Figure 8. Same as Figure 7 but for two typical bright-surface regions: (a-c) Beijing, China and (d-f) the Sahara Desert in northern Africa.
Comparison with other AI models
Finally, we compare the performance of our ART-GCT-GEE model with ten popular AI models through spatiotemporally independent validation, i.e., training these models on data from 2015–2020 and validating them on data from the remaining years (Table 3). All machine-learning models show good performance in predicting global AOD levels, with strong correlations with surface observations (R = 0.72–0.84), MAE and RMSE values generally below 0.08 and 0.014, and a relatively high proportion of retrievals meet the EE (64–77%) and GCOS (27–44%) criteria. Moreover, boosting-based models (i.e., XGBoost, CatBoost, and LightDBM) outperform bagging-based ensemble-learning methods (Extra Trees and Random Forest) in aerosol retrievals due to their continuous optimization functions, which correct the residual more effectively. However, DL, with its enhanced data-mining capabilities and improved ability to tackle nonlinear problems, exhibits improved predictive accuracy (e.g., R = 0.82–0.85,\(f_{\text{EE}}\) = 74–80%), especially the latest models like MLP and ResNet. The Transformers model represents a recent powerful state-of-the-art DL approach, incorporating spatial information and, notably, time series data, surpassing all other machine and DL methods. This superiority is evident in its highest correlation, near-zero estimation bias, lowest MAE and RMSE values, and the highest proportions of retrieval samples meeting the EE and GCOS criteria. The 10-CV results performed at sample, spatial, and temporal scales yield similar conclusions (Table S8), providing further evidence of the Transformers model’s effectiveness. This study presents the first attempt to apply the Transformers model to the complex problem of aerosol retrieval, effectively decoupling the Earth-atmosphere nonlinear problem.
Table 3. Comparison in performance across various machine- and deep-learning models for retrieving AOD from Landsat imagery.