Multi-scale emission source inversion method based on three-dimensional UNet structure
By constructing a multi-scale emission source inversion method with a three-dimensional UNet structure, and combining ensemble Kalman filtering and deep learning, the problems of high computational cost and insufficient feature extraction of EnKF are solved, achieving efficient and accurate emission source inversion and improving the accuracy of atmospheric pollutant forecasting.
Patent Information
- Authority / Receiving Office
- CN Β· China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2026-02-13
- Publication Date
- 2026-06-12
AI Technical Summary
Existing EnKF methods are computationally expensive and difficult to deploy in business applications. Deep learning models do not fully extract features and do not make sufficient use of multi-scale information in emission source inversion, resulting in insufficient accuracy in air pollutant forecasts.
A multi-scale emission source inversion method based on a three-dimensional UNet structure is adopted. By constructing a three-dimensional spatial multi-scale feature fusion UNet model and combining it with an ensemble Kalman filter assimilation system, emission source inversion is performed quickly and with high accuracy using PM2.5 observation data and concentration forecast background field. A multi-scale feature fusion module is constructed and a multi-scale composite loss function is introduced to optimize the model training process.
It achieves emission source inversion with high computational efficiency and excellent inversion accuracy, can capture the multi-scale spatial structure of pollutants, improves the accuracy of pollutant numerical forecasting, reduces computational resource consumption, and is applicable to emission source inversion of various air pollutants.
Smart Images

Figure CN121706620B_ABST
Abstract
Citation Information
Patent Citations
Method and device for estimating PM2.5 concentration in three-dimensional space based on improved UNet
CN119418229A
Navigation industry carbon emission prediction method based on fusion of multi-scale feature extraction and time sequence dependence analysis
CN120725212A