Remote estimation of early-stage <i>Ulva prolifera</i> biomass in the Yellow Sea using high-resolution satellite imagery

Large-scale green tides dominated by <i>Ulva prolifera</i> in the Yellow Sea are the most severe marine ecological disaster in China. Many studies have used satellite data to map <i>Ulva</i> distribution and estimate biomass at peak-bloom stages. However, detailed analyses of the spectral reflectance of early-stage <i>Ulva</i> and the development of dedicated biomass estimation models remain scarce, constraining effective early detection and quantitative monitoring. In this study, based on in situ experiments on the Radial Sand Ridges off the Jiangsu coast, we found that early-stage <i>Ulva</i> exhibited lower reflectance peaks in the green (∼550 nm) and near-infrared bands compared to peak-bloom <i>Ulva</i>. Senescent <i>Ulva</i> showed a reflectance peak shifted to ∼590 nm and a rightward shift of the red-edge feature. Using water tank measurements, we developed biomass estimation models specifically for early-stage <i>Ulva</i> by relating biomass per unit area (BPA) to several spectral indices, including ratio vegetation index (RVI), normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), virtual-baseline floating macroalgae height index (VBFAH), tasseled-cap greenness index (TCG), and floating algae index (FAI). These indices were derived from in situ surface reflectance (<i>Ref</i>) and simulated Rayleigh-corrected reflectance (<i>R</i>_rc) under different aerosol optical depth for high-resolution satellite sensors (GF1-WFV, GF2-MSS, and Sentinel2-MSI). The VBFAH index showed the best performance for GF1-WFV and GF2-MSS (R² > 0.96, MAPE < 14%), while the FAI index performed best for Sentinel2-MSI (R²> 0.95, MAPE < 14%). Applying these models to high-resolution GF1-WFV and Sentinel2-MSI images accurately mapped pixel-level BPA and total biomass of early-stage <i>Ulva</i> blooms. The inter-sensor uncertainty was relatively low (14.76%), and total biomass estimates from <i>Ref</i>- and <i>R</i>_rc-based models showed minor differences (6.36% for GF1-WFV and 3.9% for Sentinel2-MSI). These findings fill a critical gap in the quantitative monitoring of early-stage <i>Ulva</i> blooms and support early warning and mitigation efforts.