High concentration of atmospheric particulate matter can cause the reduction of atmospheric visibility and has a negative impact on human health. Therefore, continuous monitoring, estimation and prediction of the temporal and spatial variation characteristics of particulate matter concentration are of great significance for environmental pollution assessment and treatment. In this work, by using the meteorological data of Rencheng District, Jining City, the air quality data from the Electrochemical Plant Station of Jining national environmental air quality monitoring point and the surface aerosol extinction coefficients inverted from lidar signal over Rencheng District from December 2021 to February 2022, a model of particle concentration estimation is established based on machine learning. First, the sliding average of 100 of the air quality index (AQI) in the previous 24-hour is used as the threshold value to divide the samples into two background conditions: clean and polluted. Then, the importance of input factors is ranked using random forest (RF) algorithm, and according to this importance rank, the factors are put into RF, mind evolution algorithm-back propagation neural network (MEA-BPNN), generalized regression neural network (GRNN) and wavelet neural network (WNN) models, respecitively. And then, by comparing the root mean square errors (ERMS) of different models, the optimal model and the optimal number of input factors for particulate matter concentration inversion under different atmospheric background are established. Finally, the particle estimation model developed is applied to the actual atmospheric observation data to evaluate the horizontal distribution of particle concentration under different atmospheric backgrounds based on the optimal model and input factors.