Precipitation estimation using remote sensing and numerical simulation through neural networks

Abstract

Precipitation is one of the most critical components of the hydrological balance, and its accurate quantification is essential for reliable modeling. Due to the limited availability of in-situ observational data, remote sensing is used to enhance the spatial and temporal coverage of atmospheric conditions. Additionally, statistical and machine learning models are employed to link remotely sensed variables with precipitation. In this work, a deep neural network model was developed for quantitative precipitation estimation. The network is fed with infrared brightness temperature data from the GOES-16 satellite, and numerical forecasts of precipitable water from the GFS were also considered as input. The proposed model is trained against instantaneous precipitation rates estimated from the GPM satellite’s onboard radar. Results using only brightness temperature as input were satisfactory; however, the inclusion of the numerical forecasts improved performance on the test dataset and allowed for a better representation of maximum precipitation.