A novel weighted loss function for improved deep learning-based detection of floating macroalgae blooms in multispectral imagery

Frequent blooms of macroalgae in China’s marginal seas have raised growing ecological and economic concerns. Deep learning techniques have been employed to detect floating macroalgae from multispectral ocean colour imagery, enabling efficient monitoring of their spatial and temporal dynamics. However, macroalgae pixels are vastly outnumbered by seawater pixels, resulting in severe class imbalance. Additionally, a key challenge is distinguishing macroalgae from noise, such as sun glint and imaging artefacts, due to similar spectral features. To address these challenges, we proposed a novel weighted loss function, which prioritizes algal feature learning within specific spectral index ranges while enhancing discrimination against spectrally similar noise. The performance was tested on two deep learning architectures, namely UNet-based and attention-based models, using multispectral imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imaging Radiometer Suite (VIIRS), and Ocean and Land Colour Instrument (OLCI) imagery. Experiments showed improved Precision, F1 score, and IoU, with the most significant gains on OLCI, likely due to its higher spatial resolution that captures algae-like noise patterns. The loss function was further applied to macroalgae blooms in the Caribbean Sea, achieving consistent gains. This study underscores the importance of tailored loss functions in enhancing deep learning-based monitoring of macroalgae blooms.