PM2.5 is a major airborne pollutant affecting air quality and public health and exhibits complex spatiotemporal patterns driven by multi-factor interactions. We propose Deep Convolutional Boosted Forest (DCBF), an enhanced deep learning framework integrating multi-source data including satellite observations, meteorological records, and auxiliary data for seamless day-night hourly PM2.5 inversion. The method replaces conventional multi-grained scanning with an Efficient Channel Attention-enhanced 1D convolutional neural network (1DCNN-ECA) to improve local feature extraction, while employing CatBoost to reconstruct cascade forest structures for enhanced global feature interactions and generalization capability. Experimental results demonstrate that DCBF shows significant improvement over baseline models. When applied to analyze a heavy pollution episode in Shijiazhuang, DCBF successfully captured the spatiotemporal evolution patterns of PM2.5, validating its practical utility in environmental monitoring.