At present, in the simulation of various atmospheric pollutants, the simulation of independent trace gases is constrained by the insufficient resolution of key remote sensing products, resulting in insufficient simulation reliability. This study combines spatial sampling and parameter convolution to optimize LightGBM using ground observations, remote sensing products, meteorological data, aid data, and random IDs. Through the above techniques and atmospheric pollutant sequence simulation, we obtained seamless products with a daily resolution of 1 kilometer for SO2in most parts of China from 2018 to 2020. Through random sampling, random site sampling, specific area validation, comparison of different models, and horizontal comparison of different studies, we have verified that our simulation of the spatial distribution of various atmospheric pollutants is reliable and effective.
The data used in this study includes daily ground monitoring data of Chinese SO2. In addition, remote sensing data, meteorological data, and auxiliary data were also used.
A multi pollutant universal machine learning model based on random IDs, spatial adoption, parameter convolution, and other methods can better consider multiple factors when predicting changes in atmospheric pollutant concentrations and optimize the estimation of pollutant spatial distribution. We use CV and visual qualitative analysis to evaluate the model results. Compare LightGBM, LSTM, and RF Ps with our model to evaluate their performance. Finally, we use SHAP to attempt to explain the output results of the model.
The data quality is good.
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
CSTR:11738.11.NCDC.ZENODO.DB4168.2024
10.5281/zenodo.7580714