Explainable AI-Driven Domain Adaptation for PM2.5 Estimation from Aerosol Optical Depth

Satellite-based estimation of fine particulate matter (PM2.5) increasingly relies on machine learning (ML), yet the transferability of these models across space and time remains poorly understood. This limitation is critical for remote sensing applications where models trained in one domain are routinely applied elsewhere without ground validation. In this study, we ask: Which PM2.5 predictors exhibit consistent influence on model performance across domains, and how can this knowledge be used to improve model generalisability? We develop Random Forest models for PM2.5 estimation using satellite aerosol optical depth (AOD) and meteorological data for two African cities, Kampala and Nairobi, representing contrasting pollution profiles. Using five years (2020–2024) of data, we systematically evaluate model transferability across cities and seasons. We first quantify domain shift using two-sample Kolmogorov–Smirnov tests (KS tests) and t-distributed stochastic neighbour embedding (t-SNE), and use SHapley Additive exPlanations (SHAP) to identify predictors with consistent influence across domains. We evaluate three transfer strategies: naive transfer, supervised feature augmentation (FA), and our proposed SHAP-guided supervised FA, which combines SHAP importance, domain-weighted training, and supervised FA. Results from KS-tests and t-SNE visualisations show that large distributional differences exist across spatial domains, with clearer separation between cities than between seasons. SHAP analyses revealed that while AOD remains a dominant predictor, the relative influence of meteorological variables varies across domains, reflecting their sensitivity to local atmospheric dynamics. Naive cross-city transfer results in performance degradation, with NRMSE increasing from 9.9 to 33% when transferring the model from Kampala to Nairobi and from 10 to 16.8% when transferring the model from Nairobi to Kampala. In contrast, SHAP-guided FA reduces NRMSE to 7.6% and 6.9%, respectively. For cross-season transfer, SHAP-guided FA similarly improves NRMSE from 11.3% to 6.9% for wet-to-dry season transfer and from 10.5% to 6.9% for dry-to-wet season transfer. These results demonstrate that xAI can serve not only for interpretability but also as a practical tool to support domain adaptation in satellite-based air quality modelling. Although evaluated using Random Forest models in two African cities, the framework is model-agnostic and provides a basis for future evaluation across additional geographic domains and ML architectures. Another limitation of the study is the relatively small dataset used (1617 and 853 samples for Kampala and Nairobi, respectively); however, this reflects the realities of data availability in many low-resource settings.