A Deep Learning Framework for Precipitation Estimation from GOES-16 multispectral satellite imagery - Application of the conditional Generative Adversarial Networks (cGANs)

Improving the accuracy of precipitation estimation from remotely sensed data is a more achievable goal with recent advancements in satellite remote sensing technologies and data-driven methods. In this study, a state-of-the-art precipitation estimation framework from remotely sensed satellite information using advances in Deep Learning (DL) is presented. Several Spectral bands of geostationary satellite data (GOES-16) with high temporal, spatial, and spectral resolutions along with the elevation information as an ancillary data are utilized to first provide a Rain/No Rain (R/NR) binary mask by classification of the pixels and then to apply a regression to estimate the amount of rainfall for rainy pixels. A fully convolutional network (U-Net) is used as a regressor to predict precipitation estimates. The network is trained using both the conditional Generative Adversarial Network (cGAN), and the mean squared error (MSE) loss terms to generate results that better learn the complex distribution of precipitation from the ground truth data. Most common performance measures are used to evaluate the accuracy of both R/NR classification and real-valued precipitation estimates. Statistics and visualizations of the metrics represent an improvement of the precipitation retrieval accuracy for the proposed algorithm compared to the PERSIANN-CCS which is an operational product, and a baseline model trained using the conventional Mean Squared Error (MSE) loss term.