Dissertation defense (November 14, 2024): Augusto José Moreira da Fonseca
Student: Augusto José Moreira da Fonseca
Title: Precipitation Interpolation using Spatiotemporal Graph Convolutional Networks
Advisors: Eduardo Bezerra da Silva (Advisor) and Fabio Andre Machado Porto (Co-advisor)
Committee: Eduardo Bezerra da Silva (Cefet/RJ – PPCIC), Fabio Andre Machado Porto (LNCC), Eduardo Soares Ogasawara (Cefet/RJ – PPCIC), Mariza Ferro (UFF), Leonardo Silva de Lima (UFPR)
Day/Time: November 14, 2024 / 10 a.m.
Abstract: The monitoring and forecasting of atmospheric weather conditions have significant implications across diverse fields, including agriculture, transportation, and public safety. Accurate weather forecasting can help mitigate and prevent the adverse impacts of severe weather events, including the loss of human lives. With technological advancements, several instruments are available for atmospheric observation, such as satellites, radars, and in-situ stations. These instruments provide real-time observations and, in some cases, high spatial and temporal resolution, making them ideal input for training Machine Learning (ML)-based weather prediction models. However, one challenge regarding meteorological data is the availability and spatial distribution of in-situ stations. Due to their uneven distribution, certain areas end up with insufficient monitoring. Interpolation methods are commonly used to infer precipitation values for areas not covered by in-situ stations. However, these methods often overlook atmospheric context and physics, leading to discrepancies between interpolated and observed data. Additionally, these methods do not simultaneously account for both spatial and temporal dimensions. This study aims to implement a Spatio-Temporal Graph Convolutional Network (STGCN) to interpolate precipitation data, a significantly imbalanced variable. The STGCN is adapted to perform precipitation interpolation by learning atmospheric patterns and physics in multivariate time series from multiple meteorological instruments. The objective is to improve interpolation accuracy, especially in areas with limited in-situ station coverage. We compare the results achieved by the STGCN with those of a traditional interpolation method, Inverse Distance Weighting (IDW). The results indicate that, compared to IDW, the STGCN achieves higher accuracy in extreme precipitation events and promising results in events of lower magnitude. In some cases, our model achieved accuracy improvements for extreme events ranging from 40% to 80% over the IDW method. However, we identified that the imbalance in precipitation data and the relatively low number of extreme event examples could hinder STGCN accuracy in some scenarios. Our method proves promising and paves the way for future research.
