Dissertation Defense: “A Weather Radar Nowcasting Tool"

Date

Horário de início

14:00

Local

Auditório 1 - Kenkichi Fujimori (P 217) - Bloco Principal - IAG/USP

Dissertation defense: 
Student: Andrea Salomé Viteri López
Program: Meteorologia
Title: “A Weather Radar Nowcasting Tool”
Advisor: Prof. Dr. Carlos Augusto Morales Rodriguez  - IAG/USP

 

Judging Comitee:

  1. Prof. Dr. Carlos Augusto Morales Rodriguez – Presidente e Orientador - IAG/USP
  2. Prof. Dr. Alexandre Gonçalves Evsukoff – UFRJ (videoconferência)
  3. Dr. Thiago Souza Biscaro – INPE (videoconferência)
  4. Prof. Dr. Leonardo Calvetti – UFPEL (videoconferência) 
  5. Dr. Alan James Peixoto Calheiros - INPE (videoconferência)

 

Abstract:

 

This study proposes a nowcasting tool based on artificial intelligence that combines the TITAN (Thunderstorm Identification, Tracking, Analysis, and Nowcasting) algorithm with recurrent neural networks of the LSTM (Long Short-Term Memory) type. The main goal is to predict the temporal and spatial evolution of the rain area and the precipitation intensity using rainfall fields estimated by a polarimetric weather radar. In this model, TITAN-LSTM, the TITAN algorithm tracks the precipitating systems observed by the radar, feeds the LSTM algorithm, and predicts the position of the storm centroids up to 30 minutes ahead. The LSTM performs the prediction of the rain area, shape, and intensity. The network architecture consists of three LSTM layers with a dropout of 0.2 and a linear layer, using the ADAM optimizer with a learning rate of 0.0001. In this work, data from the DAEE's S-band dual-polarization Doppler radar (SPOL), operated by FCTH, with a temporal resolution of 5 minutes, in the state of São Paulo, between 2016 and 2019, were used. A total of 439 storms were selected based on trajectory, duration, location, and proximity criteria. For the construction of the algorithm, 307 (70%) storms were used, while 132 (30%) were used for validation. The root mean square error (RMSE) and the mean squared error (MSE) of the predicted rain area as a function of storm duration decreases by 6% (36%) for storms with a duration between 20-40 minutes when between 10 and 15 (30-35) minutes of input are used. The mean bias error (MBE) indicates that the model overestimates (underestimates) the rain area for storms with durations between 20-40 (> 60) minutes. The algorithm presented a maximum probability of detection (POD) of 89% in the first 15 minutes of the forecast, while the false alarm rate (FAR) was less than 20%. For the remaining forecast times, the POD varied between 50-70%, but the FAR increased to 30-60% depending on the input time interval. The best performance for rainfall intensity showed a RMSE of 3.5 mm/h to 4.5 mm/h when 20 minutes of input were used. Overall, the model proved to be a good tool for predicting both rain area and precipitation, with good spatial and temporal performance. The main advantages of the methodology adopted include the use of rainfall rate fields estimated by weather radar, the prediction of the evolution of storms with their physical properties and the low computational cost of processing.


Keywords: TITAN, LSTM, weather radar