Beyond the Black Box: Interpretable Machine Learning for Airplane Crash Severity Classification

Airplane crash severity classification is crucial for enhancing aviation safety and risk assessment. This study proposes a method to classify airplane crashes as Severe or Non-Severe based on the number of fatalities, using a dataset containing various crash details, including Date, Location, Operator, Fatalities, and more. A new feature, Crash_Severity, is introduced to label crashes with more than 15 fatalities as Severe, and others as Non-Severe. Four Classification models were employed to classify crash severity. The models achieved accuracies of 97.4%, 93.8%, 88.75%, and 58.23%, respectively. The results indicate that Logistic Regression and Random Forest performed very well, possibly due to the clear distinction between the two classes based on fatalities. In contrast, SVC struggled, possibly due to the linear separability limitations in the dataset. These findings demonstrate that fatalities serve as a strong predictor for crash severity, with Random Forest and Logistic Regression being the most effective models for this task. Future work may explore incorporating additional features to further enhance model performance