Classification of Low-Level Wind Shear on Imbalanced Radiosonde Data Using a Hybrid CNN-LSTM Architecture

Abstract

Low-Level Wind Shear (LLWS) during the takeoff and landing phases poses a severe and dynamic threat to aviation safety. However, it remains inherently challenging to detect directly from surface observations due to its highly non-linear, transient, and sequential nature in the lower atmosphere. This study proposes an advanced deep learning approach to classify LLWS events based on high-resolution vertical radiosonde profile data spanning the 2024–2025 period at Soekarno-Hatta International Airport, Indonesia. Addressing the critical challenge of a highly imbalanced dataset—where hazardous meteorological events are exceptionally rare compared to normal atmospheric conditions—the classification target was simplified into an event-level binary scheme. This scheme utilizes a hazard intensity threshold of ≥ 4 knots/100 ft, adhering strictly to the International Civil Aviation Organization (ICAO) moderate-to-severe hazard categories. The comprehensive research methodology involved fixed-grid interpolation into 21 uniform vertical observation points (spanning 0 to 2000 ft), rigorous wind vector transformation into zonal and meridional components, and circular encoding for wind direction to maintain absolute numerical stability during model training. A hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) architecture was developed to capture both local spatial gradients and long-term vertical dependencies, which was then systematically compared against a Transformer baseline equipped with a global self-attention mechanism. The empirical results and operational discussion demonstrate that the proposed CNN-LSTM model significantly outperformed the Transformer baseline, achieving an outstanding Area Under the Precision-Recall Curve (PR-AUC) of 0.9672 compared to the Transformer's 0.7533. Furthermore, the confusion matrix analysis conclusively proved the superior capability of the CNN-LSTM network in minimizing false negatives (missed detections) from 17 down to merely 7 critical events. In conclusion, the integration of local feature extraction through CNN and sequential dependency modeling via LSTM is highly effective for addressing imbalanced meteorological datasets, providing a highly robust and physically consistent computational foundation for automated aviation safety early warning systems.