Summary and conclusions
The Landsat series of satellites provides Earth observations dating back to the 1970s with a high spatial resolution of 30 m every 16 days, making them highly valuable for a variety of applications, such as land use/cover classification and change detection, resource assessment, and environmental monitoring. However, a significant challenge faced is the interference caused by atmospheric aerosols, largely affecting the accuracy of extraction and retrieval of surface parameters. To tackle this issue, this study first identifies the critical input features for aerosol retrieval based on the fundamental principles of an atmospheric radiative transfer (ART) model. Subsequently, we introduce the Transformers model, for the first time, to address numerous intractable nonlinear problems inherent in the complex processes of decoupling the Earth-atmosphere system. Additionally, multidimensional spatial, temporal (time series), and altitude information are incorporated into the Transformers model (GeoChronoTransformers, or GCT) to improve the model performance. Finally, all Landsat data preprocessing, spatiotemporal matching, and aerosol retrieval tasks are efficiently executed using the Google Earth Engine (GEE) cloud platform. In this effort, the study gathered all available images (~20,755 cloud-screened scenes) from Landsat 8 and 9, spanning from their launch to 2022, matching approximately 470 ground monitoring stations. They then were used to develop a robust model capable of conducting global aerosol retrieval tasks for Landsat imagery in an automatic and operational manner.
We utilized the XAI (Explainable Artificial Intelligence)-SHAP (SHapley Additive exPlanations) method to elucidate the internal mechanisms of our developed ART-GCT-GEE model, revealing the significance of multi-band spectral channels, observation geometry, and multi-dimensional spatiotemporal information in the aerosol retrieval, contributing 58%, 19%, and 19%, respectively. The 10-fold cross-validations demonstrated a high level of agreement between our retrievals and observations, with sample-based, station-based, and month-based correlation coefficients (R) [root-mean-square error, RMSE] of 0.902 (0.086), 0.826 (0.111), and 0.871 (0.097), respectively. Furthermore, independent spatiotemporal validations demonstrated the effectiveness of our model in predicting aerosol optical depth (AOD) levels in regions (with-continent hold) and periods (with-year hold) where observations are unavailable, as well as historical and future AOD levels (e.g., R = 0.863, RMSE = 0.096) over land. Moreover, the model exhibits a remarkable level of robustness, with minimal sensitivity to changes in surface conditions like surface reflectance, elevation, and land-use cover. Importantly, it outperforms most widely used traditional machine- and deep-learning models. Aerosol retrieval experiments conducted in four representative regions worldwide have confirmed the model’s capacity to accurately capture variations in AOD concentrations over surfaces ranging from dark to bright. The model excels in providing high-resolution and highly detailed spatial distributions of AOD, particularly in urban areas characterized by high levels of anthropogenic emissions. It also performs effectively in scenarios involving elevated AOD pollution levels, like in situations where smoke and dust are present. In summary, our developed ART-GCT-GEE model demonstrates that it can provide precise aerosol retrievals from Landsat imagery over land, offering valuable insights into environmental and air quality assessments. Particularly noteworthy is that our model can be used with data from instruments on other satellites (e.g., MODIS and VIIRS) as long as their data becomes available on the GEE platform in the future.
Data availability
The ART-GCT-GEE aerosol retrieval cloud model for Landsat imagery is free to all users and available at https://weijing-rs.github.io/product.html. This online framework will be made publicly available once the paper is accepted. The authors greatly thank Kaitao Li and Li Li from the Chinese Academy of Sciences for their help in data collection and processing.