A knowledge-guided deep learning framework for Ulva prolifera green tide detection and quantification on MODIS imagery

Ulva prolifera (U. prolifera) green tides have become a recurring ecological issue in the Yellow Sea, posing serious threats to coastal ecosystems and regional economies. Traditional optical satellite index-threshold methods often require scene-specific threshold adjustment under complex marine environments, which may constrain their automation in large-scale applications. In this study, we present Green Tide Detection Network (GTD-Net), a knowledge-guided deep learning model for detecting and quantifying U. prolifera biomass using MODIS imagery. A key innovation of this study is the construction of a highly heterogeneous, multidimensional training dataset that captures variations in observation conditions, environmental backgrounds, and algae distribution patterns. By translating domain knowledge into learnable sample constraints, the dataset effectively guides model learning and improves robustness under heterogeneous observation conditions. Architecturally, GTD-Net augments the classical U-Net with multi-scale Inception modules, Residual Blocks, and the Convolutional Block Attention Module (CBAM), together with a hybrid weighted loss function to improve the delineation of boundaries and the detection of small or sparse targets. Ablation experiments showed that all components contributed positively to model performance, with the Inception and CBAM modules playing particularly important roles in multi-scale green tide feature extraction and background suppression. Comparative experiments showed that GTD-Net consistently outperforms traditional index-based methods and classical deep learning models, achieving an F1-score of 0.837 under complex observation conditions. When applied to MODIS time series data from 2007 to 2024, GTD-Net enables long-term, high-precision biomass estimation. For example, in 2019, its annual estimates were approximately 14% higher than those from the index-based method. These findings highlight the effectiveness and applicability of GTD-Net for large-scale monitoring of U. prolifera green tides.