Improving Indian summer monsoon rainfall prediction using deep learning up to two years in advance

ABSTRACTThe present study compares the Indian summer monsoon rainfall (ISMR) prediction skill of monsoon mission climate forecast system version 2 (CFSv2‐T382) with that of the seasonal prediction models participating in US National Multi‐Model Ensemble (NMME) project. In general, the present‐day models simulate cooler than observed sea surface temperature (SST) in majority of the Tropics and extratropics. The model rainfall has strong dry bias over major continental regions and wet bias over tropical oceans. Meanwhile, prediction of the boundary forcing such as SST is essential for driving the atmospheric response through teleconnections. It is noted that even though the prediction skill for SST boundary forcings like El Niño‐Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) is not at the best in CFSv2‐T382 compared to a few of the NMME models, it shows better skill for ISMR hindcasts initialized at 3‐month lead time (February IC). This may be attributed to the better teleconnection pattern of ENSO and IOD in CFSv2‐T382, which has minimum biases in equatorial Indo‐Pacific region. It also has a better ISMR–SST teleconnections in the Tropics with a pattern correlation of around 0.6. In many of the NMME models, the better prediction skill of the inter‐annual variability of SST indices is not transformed into the improvement of ISMR skill through teleconnections. It is therefore concluded that having good prediction skill for major SST boundary forcings is not sufficient, but capturing the appropriate teleconnections of these SST boundary forcings in the model is critical for the better prediction of ISMR. The study points out that the present‐day seasonal prediction systems need to be improved in their simulation of tropical SST–monsoon teleconnections, which can improve the seasonal prediction skill of Indian summer monsoon further. One area where the immediate focus is required is the Indian Ocean SST and ISMR teleconnection.

Eurasian snow is one of the slowly varying boundary forcings known to have significant influences on the mean and variability of the Indian summer monsoon rainfall (ISMR). A multilayer complex snow scheme, incorporated into the state‐of‐the‐art coupled Climate Forecast System version 2 (CFSv2) showed significant improvements in the simulation of mean ISMR, snow, and Northern Hemisphere surface and tropospheric temperature. Here we show that a realistic simulation of high‐latitude snow decreases the north–south temperature gradient, which in turn decreases the meridional transport of energy from the Equator to the Pole, consequently affecting the tropical sea surface temperature (SST) and air–sea interactions. The global teleconnections of the ISMR with SST and 2 m temperature over land are also improved considerably in association with improved simulation of the oceanic natural modes of variability. Our findings provide new insights for the relationship between the winter Eurasian snow and the following ISMR, namely that the same relationship may be understood through a framework of meridional atmospheric energy transport and its effects on the tropical air–sea interactions. The improvements in the global teleconnection in the modified version of CFSv2 may have implications in the ISMR predictability and prediction skill.

Century-long efforts have been devoted to seasonal forecast of Indian summer monsoon rainfall (ISMR). Most studies of seasonal forecast so far have focused on predicting the total amount of summer rainfall averaged over the entire India (i.e., all Indian rainfall index-AIRI). However, it is practically more useful to forecast anomalous seasonal rainfall distribution (anomaly pattern) across India. The unknown science question is to what extent the anomalous rainfall pattern is predictable. This study attempted to address this question. Assessment of the 46-year (1960–2005) hindcast made by the five state-of-the-art ENSEMBLE coupled dynamic models’ multi-model ensemble (MME) prediction reveals that the temporal correlation coefficient (TCC) skill for prediction of AIRI is 0.43, while the area averaged TCC skill for prediction of anomalous rainfall pattern is only 0.16. The present study aims to estimate the predictability of ISMR on regional scales by using Predictable Mode Analysis method and to develop a set of physics-based empirical (P–E) models for prediction of ISMR anomaly pattern. We show that the first three observed empirical orthogonal function (EOF) patterns of the ISMR have their distinct dynamical origins rooted in an eastern Pacific-type La Nina, a central Pacific-type La Nina, and a cooling center near dateline, respectively. These equatorial Pacific sea surface temperature anomalies, while located in different longitudes, can all set up a specific teleconnection pattern that affects Indian monsoon and results in different rainfall EOF patterns. Furthermore, the dynamical models’ skill for predicting ISMR distribution primarily comes primarily from these three modes. Therefore, these modes can be regarded as potentially predictable modes. If these modes are perfectly predicted, about 51 % of the total observed variability is potentially predictable. Based on understanding the lead–lag relationships between the lower boundary anomalies and the predictable modes, a set of P–E models is established to predict the principal component of each predictable mode, so that the ISMR anomaly pattern can be predicted by using the sum of the predictable modes. Three validation schemes are used to assess the performance of the P–E models’ hindcast and independent forecast. The validated TCC skills of the P–E model here are more than doubled that of dynamical models’ MME hindcast, suggesting a large room for improvement of the current dynamical prediction. The methodology proposed here can be applied to a wide range of climate prediction and predictability studies. The limitation and future improvement are also discussed.

This study has identified probable factors that govern ISMR predictability. Furthermore, extensive analysis has been performed to evaluate factors leading to the predictability aspect of Indian Summer Monsoon Rainfall (ISMR) using uncoupled and coupled version of National Centers for Environmental Prediction Coupled Forecast System (CFS). It has been found that the coupled version (CFS) has outperformed the uncoupled version [Global Forecast System (GFS)] of the model in terms of prediction of rainfall over Indian land points. Even the spatial distribution of rainfall is much better represented in the CFS as compared to that of GFS. Even though these model skills are inadequate for the reliable forecasting of monsoon, it imparts the capacious knowledge about the model fidelity. The mean monsoon features and its evolution in terms of rainfall and large-scale circulation along with the zonal and meridional shear of winds, which govern the strength of the monsoon, are relatively closer to the observation in the CFS as compared to the GFS. Furthermore, sea surface temperature–rainfall relation is fairly realistic and intense in the coupled version of the model (CFS). It is found that the CFS is able to capture El Nino Southern Oscillation ISMR (ENSO-ISMR) teleconnections much strongly as compared to GFS; however, in the case of Indian Ocean Dipole ISMR teleconnections, GFS has the larger say. Coupled models have to be fine-tuned for the prediction of the transition of El Nino as well as the strength of the mature phase has to be improved. Thus, to sum up, CFS tends to have better predictive skill on account of following three factors: (a) better ability to replicate mean features, (b) comparatively better representation of air–sea interactions, and (c) much better portrayal of ENSO-ISMR teleconnections. This study clearly brings out that coupled model is the only way forward for improving the ISMR prediction skill. However, coupled model’s spurious representation of SST variability and mean model bias are detrimental in seasonal prediction.

The Northwest Pacific (NWP) circulation (subtropical high) is an important component of the East Asian summer monsoon system. During summer (June–August), anomalous lower tropospheric anticyclonic (cyclonic) circulation appears over NWP in some years, which is an indicative of stronger (weaker) than normal subtropical high. The anomalous NWP cyclonic (anticyclonic) circulation years are associated with negative (positive) precipitation anomalies over most of Indian summer monsoon rainfall (ISMR) region. This indicates concurrent relationship between NWP circulation and convection over the ISMR region. Dry wind advection from subtropical land regions and moisture divergence over the southern peninsular India during the NWP cyclonic circulation years are mainly responsible for the negative rainfall anomalies over the ISMR region. In contrast, during anticyclonic years, warm north Indian Ocean and moisture divergence over the head Bay of Bengal-Gangetic Plain region support moisture instability and convergence in the southern flank of ridge region, which favors positive rainfall over most of the ISMR region. The interaction between NWP circulation (anticyclonic or cyclonic) and ISMR and their predictability during these anomalous years are examined in the present study. Seven coupled ocean–atmosphere general circulation models from the Asia-Pacific Economic Cooperation Climate Center and their multimodel ensemble mean skills in predicting the seasonal rainfall and circulation anomalies over the ISMR region and NWP for the period 1982–2004 are assessed. Analysis reveals that three (two) out of seven models are unable to predict negative (positive) precipitation anomalies over the Indian subcontinent during the NWP cyclonic (anticyclonic) circulation years at 1-month lead (model is initialized on 1 May). The limited westward extension of the NWP circulation and misrepresentation of SST anomalies over the north Indian Ocean are found to be the main reasons for the poor skill (of some models) in rainfall prediction over the Indian subcontinent. This study demonstrates the importance of the NWP circulation variability in predicting summer monsoon precipitation over South Asia. Considering the predictability of the NWP circulation, the current study provides an insight into the predictability of ISMR. Long lead prediction of the ISMR associated with anomalous NWP circulation is also discussed.

Skilful prediction of the seasonal Indian summer monsoon (ISM) rainfall (ISMR) at least one season in advance has great socio‐economic value. The ISM is a lifeline for about a sixth of the world's population. The ISMR prediction remained a challenging problem with the subcritical skills of the dynamical models due to a limited understanding of the interaction among clouds, convection and circulation. In this study, we have analysed the seasonal mean of high cloud fraction, ice mixing ratio and ice cloud fraction from satellite and reanalysis and demonstrated their importance for ISM. The variability of the mixing ratio of cloud ice in different time scales (3–7 days, 10–20 days and 30–60 days bands) is also examined from reanalysis during ISM. Here, we have shown the teleconnection of different cloud variables over the ISM region with global sea surface temperature. We found that they are tied with slowly varying forcing (e.g., El Niño and Southern Oscillation). Besides, the correlation of cloud ice with different indices (Niño, Pacific Decadal Oscillation, North Atlantic Oscillation and Extratropics) may enhance the potential predictability of ISMR. The representation of deep convective clouds, which involve the ice‐phase processes in a coupled climate model, strongly modulates ISMR variability in association with global predictors. The results from the two sensitivity simulations using coupled global climate model (CGCM) demonstrate the importance of the cloud ice on ISM rainfall predictability. Therefore, this study provides a scientific basis for improving the simulation of the seasonal ISMR by developing the physical processes of the cloud on a subseasonal time scale and motivating further research in this direction.

The dominant interannual SST variability in the eastern equatorial Atlantic is referred to as the Atlantic Zonal Mode (AZM), which peaks in boreal summer impacts global weather patterns. The cold (warm) phase of this ocean-atmospheric coupled phenomenon enhances (weakens) the intensity of the Indian Summer Monsoon Rainfall (ISMR). Observational studies show a strengthening relationship between AZM and ISMR in recent decades, providing a predictive signal for the ISMR. However, a suite of Coupled Model Intercomparison Project Phase 6 (CMIP6) model simulations in the highest emission scenario (SSP58.5) show a weakening relationship between ISMR and AZM in the future (2050–2099). The strengthening of atmospheric thermal stability over the tropical Atlantic in the warming scenario weakens the associated convection over the eastern equatorial Atlantic in response to the warm phase of AZM. This leads to weakening velocity potential response over the Indian subcontinent, resulting in a weak AZM–ISMR relationship. There is no convincing evidence to indicate that either the tropical Atlantic SST bias or the AZM–ISMR teleconnection bias plays a crucial role in the potential weakening of this relationship. These results imply that ISMR prediction will become more challenging in a warming scenario as one of the major external boundary forces that influence monsoon weakens.

This thesis focuses on improving seasonal rainfall forecasts through post-processing techniques, with a particular emphasis on Java, Indonesia. The primary objective of this research is to develop and evaluate bias correction methods for seasonal precipitation forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) Seasonal Forecasting System, Version 5 (SEAS5). By improving forecast skills for critical agricultural months, this study aims to provide insights and tools that support better decision-making and planning. Chapter one provides the background and introduction, focusing on the importance of improving precipitation model forecasts with post-processing techniques and the significance of seasonal forecasting. This chapter lays the groundwork for the rest of the thesis. Chapter two attempts to correct the biases in seasonal precipitation forecasts from the ECMWF’s SEAS5 system for Java, Indonesia, using empirical quantile mapping (EQM). The study demonstrates that bias correction enhances forecast accuracy, particularly during critical agricultural months (July-September), and could support agricultural planning. Chapter three continues with the post-processing of seasonal forecasts, comparing a more advanced statistical method with the traditional EQM approach. It also investigates the impact of climate factors such as El Niño-Southern Oscillation (ENSO), Indian Dipole Mode (IOD), Madden-Julian Oscillation (MJO), regional Sea Surface Temperature (SST), and geographical features on forecast accuracy, evaluating forecasts from 1981 to 2010, focusing on July to October. Chapter four emphasizes the importance of seasonal forecasts for hydrological models, particularly in predicting streamflow. It evaluates the calibration of streamflow forecasts with lead times up to four months, using EQM-corrected rainfall data as the primary input. Various metrics, including Continuous Ranked Probability Score Skill Score (CRPSS), Brier Skill Score (BSS), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Relative Operating Characteristic Score (ROCS), are used for verification. This chapter marks a pioneering effort in integrating hydrological models with seasonal rainfall forecasts in Indonesia. Chapter five serves as a comprehensive overview of the primary findings and discussions, exploring how the EQM bias correction method can improve the seasonal rainfall forecasts of the ECMWF model for Java and the potential forecast skill improvements when incorporating multiple predictors in the statistical postprocessing of SEAS5 rainfall forecasts. This chapter also evaluates the significance of these bias correction methods on seasonal rainfall and streamflow forecasts. Additionally, it outlines future research directions to enhance seasonal forecasting in Indonesia.

A detailed analysis of sensitivity to the initial condition for the simulation of the Indian summer monsoon using retrospective forecast by the latest version of the Climate Forecast System version-2 (CFSv2) is carried out. This study primarily focuses on the tropical region of Indian and Pacific Ocean basin, with special emphasis on the Indian land region. The simulated seasonal mean and the inter-annual standard deviations of rainfall, upper and lower level atmospheric circulations and Sea Surface Temperature (SST) tend to be more skillful as the lead forecast time decreases (5 month lead to 0 month lead time i.e. L5–L0). In general spatial correlation (bias) increases (decreases) as forecast lead time decreases. This is further substantiated by their averaged value over the selected study regions over the Indian and Pacific Ocean basins. The tendency of increase (decrease) of model bias with increasing (decreasing) forecast lead time also indicates the dynamical drift of the model. Large scale lower level circulation (850 hPa) shows enhancement of anomalous westerlies (easterlies) over the tropical region of the Indian Ocean (Western Pacific Ocean), which indicates the enhancement of model error with the decrease in lead time. At the upper level circulation (200 hPa) biases in both tropical easterly jet and subtropical westerlies jet tend to decrease as the lead time decreases. Despite enhancement of the prediction skill, mean SST bias seems to be insensitive to the initialization. All these biases are significant and together they make CFSv2 vulnerable to seasonal uncertainties in all the lead times. Overall the zeroth lead (L0) seems to have the best skill, however, in case of Indian summer monsoon rainfall (ISMR), the 3 month lead forecast time (L3) has the maximum ISMR prediction skill. This is valid using different independent datasets, wherein these maximum skill scores are 0.64, 0.42 and 0.57 with respect to the Global Precipitation Climatology Project, CPC Merged Analysis of Precipitation and the India Meteorological Department precipitation dataset respectively for L3. Despite significant El-Nino Southern Oscillation (ENSO) spring predictability barrier at L3, the ISMR skill score is highest at L3. Further, large scale zonal wind shear (Webster–Yang index) and SST over Nino3.4 region is best at L1 and L0. This implies that predictability aspect of ISMR is controlled by factors other than ENSO and Indian Ocean Dipole. Also, the model error (forecast error) outruns the error acquired by the inadequacies in the initial conditions (predictability error). Thus model deficiency is having more serious consequences as compared to the initial condition error for the seasonal forecast. All the model parameters show the increase in the predictability error as the lead decreases over the equatorial eastern Pacific basin and peaks at L2, then it further decreases. The dynamical consistency of both the forecast and the predictability error among all the variables indicates that these biases are purely systematic in nature and improvement of the physical processes in the CFSv2 may enhance the overall predictability.

Indian summer monsoon rainfall prediction: A comparison of iterative and non-iterative approaches

ABSTRACTEarlier studies have identified a teleconnection between the Atlantic zonal mode (AZM) and Indian summer monsoon rainfall (ISMR), both of which are active during the boreal summer (AZM: June–August; ISMR: June–September). It is known that El Niño‐Southern Oscillation (ENSO)‐like coupled dynamics are operational in the tropical Atlantic during the AZM events. Our goal here is to extend this process understanding to seek a predictive relation between the tropical Atlantic and the ISMR based on these known teleconnections. Monthly composite analysis of the zonal surface winds, heat content, and sea surface temperature (SST) in the equatorial Atlantic tells us that signatures of a warm or cold AZM event begin to emerge as early as January of that year. We found significant correlations between the ISMR and the low level zonal winds in the western equatorial Atlantic and heat content in the eastern equatorial Atlantic in the boreal spring season. Tracking coherent changes in these winds and the evolution of the heat content in the deep tropical Atlantic in the boreal spring may offer the potential for skillful predictions of the ensuing summer monsoon anomalies, especially during non‐ENSO years when the predictability of ISMR tends to be low.

Chapter 2 - Interannual variation of the Indian summer monsoon, ENSO, IOD, and EQUINOO

We use seasonal forecasts from the Climate Historical Forecast Project (CHFP) to study the skill of multiple climate models in predicting Indian summer monsoon precipitation. The multi-model average of seasonal forecasts from eight prediction systems shows statistically significant skill for predicting Indian monsoon precipitation at seasonal lead times. Rapid convergence of tropical rainfall skill with ensemble size suggests that the skill of seasonal monsoon rainfall forecasts improves only marginally when using multi-model ensemble (MME) means as compared to the single most skillful system. There is also a large range in the skill of individual models. Some individual models show correlation skill as high as 0.6, which is similar to the MME mean, while others show low skill. We also investigate the effect of spatial averaging on the skill of predicting monsoon rainfall and show that the predictions averaged over a larger area than the verifying observations can yield higher skill due to the extended spatial coherence of monsoon rainfall variability. We also document current errors in seasonal prediction systems and show that these are more strongly related to the errors in El-Nino Southern Oscillation (ENSO) teleconnections than they are to mean rainfall biases. Finally, we examine the ENSO-monsoon relationship and confirm that this relationship is likely to be stationary, despite fluctuations in the observed relationship, which can simply be explained as sampling variability on an underlying stationary teleconnection between ENSO and the Indian monsoon.

The potential predictability of the Indian summer monsoon rainfall (ISMR), soil moisture, and sea surface temperature (SST) is explored in the latest version of the NCEP Climate Forecast System (CFSv2) retrospective forecast at five different lead times. The focus of this study is to find out the sensitivity of the potential predictability of the ISMR to the initial condition through analysis of variance technique (ANOVA), information‐based measure, including relative entropy (RE), mutual information (MI), and classical perfect model correlation. In general, the all methods show an increase in potential predictability with a decrease in lead time. Predictability is large over the Pacific Ocean basin as compared to that of the Indian Ocean basin. However, over the Indian land region the potential predictability increases from lead‐4 to lead‐2 and then decreases at lead‐1 followed by again increase at lead‐0. While the actual ISMR prediction skill is highest at lead‐3 forecast (second highest at lead‐1), the potential predictability is highest at lead‐2. It is found that highest and second highest actual prediction skill of the ISMR in CFSv2 is due to the combined effects of initial Eurasian snow and SST over Indian, west Pacific and eastern equatorial Pacific Ocean region. While the teleconnection between the ISMR and El Niño‐Southern Oscillation is too strong, the ISMR and Indian Ocean dipole have completely out of phase relation in the model as compared to the observation. Furthermore, the actual prediction skill of the ISMR is now very close to the potential predictability limit. Therefore, in order to improve the ISMR prediction skill further, development of model physics as well as improvements in the initial conditions is required.

The prediction of Indian summer monsoon rainfall (ISMR) on a seasonal time scales has been attempted by various research groups using different techniques including artificial neural networks. The prediction of ISMR on monthly and seasonal time scales is not only scientifically challenging but is also important for planning and devising agricultural strategies. This article describes the artificial neural network (ANN) technique with error- back-propagation algorithm to provide prediction (hindcast) of ISMR on monthly and seasonal time scales. The ANN technique is applied to the five time series of June, July, August, September monthly means and seasonal mean (June + July + August + September) rainfall from 1871 to 1994 based on Parthasarathy data set. The previous five years values from all the five time-series were used to train the ANN to predict for the next year. The details of the models used are discussed. Various statistics are calculated to examine the performance of the models and it is found that the models could be used as a forecasting tool on seasonal and monthly time scales. It is observed by various researchers that with the passage of time the relationships between various predictors and Indian monsoon are changing, leading to changes in monsoon predictability. This issue is discussed and it is found that the monsoon system inherently has a decadal scale variation in predictability.