Towards Location-Specific Precipitation Projections Using Deep Neural Networks

TL;DR

The paper addresses the challenge of obtaining accurate location-specific precipitation from coarse gridded data by comparing Kriging with two deep neural network architectures. The authors assemble a large, multi-source dataset (IMD station data and ERA5 reanalysis) and evaluate two input configurations, training with 35 years and testing on five representative years, using metrics such as $CC$, $RMSE$, $bias$, and $SS$. Their results show that deep neural networks, particularly the model incorporating additional ERA5 inputs, outperform Kriging across spatial and year-wise analyses, demonstrating strong potential for localized precipitation estimation. The work signifies a step toward rapid, high-resolution, location-specific forecasts and proposes avenues for further enhancement in spatial reconstruction using differentiable programming and related techniques.

Abstract

Accurate precipitation estimates at individual locations are crucial for weather forecasting and spatial analysis. This study presents a paradigm shift by leveraging Deep Neural Networks (DNNs) to surpass traditional methods like Kriging for station-specific precipitation approximation. We propose two innovative NN architectures: one utilizing precipitation, elevation, and location, and another incorporating additional meteorological parameters like humidity, temperature, and wind speed. Trained on a vast dataset (1980-2019), these models outperform Kriging across various evaluation metrics (correlation coefficient, root mean square error, bias, and skill score) on a five-year validation set. This compelling evidence demonstrates the transformative power of deep learning for spatial prediction, offering a robust and precise alternative for station-specific precipitation estimation.

Figures (9)

• Figure 1: Depiction of the trade-off between computational expense and forecast precision, as well as minimizing superfluous data generation for providing location-specific forecast.
• Figure 2: A representation of the nearest grid points surrounding the desired location.
• Figure 3: Representation of spatial plots CC, biases, RMSE and skill score obtained from three models.
• Figure 4: Performance evaluation of three methods using density plots for 5-years test data.
• Figure 5: The station locations where the model performed exceptionally well, yielding very high correlation values. The total count of stations where the correlation coefficient (CC) value exceeded the threshold of 0.90 is provided for all three models are also indicated.