Wildfires are increasingly recognised as major contributors to atmospheric pollution, yet their spatial–temporal influence on regional air quality remains insufficiently understood. This study quantifies the impact of wildfire activity on atmospheric gas concentrations by integrating exposure modelling, correlation analysis, and advanced machine learning techniques. Two datasets were employed: (i) continuous atmospheric observations of CO, CO\(_2\), CH\(_4\), and BC alongside meteorological parameters; and (ii) a wildfire dataset containing fire locations, burned area, and duration. A novel Fire Exposure Index (FEI) was developed to quantify the dynamic likelihood that wildfire plumes influence the observatory, incorporating fire distance, burned area, and wind characteristics. Correlation analyses across distance bands and daily lags (up to six days) revealed clear distance and wind-dependent relationships, with short-term lag effects primarily within one to two days. Baseline machine learning models (Gradient Boosting, Random Forest, Decision Tree) achieved moderate accuracy, while recurrent neural networks (LSTM, GRU, BiLSTM) captured stronger temporal dependencies, particularly for CO\(_2\) and CO. A stacked ensemble architecture was subsequently implemented, combining LightGBM, LSTM, GRU, and BiLSTM with a LightGBM meta-learner. The hybrid framework achieved substantial performance improvements within the study domain (\(\textrm{R}^2\) values ranging from 0.959 to 0.9897 across CO, CH\(_4\), and CO\(_2\)), demonstrating strong predictive capability under the specific meteorological and geographic conditions examined. The proposed approach demonstrates that integrating exposure metrics with hybrid ensemble learning provides a promising and interpretable strategy for predicting wildfire-induced atmospheric variability and supporting air quality management in fire-prone regions.
