Research On Machine-learning-based Urban-scale PM_2.5 Concentration Mapping Method

Atmospheric PM_2.5 poses significant threats to human health,ecological environment,and global climate.It is important and necessary to monitor PM_2.5concentration with high precision and high resolution.However,due to the limited resolution of existing satellite aerosol products,current PM_2.5 inversion algorithms can only estimate the concentration at a kilometer scale,which is far from enough for fine PM_2.5 monitoring at the city scale.Therefore,this study focuses on the problem of urban-scale PM_2.5 concentration mapping at the spatial resolution of hundred-meter,and our goal is to achieve high-resolution,high-precision,and high-coverage PM_2.5concentration estimation for pollution monitoring inner the city.The research mainly focuses on the following four aspects:（1）Considering that there are currently some PM_2.5 products with kilometer-level resolution,we attempted to use a machine learning-based downscaling approach to improve the spatial resolution of existing products.With high-resolution terrain data as ancillary data,we developed a PM_2.5 product downscaling model based on a cascaded random forest algorithm that fused random forest model with cascade learning.We conducted experiments nationwide,and the results showed that the downscaled model can improve the spatial resolution of existing PM_2.5 products by more than 30 times,and the downscaled PM_2.5 products at the scale of hundreds of meters can accurately depict PM_2.5 pollution at the city scale.However,the quality of the downscaled product heavily relies on the original product,and the defects of the original product will be inherited by the downscaled product,resulting in limited accuracy.（2）To address this issue,we attempted to achieve sub-kilometer level PM_2.5 mapping with retrieval method.We found that dominant factors of PM_2.5’s spatial variation changed with scale.Based on this,we developed a dual-scale retrieval model that considered scale effects.The model used different variables and constructed different models at different scales to obtain high-resolution PM_2.5 products step-by-step.Nationwide experiments showed that the dual-scale retrieval model yielded higher accuracy than traditional single-scale retrieval method,producing higher-resolution PM_2.5 concentration maps.However,the proposed model still relied on satellite aerosol data as input,leading to serious data gaps in the retrieval results,particularly in urban areas,which affected the practical application value of the product.（3）In order to obtain high-resolution PM_2.5 products with higher spatial coverage,we attempted to directly estimate PM_2.5 concentrations using reflectance data.Due to the more complex relationships between PM_2.5 and reflectance data compared to the relationships between PM_2.5 and aerosol products,we employed a deep neural network model with powerful fitting ability.In addition,we conducted in-depth discussions on key issues in the inversion process,such as band selection,cloud influence,scale effect,model selection,and model evaluation.And a multidimensional evaluation system and a mapping quality evaluation method based on semivariance variance function and frequency distribution chart was proposed.Experimental results showed that this method can obtain high spatial coverage and high-precision PM_2.5 products with a resolution of 100 meters.The product can accurately reflect the changes in PM_2.5 pollution in urban train stations,factories,parks,residential areas,hospitals,and other locations during the COVID-19 lockdown period,revealing the close relationships between human activities and PM_2.5 pollution.（4）We found that the approach of using reflectance data for inversion is also applicable to other air pollutants,and different pollutants are closely related.With this knowledge,we proposed to use multi-task learning techniques to build a unified retrieval model that can synchronously and jointly estimate surface concentrations of six criteria atmospheric pollutants.We constructed a multi-task deep neural network.We added a physical-informed concatenate layer into the model considering the interrelationships between air pollutants,and designed a physical-guided loss function to improve the model accuracy.Experimental results show that our proposed physically-informed multi-task deep neural network model can accurately estimate six atmospheric pollutants simultaneously,with accuracy comparable to that of single-task learning models.Meanwhile,the model efficiency has been improved by several times.Overall,from the downscaling model based on cascade random forest,to the dual-scale inversion model,to the reflectance-based inversion model,we have gradually achieved high-resolution,high-precision,and high-coverage PM_2.5 concentration mapping,making it possible to finely monitor PM_2.5 at the urban scale.The joint inversion model based on multi-task learning,as an extension of the reflectance-based inversion model,further alerted the traditional modeling approach that targeted at single pollutant and improved the modeling efficiency of atmospheric pollutant remote sensing inversion.