Estimation of high spatial resolution ground-level ozone concentrations based on Landsat 8 TIR bands with deep forest model.
Researchers
Journal
Modalities
Models
Abstract
In recent years, China has been facing severe ozone (O3) pollution, which poses a remarkable threat to human health. Most estimation methods only provide ozone products at a relatively coarse resolution, such as 5 km, but high-resolution ozone data are essential for ozone pollution prevention and control. To this end, we proposed a new framework for estimating ozone concentrations at 300 m resolution in China based on Landsat 8 infrared (IR) bands and meteorological data using a deep forest (DF) model. DF combines the excellent performance of tree integration with the expressive power of hierarchical distributed representations of neural networks. The accuracy and mapping results of DF are considerably better than some widely used machine learning methods (generalized regression neural regression network and random forest). The sample-based cross-validation (CV), station-based CV, time-based CV, and extrapolation validation show that the estimations of DF are in high agreement with the station observations with determination coefficient values of 0.938, 0.926, 0.687, and 0.660, respectively. The proposed method was used to analyze the spatial and temporal ozone variations at fine scales in three typical Chinese cities (Beijing, Wuhan and Guangzhou), where the mean ozone concentrations during the polluted season are consistent with the land use and urban heat island distribution. The rationality of ozone estimates was verified, and the advantages of high-resolution mapping was demonstrated by comparing the monitoring data from municipal controlling stations in Beijing, 10 km ozone products, and satellite images. Our product can represent spatial details and locate local pollution sources, such as temples. The proposed method has important implications for the fine-scale monitoring of ozone pollution.Copyright © 2022 Elsevier Ltd. All rights reserved.