Impact of features in rainfall prediction with Explainable Artificial Intelligence - SHAP

Predicting when rain will fall is essential for managing water resources and mitigating hydrological disasters. This work was performed to identify the impact of the feature elements using an explainable deep learning strategy that causes heavy rainfall. First, we trained the rainfall prediction model with an artificial neural network (ANN). In order to assess the relevance of the input meteorological parameters to the target parameter, we computed the attribution values for the parameters using SHapley Additive exPlanations (SHAP) framework. The hourly rainfall data for the eight cities in India with the highest population was used to train the ANN model. The key conclusion from the experiment is that the most influential factors in rainfall prediction are those features whose values show a strong positive or negative correlation with their corresponding attribution (SHAP) values. According to the results, the most important positive features are FeelsLikeC, Humidity, CloudCover, Pressure, and moon illumination, and the most important negative features are WindSpeedKmph, TempC, Visibility, HeatIndexC, and WindChillC, which will play an important role in rainfall prediction. The importance of this study involves improving interpretability by identifying important features.

کلیدواژه ها:

Rainfall ، Explainable Artificial Intelligence ، Artificial Neural Network ، Remote Sensing ، SHAP