Salt marsh carbon stock estimation using deep learning with Sentinel-1 SAR of the Yangtze River estuary, China

Salt marshes are pivotal in the global carbon cycle, serving as significant contributors to the blue carbon sink. Accurately estimating carbon stock in salt marshes relies on precise vegetation classification. Here, we developed the Salt Marsh Vegetation Classification Network (SVCN), a deep learning algorithm designed to classify three primary vegetation canopies ( S. alterniflora , P. australis , and S. mariqueter ) spanning from 2016 to 2023 over the Yangtze River estuary, China. The SVCN was initially trained using 412 vegetation samples and Sentinel-1 Synthetic Aperture Radar (SAR) data in 2018. Additionally, we trained three traditional machine learning models – Classification and Regression Trees, Random Forests, and K-Nearest Neighbors – to facilitate a comparative analysis of model performance. Leveraging the classified vegetation outcomes, we conducted salt marsh carbon stock estimations using the InVEST model. The results showed that the SVCN model outperformed the other three models, achieving an overall accuracy of 0.97. Salt marsh carbon stocks in the Yangtze River estuary exhibited an overall increasing trend from 2016 to 2023, with an average annual increase of 3.13 × 1 0 3 Mg ⋅ year − 1 . However, there was a notable decrease of 10.36% in 2017, primarily attributed to reductions in the area covered by S. alterniflora and P. australis , which decreased by 11.18% and 10.11%, respectively. These findings highlight the potential of deep learning models and the incorporation of salt marshes in carbon stock estimates to enhance accuracy. • A deep learning model SVCN for accurately classifying vegetation in salt marshes. • The SVCN model outperformed other machine learning models, achieving an accuracy of 0.97. • Salt marsh carbon stocks in the Yangtze estuary grew from 2016 to 2023, with a dip in 2017. • Contributions to salt marsh carbon stocks is: S. alterniflora > P. australis > S. mariqueter .