at large but also to examine the skills of the forecasts for different criteria of rainfall used for disseminating the operational forecast to the public on a daily basis. All verification skill scores are computed over the Odisha state, and IMD rainfall gridded data (25 km) are brought to model horizontal resolution (3 km) by using the bilinear interpolation method for carrying out the validation. These analyses will provide comprehensive verification measures for model deterministic and ensemble forecasts in terms of rainfall characteristics, namely intensity, amount, distribution, timing, and location.

Figure 2
figure 2

Box and whisker plot of skill scores showing column wise Equitable Threat Score (ETS), Critical Success Index (CSI), Probability of Detection (POD), False Alarm Ratio (FAR) and Bias of rainfall for all 525 simulations.

Full size image

It is seen that equitable threat score (ETS) and other skill scores decrease from day-1 to day-3 and as we move to lower to higher rainfall thresholds (figure 2a, f, k). The ETS provides us with the magnitude of correctly predicted rainfall events under a certain threshold, which is necessary for today to correctly predict the hyperlocal rainfall in terms of intensity, location, and time of occurrence. The maximum values are up to 0.4 (95 percentile values), above which the values are considered outliers (figure 2a, f). It is to note that few combinations of parameterizations show values below 0.1 for the same day and same rainfall thresholds values, suggesting the impact of the choice of parameterization on rainfall predictions. It is also noticed that ETS and Bias for day-1 show the robust performance of the model (ETS > 0.2 and bias < 2) for all rainfall thresholds (figure 2a, e).

Further, the critical success index (CSI) and probability of detection (POD) are also shown in figure 2 in columns 2 and 3, respectively. These two skill scores considered the forecast that counts by removing the correct false alarms. The CSI can be viewed as the fraction of observed and simulated events correctly predicted, thus ignoring the correct negatives. Figure 2(b, g) show up to 70% (0.7) of the LR events that occur are correctly predicted over Odisha; however, in the same category, the skill for day-3 and VHR, respectively, is very poor. These results highlight the poor reality of the models in predicting the high amount of rainfall that too with a lead time of day-3 particularly at the local scale; however, due to the rise in extreme rainfall events, there is a need to improve these skills on priority for researchers and operational forecasters. In addition, to statistical skill score results, here we also present the performance of the model in replicating the special distribution of rainfall over the study region. We have presented one monsoon case (case 04, i.e., 07–10 August 2018) out of 525 cases as an example to show the spatial distribution of day-3 accumulated model forecasted and IMD observed rainfall (figure 3). As shown by IMD, it is seen that the monsoon DD has contributed over 300 mm rainfall in 3 days over the Coastal Odisha region (figure 3z). Further, it is found that most of the deterministic simulations have successfully captured the event at large over coastal Odisha. However, the distribution of rain significantly varies with the variety of schemes, and out of these, YSU (figure 3x) and ACM2 (figure 3d) combined with Thompson microphysics simulation have the highest resembles with IMD rainfall (figure 3z). In addition, ensemble spatial distribution was added in the analysis for analyzing the spatial distribution of the multi-physics ensemble approach. It is noted that all ensemble members (E1, E2, and E3) improved the spatial distribution for intense pocket (rainfall > 200 mm) for the study domain. The distribution is significantly approved in ensemble members (E1, E2, and E3) as compared to some of the combinations overland (BouLAC and Goddard, BouLAC and WDM6 and MRF and Thompson, respectively). This result is for a specific case, in general, from comprehensive validation suggests, ensemble members (E1, E2, and E3) and deterministic forecast using YSU and MRF PBL combinations along with Thompson and WDM6 microphysics provided the best result.

Figure 3
figure 3

(a–z, E1, E2, E3). Spatial distribution of accumulated rainfall for all 25 physics combinations validated against IMD gridded rainfall data. The case shown here is a monsoon depression case, i.e., Case 04 (07–10 August 2018).

Full size image

At this stage, we are introducing the results for ensemble forecasting and discussing the same compared to individual models. Statistical methods for category (yes/no) forecasts, such as the threat score, are typically used to verify precipitation forecasts. However, these standard methods are affected by the problem caused by slight displacements in either space or time concerning the observations. Hence, spatial techniques have recently been developed to help solve this problem. In this context, the fractional skill score (FSS) is computed to assess the efficiency of the spatial distribution of rainfall for all the cases considered in this study. FSS is a spatial verification measure used to assess the performance of precipitation forecast from the model output. Conceptually, the spatial distribution of events within a small area is treated probabilistically rather than deterministically. An event is the binary (yes or no) occurrence of rainfall. For any grid size, FSS considers a perfect forecast to be one with the same frequency of events as observed within the same grid point, irrespective of their particular placement within the grid. The FSS is computed as the Fraction Brier Score (a variation of the Brier skill score used to verify probabilistic forecast), divided by the sum of the mean squared forecast and observed fractions. Previous studies have shown that the FSS can directly measure the error in the placement of the rain, and a value above 0.5 indicates a good forecast (Roberts and Lean 2008).

Figure 4(a–d) shows the box and whisker plot of FSS and the top five performing schemes, along with the three ensembles forecasts (table 5) for all the 525 simulations. Here the FSS scores are segregated based on 13 MDs (figure 4a, b) and 8 DDs (figure 4c, d). In general, it is seen that FSS values decrease from day-1 (0.61) to day-3 (0.35) by the median values in the box whisker plot. As mentioned before, based on the performances of two PBL schemes (YSU and MRF), ensemble forecasts (E1 and E2) are designed including all cloud microphysics schemes, further, ensemble E3 is created by considering the best five deterministic member models (table 5). For DD cases, the FSS of ensemble forecasts showing a significant improvement (E1: 0.63, E2: 0.65 and E3: 0.62) for the day-2 rainfall compared to all the deterministic models with the highest skills of M03 (Model number 3 out of best 5 models, table 5) with skill 0.6 (figure 4d) highlighting that the ensemble members have improved the spatial distribution of rainfall compared to individual deterministic members. We have also noted similar pattern is maintained up to day-3 (i.e., E1: 0.51, E2: 0.57, E3: 0.55, and M03: 0.54). The results for the 13 MD cases are in a similar pattern to that of DD except for the skills of the ensemble forecast has been slightly reduced throughout day-1 to day-3 but better than best deterministic models (i.e., M01 and M03). The skills for day-2 (i.e., E1: 0.55, E2: 0.56, E3: 0.57 and M01: 0.58) and day-3 (i.e., E1: 0.41, E2: 0.4, E3: 0.4 and M03: 0.37) for respective models suggest the performance of multi-physics ensemble is superior to any deterministic model simulations (figure 4b).

Figure 4
figure 4

(a–d) Box and whisker plot of Fraction Skill Score (FSS) of all the cases considered in this study. Second column represents the best five scheme among the 25 combinations used along with the three ensembles members.

Full size image
Table 5 Best five deterministic model and three ensemble members chosen based on rainfall validation.
Full size table

Currently, many societal applications need rainfall forecast at higher temporal resolution (1 hrly); therefore, efforts are made to examine the average skills at hourly rain rate for both the MDs and DD simulations for the 72 hr duration validated against GPM estimates (figure 5a, b). It is found that deterministic models do not show any distinct improvement analysis averaged for the entire inner domain with maximum skills for MD (DD) in terms of standard deviation and the correlation coefficient is 2.7 (1.5) and 0.32 (0.62) for M04 (figure 5b). Whereas ensemble forecast demonstrates a better standard deviation and correlation coefficient skills. For MDs (figure 5a), ensemble forecast (standard deviation, correlation coefficient) are E1 (2.58, 0.62), E2 (2.9, 0.43) and E3 (4.5, 0.44). Similarly, for DD, E1 (1, 0.62), E2 (1.3, 0.71), and E3 (1.4, 0.57), suggesting that multi-physics ensemble forecast has performed better than any deterministic models consistently even at 1 hr rain rate temporal scale. This is also to highlight that E1 has performed better compared to the other two ensembles (E2 and E3), suggesting that ensembles created through combinations of top-performing deterministic models might not always have proven to be best, rather an ensemble created based on optimal physics (i.e., E1) might have better hand particularly in terms of rainfall forecast over the region of study.

Figure 5
figure 5

Taylor diagram of hourly rainfall rate validated against GPM half hourly rain data for all the simulations showing best five deterministic and three ensemble members (a) MD and (b) DD.

Full size image

3.2 Block level rainfall

The city of Bhubaneswar is one of the smart city of India and always suffer from water logging issues due to monsoon rainfall. In this context, we attempted to validate rainfall forecast at the hyper-local block-level stations in Khurda district up to day-3 lead time. Ten block-level rainfall observations with daily rainfall are used for validation (figure 6). Results discussed in this section are obtained from 228 simulations (i.e., 156 MDs, 72 DDs) consisting of top five deterministic models (i.e., M1–M5) and three ensembles (i.e., E1–E3) for 19 cases combined (station data not available for two DDs, i.e., Case 7 and 8). As the block-level datasets are available on the accumulated daily scale (24 hrs), discussions are restricted on the daily scale only. Here model forecast is bi-linearly interpolated at the station locations for validation. Figure 7(a, c, e) shows the difference for MD cases (156 simulations) averaged for thirteen cases for ten stations for five best deterministic and three multi-physics ensemble members. It is noted that the average bias is between 15–40 mm/day. Specifically, the ensemble model (E2) has produced the least bias of 15 mm/day compared to the deterministic model M5, which produced the least bias of 18 mm/day on day-3. Similarly, for DD cases, it is noted that the average bias of forecast is between 8–30 mm/day, and the least bias forecast is obtained from the E2 with a value 8 mm/day compared to deterministic model M2 that has produced a bias of 11 mm/day for day-3. A similar distinct pattern of superior performance is noted for ensemble models for day-1 and day-2, suggesting that localized rainfall forecasts better captured these optimal ensemble combinations compared to any deterministic models. Among the ensemble, E1 and E2 have the best performing models univocally, again re-emphasizing the optimal combination of physics ensemble is better than the just ensemble of best deterministic models. In general, the skills of DD cases are better than MDs.

Figure 6
figure 6

Ten block station name and location in Khurda district (Odisha).

Full size image
Figure 7
figure 7

Station rainfall difference (MODEL–OBS) for daily accumulated rainfall up to day-3 for best five deterministic and 3 multi-physics ensemble members. Left column represents monsoon depression (MD) and right column deep depression (DD) cases respectively.

Full size image

3.3 Validation with IMD operational forecast

It is always essential and challenging to provide an objective way to determine the performance of the real-time forecast of operational agencies. In this context, efforts are made to examine the performance of high-resolution model forecasts (up to day-3) obtained from IMD for 3 MDs with exact initial condition dates and forecasting duration. Three MD cases considered are case-3, case-4, and case-5 (table 5, based on IMD data availability). It is to be noted that IMD real-time forecasts are available for 9 km resolution; therefore, those data are interpolated to 3 km resolution (our model grid) for carrying out verification skills. The validation is made based on 4 rainfall categories, i.e., LR, MR, HR, VHR, which are extensively used for operational forecasting. Overall, it is seen that 3 km simulated ensemble models consistently perform better than the 9 km operational IMD product in terms of ETS and FAR for all rainfall categories and throughout day-1 to day-3 (figure 8). This result re-emphasizes the need for an optimized multi-physics ensemble for better rainfall forecasting. Further, it is noted that ensemble scores for LR forecast skills scores such as ETS, FAR, CSI, POD, and BIAS are found to be day-1, i.e., E1 (0.23, 0.25, 0.57, 0.62 and 1.5), E2 (0.24, 0.25, 0.55, 0.61 and 1.5), E3 (0.23, 0.25, 0.55, 0.61 and 2.5) and IMD (0.12, 0.35, 0.58, 0.83 and 2.0), suggesting that ensemble forecast consistently have higher skills in terms of rainfall prediction compared to IMD for most of the skill score from day-1 to day-3 for all rainfall thresholds. This is also to mention that IMD skills have an edge in POD for the LR category for day-1 and day-2 compared to ensemble models. This might be attributed to the incorporation of data assimilations being carried out for the operational model reducing the spin-up issues in that model compared to in-house ensemble models. However, except for this component, in all other categories, the ensemble is consistently provided higher skills at block level; particularly remarkable improvement is a note for CSI index, which is one of the key indices for determining the model performance in terms of rainfall prediction. The results are highly satisfying and encouraging.

Figure 8
figure 8

Skill scores (column wise from left to right: ETS, FAR, CSI, POD and Bias) calculated for best performing models (M01 to M05), three ensemble members (E01 to E03) and IMD NWP product.

Full size image

4 Summary

In this study, efforts are made to demonstrate the skills of optimal combination of multiple physics ensemble forecasting to improve the rainfall forecast skills at a fine spatiotemporal scale over the state of Odisha. Total 525 simulations using 5 PBL and CMP parameterization schemes are carried out over the state of Odisha using the WRF model for 21 cases of 13 MDs and 8 DDs, which has contributed a substantial amount of rainfall over the region. Extensive verification exercises are carried out using benchmark statistics such as FAR, CSI, FSS, POD, Taylor diagram, ETS, and CC for four categories of rainfall (i.e., LR, MR, HR, and VHR) at finer spatiotemporal resolution (i.e., block-level to the state level and daily to hourly scale). The IMD (daily), GPM (1/2 hourly), and Odisha block-level datasets are used for validation purposes. The aim is to explore the possibility of providing the best possible skills at the finer horizontal scale that can be practically utilized for multiple societal applications, primarily towards disaster mitigation, preparedness, water management, and agricultural practices. Another major goal of this work is to develop and demonstrate the ability of ensemble forecast products in quantifying rainfall over the finer scale with adequate lead time up to day-3. In this context, two types of ensemble forecasts are prepared using optimized parameterization schemes and the best combination of deterministic models. Here to mention that there are three best deterministic models (i.e., R24, R13, R25), which are part of an ensemble combination of best models as well as an optimized ensemble of parameterization. Further, for the first time, it is attempted to actually validate operational IMD forecast with in-house developed models to verify its fidelity in terms of rainfall forecast skills over the study region. Here the only thrust of this study is to examine the performance of the model in terms of rainfall forecast as it is which is most required by the society and policymakers, but unfortunately, our models always struggle to provide satisfactory skills at fine-scale with adequate lead time.

It is comprehensively demonstrated from 525 simulations that YSU with Thompson and MRF with Morrison are two combinations of deterministic model configuration that have provided the best results consistently across different rainfall categories (i.e., LR, MR, HR, and VHR) with a lead time up to day-3. Further, it is noted that apart from these two best combinations, we have noted three more combinations in the top five list of models, and those are ACM2 with WDM6, Boulac and Thompson, and YSU and WDM6. Afterwards, three ensembles are created and validated against these best five models for rainfall obtained from IMD (daily) and GPM (1/2 hrly) datasets over Odisha as well as at the block level station data obtained from Odisha Government. It is found that three ensembles have consistently performed better in terms of rainfall intensity, amount, distribution, and location throughout the forecast duration up to day-3. Further, it is also noted that E1 has performed better compared to E2 and E3, suggesting that an ensemble created through an optimized combination of physical parameterization is better than the just ensemble of best models. In general, it is found that the forecast skills of DDs are better than MDs, and this might be attributed to the better mesoscale organization in DDs, which is always better replicated by the models.

Finally, IMD operational forecast data valid up to day-3 are obtained for 3 MDs compared with our best 5 deterministic models as well as 3 ensemble models for four categories of rainfall. It is found that ensemble results consistently performed better up to day-3 over the region. However, it is also noted that the POD skills of IMD forecast are marginally better than in-house ensemble and deterministic models, particularly on day-1 and day-2. This might be attributed to the induction of data assimilation in IMD models, and only 3 MDs cases datasets are used for validation. Except for this limitation, the ensemble models have consistently performed better over the region not only up to day-3 but also for different categories of rainfall which is operationally forecasted in real-time. We would like to address specific mechanisms responsible for the outcome of these results in a more detailed manner in the future to provide some new insight reasoning for these results. This study has demonstrated the usefulness of an optimized combination of multiple physics parameterization in predicting rainfall at a finer spatiotemporal scale over the Odisha region. The findings of this study can be helpful towards improving local-scale rainfall forecasts for efficient disaster preparedness and management. Following are the major conclusions drawn from this study:

  • The multiple physics ensemble prediction framework significantly improves the forecast skills of localized rainfall in terms of its amount, distribution, intensity, and timing over the study region.

  • Compared to the regular ensemble approach, the optimized multiple physics ensemble technique has captured significant uncertainties of localized rainfall characteristics contributed by different synoptic-scale systems for various rainfall thresholds resulting in better forecast skills.

  • This prediction technique has used high-resolution forecast datasets obtained from the innermost domain (3 km) shows an improvement in terms of rainfall verification skill scores compared to IMD operational forecasts, thus indicating the advantage of higher resolution NWP products.