Skill Evaluation of Extended Range Forecasts for Northeast Monsoon Rainfall across Smaller Spatial Domains of South Peninsular India

Technical Summary: Skill Evaluation of Extended Range Forecasts for Northeast Monsoon Rainfall across Smaller Spatial Domains of South Peninsular India

Problem Statement
Reliable prediction of Northeast Monsoon (NEM) rainfall at extended range lead times is critical for agricultural planning, water resource management, and disaster risk reduction in South Peninsular India. Unlike the Southwest Monsoon, NEM rainfall is spatially heterogeneous and heavily concentrated in specific sub-divisions such as Tamil Nadu, Kerala, and coastal Andhra Pradesh. While previous research has advanced the understanding of NEM climatology and seasonal prediction, systematic diagnostic evaluations of Extended Range Forecasts (ERF) at finer spatial scales remain limited. Existing operational forecasts often rely on broader meteorological subdivisions, which may obscure local variability essential for region-specific decision-making. This study addresses the gap by evaluating the performance of operational ERFs over 11 smaller, more granular spatial domains within South Peninsular India across contrasting monsoon seasons.

Methodology
The study evaluates the operational Extended Range Forecast (ERF) system of the India Meteorological Department (IMD), which utilizes the CFSv2 coupled model (based on NCEP's CFSv2, with MOM4 ocean and GFS atmospheric components). The system generates weekly forecasts up to four weeks in advance, initialized every Wednesday.

• Study Area and Period: The analysis covers 11 specific regions in South Peninsular India (e.g., North/South Kerala, various Tamil Nadu sub-divisions, Coastal/Rayalaseema Andhra Pradesh, and Interior Karnataka) during the NEM seasons of 2015 (substantially above normal rainfall), 2016 (severely deficient rainfall), and 2017 (below normal rainfall).
• Data Sources: Forecasts are verified against IMD-GPM merged rainfall data at 1° resolution, combining rain-gauge and satellite estimates.
• Categorization: To account for high variability at finer scales, rainfall is categorized into three groups based on weekly departures from the Long Period Average (LPA): Above Normal (AN, >20%), Normal (NN, -19% to +19%), and Below Normal (BN, <-20%).
• Verification Metrics: Performance is assessed using categorical verification scores derived from 2x2 contingency tables. The metrics include Forecast Accuracy, Bias Score (BS), Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI), and Equitable Threat Score (ETS).

Key Results
The evaluation reveals a clear dependence of forecast skill on lead time, spatial location, and interannual monsoon characteristics:

• Lead Time Dependence: Forecast skill is highest at Week-1 and degrades systematically through Week-4. Week-1 forecasts demonstrate accuracy exceeding 65–75%, with the Normal (NN) category performing best (70–80% accuracy). By Week-4, skill declines to approximately 40–50%.
• Category Performance:
  • Normal (NN): Consistently shows the highest accuracy and POD across all lead times and years, with Bias Scores close to unity (minimal systematic error).
  • Above Normal (AN): Shows moderate accuracy but a tendency toward slight overestimation (Bias > 1), particularly at longer leads.
  • Below Normal (BN): Exhibits the lowest skill, particularly at longer leads, with a tendency toward underestimation (Bias < 1) and higher False Alarm Ratios.
• Interannual Variability:
  • 2016 (Severely Deficient): Demonstrated the highest predictability and spatial consistency among the three years, with useful skill extending to Week-3.
  • 2017 (Below Normal): Showed intermediate performance, better than 2015 but with more variability than 2016.
  • 2015 (Substantially Above Normal): Exhibited the lowest skill and highest uncertainty. This reduced predictability is attributed to dominant synoptic-scale variability and strong interactions between synoptic disturbances and low-frequency modes, which are difficult for models to capture at extended leads.
• Spatial Variability: Coastal regions generally displayed better predictability compared to interior regions, particularly at longer lead times.

Significance and Claims
The paper concludes that the operational ERF system possesses meaningful predictive skill for NEM rainfall over smaller spatial domains in South Peninsular India, particularly up to a 2–3 week lead time. The study highlights that while forecast uncertainty increases with lead time, the system remains capable of distinguishing between rainfall regimes (AN, NN, BN) with reasonable reliability, especially for normal conditions.

The authors assert that these findings support the potential application of ERFs in operational agro-advisory services and impact-based forecasting. By providing reliable weekly outlooks for specific sub-divisions, the system can assist stakeholders in optimizing agricultural activities, such as irrigation scheduling and sowing decisions, and in preparing for flood or drought risks. The study emphasizes that while the system is not perfect, its current skill level offers a valuable tool for decision-making in the region, provided users account for the degradation of skill beyond the second week and the specific challenges posed by highly active synoptic seasons like 2015.