Interpretation of black box for short-term predictions of pre-monsoon cumulonimbus cloud events over Kolkata

Abstract Thunderstorms are meso-scale systems that are characterised by deep convective cumulonimbus (Cb) clouds associated with torrential rain, lightning, hail, dust storms, strong winds, downbursts, and tornadoes. As Gangetic West Bengal is prone to thunderstorm, early forecasting is imperative i...
Ausführliche Beschreibung

Abstract Thunderstorms are meso-scale systems that are characterised by deep convective cumulonimbus (Cb) clouds associated with torrential rain, lightning, hail, dust storms, strong winds, downbursts, and tornadoes. As Gangetic West Bengal is prone to thunderstorm, early forecasting is imperative in order to protect life and property, and prevent the damage caused by these intense storms. The present study comprises two issues on model evaluation and interpretability by applying popular machine learning algorithms, viz., Extreme Gradient Boosting (XGBoost) and Logistic Regression (LR) with potentially predictive thermodynamic indices and parameters for short-term predictions of pre-monsoon thunderstorms over Kolkata. The thermodynamic indices and parameters employed include convective available potential energy (CAPE), convective inhibition (CIN), Bulk Richardson number (BRN), K-Index (KI), lifted index (LI), total totals Index (TT), Showalter index (SI),temperature (TEMP), relative humidity (RELH), dew point temperature (DWPT), wind direction (DRCT), wind Speed (SKNT), mixing ratio (MIXR), severe weather threat index (SWI), potential temperature (THTA), equivalent potential temperature (THTE), and virtual potential temperature (THTV). In the proposed approach, we place a greater emphasis on the concept of Explainable artificial intelligence (XAI) to apply SHapley Additive exPlanations (SHAP), a Shapley-value-based explanation method based on the coalitional game theory. The SHAP approach, as a primary interface, enables the identification and prioritization of features that determine the occurrences of pre-monsoon thunderstorms and compares the two different machine learning algorithms. SHAP can quantify the contribution of predictor variables to each data point and rank the importance of predictor variables in terms of their contributions to the model output. It also facilitates the computation of different plots on both global and local levels. Accordingly, it can help determine the validity of the model based on domain expertise by identifying the most important variables. The results indicate that both XGBoost and LR support the dominant positive influence of the convective available potential energy (CAPE), while the ranks and interpretations of the other predictor variables differ. Although, these two models perform well in predicting the pre-monsoon thunderstorms, they may favour different predictor variables due to their varying natures, thereby resulting in different explainability. Ausführliche Beschreibung

Autor*in:	Dutta, Debashree [verfasserIn] Pal, Sankar K.

Format:	E-Artikel
Sprache:	Englisch

Erschienen:	2022

Schlagwörter:	Thunderstorm SHapley Additive exPlanations (SHAP) XGBoost Logistic Regression CAPE

Anmerkung:	© The Author(s), under exclusive licence to Springer Nature Switzerland AG 2022

Übergeordnetes Werk:	Enthalten in: Journal of data, information and management - [Cham] : Springer Nature Switzerland AG, 2019, 4(2022), 2 vom: Juni, Seite 167-183
Übergeordnetes Werk:	volume:4 ; year:2022 ; number:2 ; month:06 ; pages:167-183

Links:	Volltext

DOI / URN:	10.1007/s42488-022-00071-9

Katalog-ID:	SPR047639903