Detecção de turbulência de céu claro por abordagem híbrida com aprendizado de máquina e modelagem atmosférica

Abstract

This thesis develops a hybrid model for severe clear-air turbulence (CAT) detection, applied to a high air-traffic region in southeastern Brazil, bounded by longitudes 43°W to 49°W and latitudes 19°S to 25°S, covering the period from 2018 to 2021. The methodology integrates atmospheric modeling using the GFS and WRF models with machine learning techniques. Predictor attribute selection was performed through hypothesis testing with false discovery rate (FDR) control, implemented in the BRB-Array Tools software, yielding 13 attributes for the GFS and 11 for the WRF. The selected attributes were used to train and evaluate 14 machine learning algorithms, including classical classifiers and MLP neural networks, through stratified crossvalidation. Using GFS data, the MLP with one layer and 10 neurons achieved the highest single-fold AUC (0.95) and the greatest stability among neural network architectures evaluated (mean AUC = 0.931). Using WRF data, the SVM obtained the best mean performance (mean AUC = 0.969). Metric stability analysis conducted via bootstrap demonstrated consistent superiority of the WRF over the GFS for the statistically most robust models, confirming that the higher spatial resolution of the WRF confers a real gain in discriminant capacity. The results demonstrate that the proposed hybrid approach is efficient and robust for severe CAT detection, with potential for operational application in aeronautical meteorological monitoring systems