Retrieval of PM2.5 Using MODIS Aerosol Products Based on Depth Learning

Abstract

Most of the current PM2.5 concentration estimation models are insufficient, such as narrow research scope, poor representativeness and complex model. To improve the above deficiencies, this paper tries to explore a simpler and more efficient PM2.5 concentration estimation model by using deep neural network. In this paper, the aerosol optical thickness (Aerosol Optical Depth, AOD), which has been widely used in the field of PM2.5 concentration estimation model, and other known variables related to the distribution of PM2.5, wind speed, relative humidity, build a small neural network using dense connection network, and estimate the PM2.5 concentration in 2020. The results proved that deep learning has efficient regularity mastery in PM2.5 estimation. After about 20 minutes of iterative training in 8 months of data, the average absolute error was 7.8μg/m3 in the annual data. And with the deepening of the number of model layers and the number of iterations, the accuracy will be steadily improved. Through the estimation of the four quarters of 2020, the relatively expected results have been achieved.