Figure 4. Density scatter plots of global AOD retrievals
derived from Landsat imagery against ground measurements over land from
2013 to 2022 using the (a) sample-based, (b) station-based, (c)
month-based 10-fold cross-validation methods, and (d) the
spatiotemporally independent validation method (i.e., using data samples
from the years 2015 to 2020 for training and the remaining years for
testing). The solid black line is the 1:1 line, and the dashed lines
outline the EE envelope.
Spatiotemporal
predictability
To further assess the model’s predictive ability in new spatiotemporal
domains, we employ a variety of spatiotemporally independent validation
methods. Initially, by isolating the spatial dependence through the
independent control of each continent (withhold), our model demonstrates
proficiency in predicting AOD values for various regions, where most of
them yielded moderate correlations exceeding 0.5 and maintained low
median biases within ± 2% (Table S5). Furthermore, approximately more
than half (49%) to 92% of the spatial predictions aligned with the EEs
for each region. Similarly, by independently controlling each year
(withhold) to reduce temporal correlations, our model effectively
captured AOD loads for the remaining years, achieving high correlations
and minimal uncertainties (e.g., R = 0.80–0.92, MAE = 0.04–0.06, and
RMSE = 0.07–0.11) compared with ground measurements (Table S6).
Moreover, a significant proportion of the temporal predictions, at least
79% and 41%, met the EE and GCOS requirements, respectively.
We also constructed the model using data from intermediate years
(2015–2020) and subsequently performed AOD retrievals and validation
for the remaining years (2013, 2014, 2021, and 2022). Across the globe,
AOD predictions exhibit notable accuracy, with moderate correlations
exceeding 0.5 at more than 74% of the sites. Moreover, approximately
76% and 78% of the sites displayed low MAE and RMSE values less than
0.08 and 0.12, respectively (Figure 5).
However, in regions with elevated
AOD levels, such as Africa and Eastern China, sites were more prone to
underestimation possibly due to the strong absorbing aerosols, showing
higher RMSE and MAE values. Nonetheless, a substantial portion (61%) of
the sites demonstrated nearly ”unbiased” estimates (within ± 2%).
In addition, more than 76% and
60% of the sites had a significant proportion of the retrievals falling
within the EE (> 70%) and GCOS (> 40%)
envelopes. Regionally, the
predictive performance was generally considerable, with relatively small
biases (Table 2). The highest
correlations were observed in Southern Africa and Eastern Asia (R =
0.895 and 0.882), while Europe and Oceania exhibited relatively lower
correlations, attributable to historically clear air with low AOD loads.
However, they had higher fractions of retrievals falling within the EE
(> 80%) and GCOS (> 46%) envelopes.
Globally, the predictive accuracy
of our model is substantial, with the correlation between retrievals and
observations reaching 0.826 and the median bias approaching zero (Figure
4d). The average MAE and RMSE
values are 0.057 and 0.096, respectively. Overall, approximately 80.73%
of the collocated points fall within the EE envelope, and approximately
half (49.64%) of them meet the GCOS requirements.
These findings highlight the
strong adaptability and stability of the ART-GCT-GEE model, which can
effectively predict both historical and future AOD levels worldwide.