Using the data-driven prediction of forest combustion PM_2.5 emissions based on machine learning and variable input

Forest combustion releases substantial amounts of PM_2.5, posing significant smoke exposure risks to firefighters operating near the firefront and adversely affecting public health across broader region. Accurate quantification of particulate matter emissions from wildfires is critical for assessing their global impact. While many current studies focus on enhancing the performance of machine learning algorithms to improve the accuracy of PM_2.5 predictions, the majority of redundant algorithms tend to compromise computational efficiency. This is disadvantageous for achieving more accurate predictions in highly dynamic forest fire scenarios, which demand both higher precision and faster model algorithms for PM_2.5 forecasting. Our research aims to enhance computational speed with reduced data input while preserving accuracy, by introducing a novelly comprehensive performance metric based on the entropy weight-TOPSIS method to evaluate both the predictive accuracy and computational speed of the models. 844 combustion experiments were employed and six machine learning algorithms were applied to analyze the impact of six variable categories-meteorology, terrain, fuel, fire behavior, flame characteristics, and fire intensity-on the prediction of PM_2.5 emissions from combustion. Results indicate that the Extra Trees (ET) algorithm achieved the highest prediction accuracy for PM_2.5 concentration (MAE =717.7 µg m⁻³) when using fire intensity, fire spread rate, meteorology, fuel, and terrain data as input variables. The Multilayer Perceptron (MLP) algorithm, when supplied with all variables, achieved the shortest computation time of 0.0025 s. Meanwhile, the MLP model incorporating fire intensity, meteorology, and fuel characteristics attained the optimal balance between prediction accuracy and computational efficiency, yielding a comprehensive performance score of 97.95. This study demonstrates that high-quality data inputs and appropriate algorithm selection are essential for achieving efficient performance, whereas variable redundancy or excessive model complexity can lead to predictive redundancy. This study provides an effective method for the rapid estimation of PM_2.5 emissions, enabling quick prediction of PM_2.5 concentrations, delineation of smoke pollution impact areas in advanced, timely issuance of health protection advisories, and facilitating efficient regional smoke management response. Furthermore, through dynamic monitoring key factors in high forest-fire-risk zones, appropriate timing for prescribed burning can be determined, enabling efficient fuel reduction. This approach mitigates the potential risks of high-intensity wildfires and associated high-concentration smoke emissions, thereby achieving reduction in particulate pollution from forest fires. These findings provide important insights for regional air quality management and forest fire risk mitigation.