Assessing the Evolution of Rainfall Trends in India through Historical Analysis, Forecasting, and Comparison with the CMIP-6 SSP585 Scenario

Technical Summary: Assessing the Evolution of Rainfall Trends in India through Historical Analysis, Forecasting, and Comparison with the CMIP-6 SSP585 Scenario

Problem Statement
Climate change is driving widespread variability in extreme weather events, posing significant challenges to water resources, agriculture, and economic stability, particularly in subtropical regions like India. Rainfall, a fundamental driver of the hydrological cycle, exhibits increasing irregularities that lead to droughts, floods, and crop failures. While previous studies have analyzed rainfall trends at global or national levels, there is a need for comprehensive, spatio-temporally extensive analyses that integrate historical trend detection with advanced machine learning forecasting and comparison against high-emission climate scenarios. Furthermore, traditional statistical methods often struggle with non-normally distributed data, autocorrelation, and the identification of structural breaks (changepoints) in long-term time series, potentially leading to misleading conclusions.

Methodology
The study conducts a rigorous analysis of rainfall data spanning 122 years (1901–2022) across India, resolved into six Homogeneous Regions (HRs) and thirty-four Meteorological Sub-divisions (MSDs). The methodology integrates four distinct analytical components:

1. Trend Detection: The study employs non-parametric statistical techniques, specifically the Mann-Kendall (MK) Test and Sen's Slope estimator, to identify monotonic trends in annual, seasonal, and monthly rainfall. To address limitations regarding autocorrelation and data homogeneity, the authors integrated homogeneity tests (Pettitt, Buishand Range, and Standard Normal Homogeneity Test) to detect changepoints (break years) before applying trend analysis.
2. Innovative Trend Analysis (ITA): To validate and refine results, the study utilizes the ITA method (Şen, 2012). This graphical non-parametric technique handles autocorrelated data and outliers without requiring strict assumptions about sample size or linearity, allowing for the detection of sub-trends and distinct trend combinations within time series.
3. Mechanistic Analysis: To understand the physical drivers of observed variations, the study analyzes ERA5 reanalysis data (1940–2022). Key variables include convective precipitation, low-cloud cover, and vertically integrated moisture divergence. A changepoint was identified in 1956, and differences in circulation patterns between the periods 1940–1956 and 1957–2022 were assessed.
4. Forecasting and Scenario Comparison: An Artificial Neural Network-Multilayer Perceptron (ANN-MLP) model was trained on historical data to forecast annual rainfall up to the year 2050. Model performance was optimized using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). These machine learning forecasts were then juxtaposed with projections from the Coupled Model Intercomparison Project Phase 6 (CMIP-6) under the Shared Socioeconomic Pathway 5-8.5 (SSP585) scenario, representing a worst-case "business-as-usual" climate trajectory. The comparison utilized a multi-model ensemble (MME) of three CMIP-6 Global Climate Models (CanESM5, BCC-CSM2-MR, and IITM-ESM).

Key Results

• Historical Trends: The MK Test and ITA revealed a significant negative trend in annual rainfall for the Northeast HR and seven specific MSDs (including Arunachal Pradesh, Assam Meghalaya, and NMMT). Conversely, nine MSDs, including Coastal Andhra Pradesh and Punjab, exhibited significant positive trends. Seasonal analysis indicated negative trends in the Northeast HR during the monsoon season, while Coastal Andhra Pradesh and Punjab showed increasing monsoon rainfall.
• Changepoints: Homogeneity tests identified that most regions experienced a structural break in rainfall patterns post-1950s. Post-changepoint analysis confirmed that the Northeast HR and several eastern subdivisions have experienced significant declines in monsoon rainfall over the last 70 years.
• Physical Drivers: ERA5 analysis attributed rainfall variations to fluctuations in convective precipitation, increased low-cloud cover, and inadequate vertical moisture divergence. Specifically, intensified northeasterly winds have reduced cooler air intrusion, decreasing winter rainfall, while increased moisture divergence in western and central regions has contributed to drier conditions.
• Forecasting (2050): The ANN-MLP model predicts an overall decline in annual rainfall across India by 2050. While the Northeast HR is expected to retain the highest average rainfall (approx. 1920 mm), the Northwest HR is projected to face the most significant reductions.
• Comparison with CMIP-6 SSP585: A critical finding is that the ANN-MLP forecasts for all six HRs and 28 of the 34 MSDs predict rainfall levels lower than the CMIP-6 MME projections under the SSP585 scenario. This suggests that actual future rainfall may fall even below the worst-case climate model projections, indicating a severe trajectory for water scarcity. Only seven MSDs, primarily coastal and some interior regions, are projected to exceed the SSP585 rainfall levels.

Significance and Claims
The authors claim that this study provides a critical, multi-faceted assessment of India's rainfall evolution by bridging historical trend analysis with future forecasting and climate model validation. The integration of changepoint detection with MK and ITA methods offers a more robust understanding of non-monotonic trends than traditional approaches alone.

The study asserts that its comparison of machine learning forecasts against CMIP-6 SSP585 projections offers a unique framework for assessing regional climate impacts, revealing that India may face rainfall deficits exceeding even the most pessimistic climate model scenarios. These findings are presented as vital for policymakers, hydrologists, and agricultural planners to formulate adaptive strategies for irrigation, water resource management, and climate resilience. The authors acknowledge limitations, noting that the study's broad spatial scale (40 regions) precludes micro-level analysis (district/block/village) and that the lack of daily data restricts the granularity of current trend detection. They suggest that future work should employ advanced deep learning (e.g., LSTM) and grid-based remote sensing to refine these insights.