This research presents a novel hybrid sampling technique, implemented at the data level, to effectively address imbalanced and noisy data in classification processes. The proposed technique expertly combines two established methods, namely, the random over sampling (ROS) and neighbourhood cleaning rule (NCL) approaches, to tackle imbalance and noise issues, respectively. The study carried out an empirical evaluation of the proposed approach using crowdsourced text data that primarily emphasized the triple bottom line (TBL) dimension of a smart social, economic, and environmental city. The study used the long short-term memory (LSTM), convolutional neural networks (CNN), and CNN-LSTM classification models to validate the efficacy of the proposed hybrid sampling technique and compare its performance with other existing approaches, including ROS oversampling, NCL undersampling, synthetic minority over sampling & tomek links (SMOTE-Tomek), and synthetic minority oversampling and edited nearest neighbours (SMOTE-ENN) hybrid sampling. The results are impressive, with the ROS-NCL hybrid sampling technique achieving high accuracy rates across all three classification models, at 97.71%, 98.01%, and 98.11%, respectively. This approach provides a robust and effective solution for handling impure data and holds great promise in identifying complex data patterns in real-world classification problems.