Rainfall regression from C-band Synthetic Aperture Radar using Multi-Task Generative Adversarial Networks

TL;DR

A data-driven approach to estimate precipitation rates from Synthetic Aperture Radar at a spatial resolution of 200 meters per pixel is introduced, introducing patch-level components and an adversarial component.

Abstract

This paper introduces a data-driven approach to estimate precipitation rates from Synthetic Aperture Radar (SAR) at a spatial resolution of 200 meters per pixel. It addresses previous challenges related to the collocation of SAR and weather radar data, specifically the misalignment in collocations and the scarcity of rainfall examples under strong wind. To tackle these challenges, the paper proposes a multi-objective formulation, introducing patch-level components and an adversarial component. It exploits the full NEXRAD archive to look for potential co-locations with Sentinel-1 data. With additional enhancements to the training procedure and the incorporation of additional inputs, the resulting model demonstrates improved accuracy in rainfall estimates and the ability to extend its performance to scenarios up to 15 m/s.

Figures (13)

• Figure 1: Wind speed distribution of the Sentinel-1/NEXRAD collocations.
• Figure 2: Geographic distribution of the patches from the Sentinel-1/NEXRAD collocations, after the reduction operation.
• Figure 3: Architecture of the multi-objective model. Inputs encompass a 2D image $x_{im}$ and scalar $x_{sc}$ data. Outputs consist of binary segmentation $y_{seg}$, regression $y_{rr}$, and discriminator output $y_D$. Convolutional layers utilize ReLU activation functions by default. ConvBlock and DeconvBlock are presented in expanded form on the side, where the parameters $a$ and $b$ correspond to the number of convolution kernels in each convolution layer.
• Figure 4: (a) Distribution of the maximum NEXRAD DPR rainfall on the patches of the test set (x-axis) against the maximum predicted rainfall (y-axis). The colored area accounts for 90% of the rain patches, while the remaining 10% are indicated as points. (b) Pixel-wise distribution of the precipitation rate from the NEXRAD dataset and the model prediction between 0 and 100 mm/h (top) and zoomed in on 0 to 30 mm/h (bottom).
• Figure 5: Co-observation from both Sentinel-1 (acquired on 2018-02-05 at 04:47:13 UTC) and NEXRAD (acquired at 04:48 UTC).