Espresso: Deep Learning Model for Global Precipitation Estimation from Satellite Data

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Sep 28 2023

In a paper published in the journal Meteorology, researchers introduced Espresso, a deep-learning model for global precipitation estimation from satellite data. Espresso, developed using geostationary satellite input and calibrated with Global Precipitation Measurement Core Observatory (GPMCO) data, offered highly accurate results. It outperformed other products in storm localization and intensity estimation, making it an operational tool at Meteo-France for real-time global precipitation estimates every 30 minutes.

Background

Precipitation is essential for Earth's climate and human activities, and accurate estimation of precipitation is crucial. Traditional methods like ground observations and radar have limitations. Satellite-based remote sensing offers global coverage, but it has challenges. Various satellite-based precipitation estimation products exist, combining different data types. While recent computing and deep-learning advances offer new possibilities for improving satellite data accuracy, deep convolutional neural networks (CNNs) have shown promise in enhancing precipitation estimation through satellite imagery.

In previous research, deep-learning techniques, specifically CNN, have been employed to estimate precipitation using satellite data. These studies have shown promise, training models on infrared (IR) data and ground radar measurements, sometimes incorporating IR and passive microwave (PMW) data, often with a regional focus like China's southeast coast. However, the untapped potential lies in extending deep learning to global-scale satellite-based precipitation estimation.

Proposed Methodology

This approach used a deep neural network called Deep Lab version 3 Plus (DeepLabv3+), widely recognized for its efficacy in image segmentation tasks. Researchers preferred DeepLabv3+ over the commonly used U-shaped network (U-Net) because it incorporated the Spatial Pyramid Pooling module and Atrous convolutions. These architectural elements enable DeepLabv3+ to capture multi-scale information, produce denser feature maps, and delineate sharper object boundaries.

Furthermore, it boasts a larger receptive field than U-Net, facilitating the analysis of a broader feature space. Notably, DeepLabv3+ possesses fewer weight parameters, rendering it less susceptible to overfitting. The implementation of DeepLabv3+ from the Keras Code Examples was utilized, with the final activation layer modified to a Rectified Linear Unit (ReLU) function for regression. Adjusted Dilution rates inside the Spatial Pyramid Pooling module reduced the network's receptive field to match the image size. The resultant model was composed of 15 million trainable parameters.

An additional weight, primarily determined by the target precipitation value, was incorporated into the modified Mean Squared Error (MSE) loss function. This weight exhibited an exponential relationship with the target precipitation, prioritizing higher weights for pixels associated with greater target precipitation values. In this context, the modification aimed to mitigate the standard MSE's tendency to diminish the impact of extreme importance, which is of critical significance.

The loss function's weight also factored in the latitude of each pixel, assigning more significance to samples from northern and southern latitudes where heavy rainfall is less common. Additionally, it considered the challenges posed by lower tropopause near the poles, affecting precipitation detection due to IR brightness temperature attenuation. This weighting mechanism addressed the imbalance from the dataset's relatively limited samples from these polar regions, incorporating a threshold.

Experimental Analysis

Performance evaluation of precipitation estimation models employed a set of metrics. The primary metric was the RMSE, apt for a regressor model. However, to gain deeper insights into the model's capabilities, rain estimates were categorized into four classes: no rain, light rain, medium rain, and heavy rain. These categories enabled the computation of the Probability of Detection (POD) and False Alarm Rate (FAR) for three rainfall thresholds, computed pixel-wise from the confusion matrix. The Fraction Skill Score (FSS) addressed the Spatial uncertainty to evaluate the model's accuracy in locating rain areas. Gridded neighborhoods were employed to compare forecasted and observed rain frequencies, alleviating the double penalty issue in traditional contingency scores.

In the experimental protocol, the examination of the influence of input data on model performance began, with a primary focus on FSS in the validation set. This investigation included assessing the advantages of employing four Geostationary Earth Orbit (GEO) images as input, evaluating the effectiveness of different channel combinations, and appraising the benefits of additional features. Training each network for 25 epochs preserved the best checkpoint based on validation set performance. The approach utilized the Adam optimizer with specific parameters. 

The performance against GPMCO data on the test set identified the most influential input data combination. The scrutiny included an examination of the confusion matrix for rain categories and an assessment of FSS, POD, and FAR variations across latitudinal zones. Subsequently, this method generated 1-hour cumulated rainfall estimation data for four radar domains to align with the hourly cumulated radar data set. Comparative analysis was conducted against six other operational products, focusing on RMSE, FSS, POD, and FAR for performance assessment.

Finally, a meticulous examination of 15 extreme rainfall events in 2022 ensued. A double-masked evaluation by Meteo-France forecasters compared Espresso with other products, culminating in their conclusions and a detailed analysis of specific cases in Montpellier and Guadeloupe. This comprehensive evaluation offered valuable insights into subjective variances across models.

Conclusion

In summary, Espresso, a deep convolutional neural network for global precipitation estimation using satellite data, exhibits strong performance, particularly in temperate and tropical regions. While facing some limitations in higher latitudes, the model shows potential for improvement by addressing data imbalances or refining model parameters. Its integration as an operational tool at Meteo-France provides real-time global precipitation estimates, aiding forecasters in monitoring areas without ground radars and enhancing weather event response capabilities. Espresso is a valuable tool in the face of evolving weather patterns due to climate change.