Feature selection is vital in data pre-processing in machine learning, and it is prominent in datasets with many features. Feature selection analyses the relevant, irrelevant, and redundant features in the dataset. Feature selection removes the irrelevant features, which improves both the accuracy and prediction performance. The significant advantages of reducing the number of features from the dataset are reducing the training time, reducing overfitting, decreasing the curse of dimensionality, and simplifying the prediction model. The filter feature selection techniques can handle the issues with the high number of features, and this paper uses the symmetric uncertainty coefficient to verify the relevance of the independent features. In this paper, a new feature selection method named as kurtosis-based feature selection has been proposed to select the relevant features which affect the air pollution. Kurtosis-based feature selection is compared with seven filter feature selection techniques on air pollution dataset and validated the performance of the proposed algorithm. It has been observed that the kurtosis-based feature selection extracts only PM2.5 as the key feature and has been compared to the accuracy of the five existing methods. The experimental results illustrate that the kurtosis-based feature selection algorithm reduces the original feature set up to 91.66%, but the existing filter feature selection techniques reduce the feature set to only 50%.