Defense in Meteorology
Student: Darsys Agüero Morell
Program: Meteorology
Title: "Precipitation estimation using the GOES-16 satellite and Artificial Neural Networks in the coastal region"
Advisor: Prof. Dr. Augusto José Pereira Filho
Judging Committee:
- Prof. Dr. Augusto José Pereira Filho – IAG/USP
- Prof. Dr. Ricardo Hallak – IAG/USP
- Prof. Dr. Nelson Jesuz Ferreira - INPE
- Prof. Dr. Otto Corrêa Rotunno Filho - UFRJ
- Prof. Dr. Hugo Abi Karam – UFRJ
Abstract:
Precipitation along the Brazilian coast is associated with recurrent hydrological impacts, yet its estimation from geostationary satellites remains challenging. One of the main limitations arises from the ambiguous relationship between cloud-top temperature and surface precipitation, particularly in shallow systems with relatively warm cloud tops. This thesis assessed in which grade incorporating large-scale meteorological variables can improve deep-learning-based precipitation estimates derived from GOES-16 observations over the coastal regions of Brazil. In the first stage, the coastal zone was objectively delineated based on the correlation between monthly IMERG precipitation time series from 2001 to 2024. The resulting “line” was then divided into four contiguous regional regimes: North, Northeast, Southeast, and South. The analysis of precipitation features observed by the GPM (Global Precipitation Measurement) Dual-frequency Precipitation Radar between 2014 and 2024, combined with brightness temperatures from GOES-16 channel 13, showed that regional differences were more closely related to the systems' frequency, vertical depth, and convective fraction than to their horizontal extent and median rainfall rate. Rainfall associated with warm cloud tops was particularly relevant in the Northeast, where it represented 18% of the precipitation features in December, January, and February and 71% in June, July, and August. In this region, rainfall rates of up to 40 mm h⁻¹ were observed at brightness temperatures between 274 and 278 K. Under such conditions, traditional radiometric criteria would tend to indicate weak rainfall or even the absence of precipitation. A four-stage selection protocol identified five ERA5 predictors for each region, with variables related to atmospheric moisture and instability prevailing among the predictors selected. In the second stage of the research, nine convolutional neural network configurations belonging to four architectural families were evaluated. The configurations included the four base models (UNet, ResAttn, ASPP, and DeepSup) and their versions incorporating meteorological variable fusion, as well as the UNet version with FiLM modulation. The networks were trained with data from 2021 to 2023, using DPR (Dual-frequency Precipitation Radar) precipitation as the reference, and evaluated in independent validation and test years, 2024 and 2025, respectively. Their performance was also compared with the operational RRQPE (Rainfall Rate / Quantitative Precipitation Estimate) product. The models outperformed RRQPE in 71 of the 72 combinations of configuration, region, and year evaluated for precipitation detection. The Critical Success Index (CSI) values ranged from 0.221 to 0.382 for the neural networks, whereas the operational product yielded values between 0.166 and 0.276. When the best-performing configuration for each region was considered, the reduction in the Mean Absolute Error (MAE) ranged from 23% to 41%. For warm rainfall, a regime for which RRQPE produced Probability of detection (POD) and CSI zero, the models achieved CSI values of up to 0.314 in the Northeast and 0.419 in the Southeast. The benefit of meteorological information varied according to the regional regime. The largest gains were observed in precipitation detection, particularly in the South, where meteorological information helped reduce false alarms, and for warm rainfall in the Northeast, where it improved the representation of a systematically underestimated regime. In contrast, its effect on rainfall-intensity estimation was small, and the inclusion of meteorological variables did not compensate for the absence of visible-channel observations at night. Among the fusion strategies evaluated, FiLM modulation produced more consistent results than direct concatenation. Concluding, the large-scale meteorological information did not uniformly improve satellite-based precipitation detection and estimation; its value was concentrated in contexts where radiometric information faces greater difficulty: overestimation in frontal regimes in the South and warm rain in the Northeast. It has also been demonstrated that regional deep learning models offer a more effective alternative to the operational RRQPE product for monitoring precipitation along the Brazilian coast. The main contribution of this thesis, therefore, lies in showing not only that the effect of the meteorological context varies according to region and rainfall type, but also in identifying where and through which mechanisms these gains occur.
Keywords: precipitation estimation, coastal precipitation, GOES-16, GPM/DPR, warm rain, deep learning, ERA5, convolutional neural networks