Related Experiment Video
Updated: Jun 10, 2026
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
11.0K
Exploring Global Aerosol Size Dynamics from 2001 to 2024 Using the Pretrained Remote sensing pIxel-based
Xing Yan1,2, Jiayi Chen1,2, Hans W Chen3
1State Key Laboratory of Earth Surface Processes and Disaster Risk Reduction, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China.
Environmental Science & Technology
|February 21, 2026
Summary
A new AI model, PRISM-DNN, accurately retrieves global fine-mode and coarse-mode aerosol optical depths from satellite data. This deep learning approach improves climate and air quality monitoring, revealing regional trends in aerosol pollution.
Area of Science:
- Earth System Science
- Atmospheric Science
- Artificial Intelligence
Background:
- Aerosols significantly impact climate and air quality, but global retrieval of their size distribution, specifically fine-mode (fAOD) and coarse-mode (cAOD) aerosol optical depths, is challenging.
- Existing satellite-based methods face limitations in accuracy and robustness for long-term aerosol component retrievals.
Purpose of the Study:
- To develop an advanced deep learning framework, PRISM-DNN, for stable and robust global retrieval of fAOD and cAOD.
- To leverage unsupervised pretraining on satellite data and supervised fine-tuning with ground measurements for improved aerosol optical depth estimation.
- To analyze 24-year global aerosol trends and regional disparities using the developed model.
Main Methods:
- Introduction of the Pretrained Remote sensing pIxel-based Spatial-teMporal Deep Neural Network (PRISM-DNN), a deep learning framework.
- Coupling of unsupervised pretraining on extensive unlabeled satellite data with supervised fine-tuning using ground-based AERONET and other network measurements.
- Extraction of spatiotemporal features for global fAOD and cAOD retrievals from 2001-2024.
Main Results:
- PRISM-DNN achieved significantly higher correlations with AERONET data compared to established products: 0.94 for fAOD (+26%) and 0.91 for cAOD (+58%).
- The model demonstrated strong performance on independent ground networks, with correlation coefficients of 0.84 for fAOD and 0.80 for cAOD.
- The 24-year dataset revealed distinct regional aerosol trends, including significant declines in eastern China and increasing fAOD in parts of India.
Conclusions:
- Integrating unsupervised pretraining enables robust, long-term aerosol component retrievals, enhancing satellite-based environmental monitoring.
- PRISM-DNN provides a powerful tool for understanding aerosol impacts on climate and air quality, paving the way for AI-driven advances in earth system science.
- The findings underscore the importance of AI in addressing complex environmental challenges and informing policy decisions.
