Kumar et al. (2026) Estimation of location-specific precipitation using Deep Neural Networks

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-03-18
• Authors: Bipin Kumar, Bhvisy Kumar Yadav, Soumyodeep Mukhopadhyay, Rakshit Rohan, Bhupendra Bahadur Singh, Rajib Chattopadhyay, Nagraju Chilukoti, Atul Kumar Sahai
• DOI: 10.1007/s00704-026-06185-z

Research Groups

• Indian Institute of Tropical Meteorology, Ministry of Earth Sciences, Government of India, Pune, India
• Earth and Atmospheric Sciences, National Institute of Technology Rourkela, Rourkela, Odisha, India
• Indian Institute of Science Education and Research, Pune, India

Short Summary

This study introduces two Deep Neural Network (DNN) architectures for location-specific precipitation estimation, demonstrating their superior accuracy and computational efficiency compared to traditional Kriging methods across various meteorological conditions in India. The DNN models, especially one incorporating additional meteorological variables, consistently outperform Kriging in capturing spatial precipitation patterns and extreme events.

Objective

• To develop and evaluate Deep Neural Network (DNN) models for accurate and computationally efficient estimation of location-specific precipitation from coarse-resolution grid data, demonstrating their superiority over traditional Kriging methods.

Study Configuration

• Spatial Scale: Location-specific (hyperlocal) precipitation estimation across the Indian landmass, utilizing data from over 6,000 IMD rain gauge stations and 0.25° x 0.25° ERA5 reanalysis grid points. Models were trained using information from the 10 nearest grid points to each target location.
• Temporal Scale: Daily temporal resolution. Data spanned 40 years (1980-2019), with 35 years used for training and a 5-year validation set (1995, 1996, 2001, 2013, 2014) selected for diverse meteorological conditions, including monsoon season (June-September) and whole-year analyses.

Methodology and Data

• Models used:
  • Deep Neural Networks (DNNs): Two architectures were developed.
    • NN_1: Utilized precipitation values, elevation, and location coordinates (latitude, longitude) from 10 nearest grid points and the target station.
    • NN2: Expanded on NN1 by incorporating five additional meteorological parameters: surface pressure, temperature, relative humidity, zonal wind speed, and meridional wind speed from 10 nearest grid points and the target station.
  • Ordinary Kriging Method (KM): Used as a traditional benchmark for spatial interpolation.
• Data sources:
  • IMD Rain Gauge Data: Daily rainfall, station elevation, latitude, and longitude from over 6,000 stations across India (1980-2019).
  • ERA5 Reanalysis Data: Daily surface pressure, temperature, relative humidity, zonal wind speed, and meridional wind speed at 0.25° x 0.25° spatial resolution over the Indian region (1980-2019).

Main Results

• DNN models consistently outperformed Ordinary Kriging across all evaluation metrics (correlation coefficient (CC), root mean square error (RMSE), bias, and skill score (SS)).
• NNModel2, which incorporated additional meteorological variables, demonstrated superior performance compared to both NNModel1 and Kriging. Its maximum CC value reached 0.9 (compared to 0.81 for NN1), and its maximum RMSE reduced to 53 mm (from 90 mm for NN1). The skill score peak increased from 0.4 (NN1) to 0.6 (NN2).
• Kriging exhibited significantly lower CC, higher RMSE, and negative biases, particularly in coastal areas and central India, indicating its struggle to capture spatial precipitation variations.
• NNModel2 showed the least bias and highest skill score across most stations, with 5342 stations showing better performance than Kriging (compared to 463 stations where Kriging was better, based on Mean Absolute Error).
• SHAP analysis revealed precipitation (30.7%), latitude (23.2%), and longitude (15.9%) as the most influential predictors, with other meteorological variables providing secondary refinement.
• The models demonstrated robust performance in detecting extreme precipitation events across different regions of India, maintaining a Critical Success Index (CSI) greater than 0.53 even for events exceeding 120 mm. Quantile loss analysis confirmed better performance at higher rainfall quantiles.
• DNN models exhibited remarkable stability and consistency in performance across different years with varying meteorological conditions, unlike Kriging which showed significant year-to-year variability.
• Computational efficiency: NN_2 inference for 1,000 prediction locations required approximately 0.09–0.12 seconds, significantly faster than Kriging's 0.20 seconds.

Contributions

• Introduces and validates Deep Neural Networks (DNNs) as a superior operational approach for location-specific precipitation estimation, offering a robust and precise alternative to traditional interpolation methods like Kriging.
• Proposes two innovative DNN architectures, demonstrating the significant value of incorporating auxiliary meteorological variables (beyond just precipitation, elevation, and location) for enhanced accuracy in hyperlocal precipitation estimation.
• Provides a comprehensive evaluation of DNNs for spatial prediction, highlighting their robustness, computational efficiency, and scalability for high-dimensional meteorological data, particularly for the Indian subcontinent.
• Offers a foundational step towards making hyperlocal weather information accessible, empowering communities and policymakers to proactively manage weather-related risks and make informed decisions.

Funding

No funding was received for conducting this study.

Citation

@article{Kumar2026Estimation,
  author = {Kumar, Bipin and Yadav, Bhvisy Kumar and Mukhopadhyay, Soumyodeep and Rohan, Rakshit and Singh, Bhupendra Bahadur and Chattopadhyay, Rajib and Chilukoti, Nagraju and Sahai, Atul Kumar},
  title = {Estimation of location-specific precipitation using Deep Neural Networks},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06185-z},
  url = {https://doi.org/10.1007/s00704-026-06185-z}
}