Cross-scenario transfer learning for estimating mangrove nitrogen and phosphorus content from field hyperspectral data to SDGSAT-1 and Sentinel-2 images

Mangroves play a critical role in maintaining biodiversity, supporting global carbon and nitrogen cycles, and contributing to the achievement of the United Nations Sustainable Development Goals (SDGs). Accurate estimation of their nitrogen and phosphorus content is essential for assessing the status of mangrove ecosystems. However, the spectral response characteristics of mangrove leaf nitrogen content (LNC) and leaf phosphorus content (LPC) remain unclear. These knowledge gaps hinder the development of robust predictive models across diverse environmental contexts. To overcome these issues, we collected 375 samples and 16,590 in situ full-spectrum hyperspectral data, and further proposed a novel Global-Fractional Order Sensitivity Analysis (G-FOSA) method. We analyzed for the first time the apparent and deep spectral characteristics of LNC and LPC for four typical mangrove species in China (Avicennia marina, Acanthus ilicifolius, Kandelia candel and Aegiceras corniculatum) using G-FOSA method. This study revealed that the LNC diagnostic wavelengths concentrated in the range of 697 nm–704 nm, while the LPC diagnostic wavelengths were mostly distributed between 691 nm–834 nm and 1869 nm–2236 nm. We developed a mechanism-guided retrieval framework based on these diagnostic wavelengths, and achieved the quantitative inversion from field diagnostic wavelengths to optical satellite (SDGSAT-1 and Sentinel-2) bands. Our experiment results confirmed that SDGSAT-1, the world's first science satellite dedicated to serving the 2030 Agenda for SDGs, performs better in estimating LNC and LPC (R² = 0.63). Finally, we utilized the advantages of cross-scenario transfer learning technology to design a novel domain adaptive transfer learning (DTL) model, which realized the cross-scenario retrieval of mangrove LNC and LPC across three typical mangrove regions, reducing estimation error (RMSE) by 0.6 %–41.1 % compared to the traditional FTL model. Our work provides new insights and a scientific basis for global mangrove conservation.