The coarse Moderate Resolution Imaging Spectroradiometer (MODIS) aerosol optical depth (AOD) product (spatial resolution: 3 km) retrieved by the dark-target algorithm always generates the missing values when being adopted to estimate the ground-level PM2.5 concentrations. In this study, we developed a two-stage random forest using MODIS 3-km AOD to obtain the PM2.5 concentrations with full coverage in a contiguous coastal developed region, i.e., Yangtze River delta–Fujian–Pearl River delta (YRD–FJ–PRD) region of China. A first-stage random forest–integrated six meteorological fields was employed to predict the missing values of AOD product, and the combined AOD (i.e., random forest–derived AOD and MODIS 3-km AOD) incorporated with other ancillary variables were developed for predicting PM2.5 concentrations within a second-stage random forest model. The results showed that the first-stage random forest could explain 94% of the AOD variability over YRD–FJ–PRD region, and we achieved a site-based cross validation (CV) R2 of 0.87 and a time-based CV R2 of 0.85. The full-coverage PM2.5 concentrations illustrated a spatial pattern with annual-mean PM2.5 of 46, 40, and 35 μg m−3 in YRD, PRD, and FJ, respectively, sharing the same trend with previous studies. Our results indicated that the proposed two-stage random forest model could be effectively used for PM2.5 estimation in different areas.
Satellite top-of-atmosphere (TOA) reflectance has been validated as an effective index for estimating PM2.5 concentrations due to its high spatial coverage and relatively high spatial resolution (i.e., 1 km). For this paper, we developed an emsembled random forest (RF) model incorporating satellite top-of-atmosphere (TOA) reflectance with four categories of supplemental parameters to derive the PM2.5 concentrations in the region of the Yangtze River Delta-Fujian (i.e., YRD-FJ) located in east China. The landscape pattern indices at two levels (i.e., type level and overall level) retrieved from 3-year land classification imageries (i.e., 2016, 2018, and 2020) were used to discuss the correlation between county-based PM2.5 values and landscape pattern. We achieved a cross validation R2 of 0.91 (RMSE = 9.06 μg/m3), 0.89 (RMSE = 10.19 μg/m3), and 0.90 (RMSE = 8.02 μg/m3) between the estimated and observed PM2.5 concentrations in 2016, 2018, and 2020, respectively. The PM2.5 distribution retrieved from the RF model showed a trend of a year-on-year decrease with the pattern of “Jiangsu > Shanghai > Zhejiang > Fujian” in the YRD-FJ region. Our results also revealed that the landscape pattern of farmland, water bodies, and construction land exhibited a highly positive relationship with the county-based average PM2.5 values, as the r coefficients reached 0.74 while the forest land was negatively correlated with the county-based PM2.5 (r = 0.84). There was also a significant correlation between the county-based PM2.5 and shrubs (r = 0.53), grass land (r = 0.76), and bare land (r = 0.60) in the YRD-FJ region, respectively. Three landscape pattern indices at an overall level were positively correlated with county-based PM2.5 concentrations (r = 0.80), indicating that the large landscape fragmentation, edge density, and landscape diversity would raise the PM2.5 pollution in the study region.
The segmentation of buildings in remote-sensing (RS) images plays an important role in monitoring landscape changes. Quantification of these changes can be used to balance economic and environmental benefits and most importantly, to support the sustainable urban development. Deep learning has been upgrading the techniques for RS image analysis. However, it requires a large-scale data set for hyper-parameter optimization. To address this issue, the concept of “one view per city” is proposed and it explores the use of one RS image for parameter settings with the purpose of handling the rest images of the same city by the trained model. The proposal of this concept comes from the observation that buildings of a same city in single-source RS images demonstrate similar intensity distributions. To verify the feasibility, a proof-of-concept study is conducted and five fully convolutional networks are evaluated on five cities in the Inria Aerial Image Labeling database. Experimental results suggest that the concept can be explored to decrease the number of images for model training and it enables us to achieve competitive performance in buildings segmentation with decreased time consumption. Based on model optimization and universal image representation, it is full of potential to improve the segmentation performance, to enhance the generalization capacity, and to extend the application of the concept in RS image analysis.