A PM_2.5 Concentration Inversion Algorithm Combining Feature Selection and Optimized Deep Forest

Accurate and spatiotemporally continuous inversion of $\mathbf{P M}_{\mathbf{2} .5}$ concentration is crucial for air pollution control, given $\mathbf{P M}_{2.5}$’s significant impact as a major atmospheric pollutant. To address the challenges posed by numerous influencing factors with complex interrelationships, this study proposes a novel machine learning framework integrating advanced feature selection and ensemble learning techniques. First, we implement a hybrid Recursive Feature Elimination with Cross-Validation (RFECV)-LightGBM algorithm for intelligent feature selection, which effectively reduces feature dimensionality while improving model accuracy in both daytime and nighttime inversion scenarios. Subsequently, we innovatively enhance the cascade forest structure using Tree-structured Parzen Estimator (TPE) optimized CatBoost, developing an improved deep forest model that demonstrates superior performance compared to single models, with significantly enhanced prediction accuracy and overfitting resistance. The proposed model demonstrates superior performance compared to other machine learning models across key metrics, including higher $\mathbf{R}^{\mathbf{2}}$, lower RMSE, and MAE. SHapley Additive explanations (SHAP) value analysis reveals that the inflence of variables on the accuracy of inversion model is quite different. Applied to a severe pollution episode in Shijiazhuang, our model successfully captures the spatiotemporal dynamics of $\mathbf{P M}_{2.5}$ distribution, providing a reliable, high-precision monitoring solution powered by cutting-edge machine learning technologies.