Figure 6. Uncertainty analysis with box plots of bias and the fraction falling within the EE envelope (curves) of global AOD from Landsat imagery retrievals against ground-based measurements as a function of (a) surface reflectance, (b) elevation (m), (c) NDVI, and (d) land-use type. Black horizontal solid lines represent zero biases. In each box, the red dots, middle, lower, and upper horizontal lines represent the AOD bias mean, median, and 25th and 75th percentiles, respectively.
Aerosol retrieval over typical regions
Given the extensive volume of Landsat imagery on a global scale, this study selected four distinct regions: Denver, USA; Madrid, Spain; Beijing, China; and the Sahara Desert to perform aerosol retrieval experiments (Figure S6). These regions encompass diverse land-use types and serve as representatives of varying surface conditions, climates, and levels of human activity. All available Landsat 8/9 OLI images from these regions from 2013 to 2022 were collected to conduct aerosol retrievals using the trained ART-GCT-GEE model on the GEE cloud platform. Figure 7 presents true-color images and corresponding AOD retrievals (550 nm) in Denver, USA, and Madrid, Spain on different dates, both illustrative examples of dark surfaces. These regions share similar land surface characteristics, primarily comprising plains and mountains, dense vegetation (forests, grasslands, croplands), low population densities, and consistently low aerosol levels throughout the year (Figure S6a-b). Our model exhibits high spatial continuity and provides detailed pollution information at a high spatial resolution (30 m), capturing subtle spatial variations within the overall low aerosol background throughout the months. Importantly, the model is capable of capturing exceptionally high smoke AOD values generated by sudden wildfires on specific days (e.g., 15 September 2018 and 4 September 2020) during fire seasons in Denver, USA, even across the entire image at relatively low pollution levels (pointed by yellow arrows in Figure 7b-c). Furthermore, besides very clean conditions, our model also works well in retrieving the AOD spatial distribution under infrequent highly polluted conditions (Figure 7f). We also conducted a validation of aerosol retrievals using data from 12 available AERONET stations within the two study areas (Table S7). A high level of agreement between observations and retrievals is revealed, with estimated biases less than ± 1% and close RMSEs equal to 0.06. Approximately 91% and 85% of the retrievals met the EE criteria, with proportions falling within the GCOS range of 67% and 55%, respectively.