Fire safety is a major global concern in subsurface environments such as underground mines. Over decades, such hazards due to fire and explosion have resulted in the loss of lives and entities. These fires not only disrupt mining operations but also endanger workers’ safety due to compromised air quality. In addition to that, these fires are also responsible for the depletion of coal, ground subsidence, and the release of greenhouse gases (GHGs) and toxic substances. To determine the fire and its intensity in the sealed-off mines several ratios are used in the literature. But all the gas ratios deal with different gas content and have their advantage and disadvantages. The current study utilizes fire data to investigate key aspects of the mine fire safety. Advanced ensemble algorithms random forest (RF), extreme gradient boosting (XGBoost), and long short-term memory (LSTM) are used for this purpose. A comprehensive dataset comprising 271 samples from sealed-off underground coal mines has been collected from the literature. This data contains measurements of the percentage of Carbon-mono-oxide (CO), Carbon-di-oxide (CO2), Oxygen (O2), Methane (CH4), Hydrogen (H2), and Nitrogen (N2) as input parameters to predict the fire class of that particular sampled region. Model hyperparameters are efficiently optimized for better performance and several performance matrices were utilized to access the model performance. The proposed intelligent decision-making framework enhances the prevention of fire risk and develops the early warning system for underground mines. This integrated framework ensures the safety of underground environments, and reduces the emission of GHGs, hence supporting sustainable development. This research work further promotes sustainability by minimizing the impact on the environment and humans.