Exploring Global Aerosol Size Dynamics from 2001 to 2024 Using the Pretrained Remote sensing pIxel-based Spatial-teMporal Deep Neural Network

Artikel i vetenskaplig tidskrift, 2026

Aerosols shape both climate and air quality, yet global understanding of their size distribution remains limited by the challenge of retrieving fine-mode (fAOD) and coarse-mode (cAOD) aerosol optical depths. Here, we introduce the Pretrained Remote sensing pIxel-based Spatial-teMporal Deep Neural Network (PRISM-DNN), an advanced deep learning framework that couples unsupervised pretraining on vast unlabeled satellite data with supervised fine-tuning using ground-based measurements. PRISM-DNN extracts rich spatiotemporal features and produces stable, robust global fAOD and cAOD retrievals for 2001-2024. Compared with established satellite products (POLDER and MISR), PRISM-DNN achieves markedly higher correlation coefficients with AERONET, reaching 0.94 for fAOD (+26%) and 0.91 for cAOD (+58%) relative to their mean performance. The model also performs well on additional ground networks excluded from training with correlation coefficients of 0.84 and 0.80 for fAOD and cAOD, respectively. The 24-year data set reveals pronounced regional disparities, with eastern China showing substantial declines in both fAOD and cAOD consistent with long-term emission control efforts, whereas parts of India exhibit increasing fine-mode aerosol loading associated with intensified human activities. These findings highlight how integrating unsupervised pretraining enables robust long-term aerosol component retrievals, strengthening satellite-based environmental monitoring and paving the way for broader AI-driven advances in earth system science.