Using GaoFen-1 to Distinguish Sargassum Horneri and Ulva Prolifera Blooms in the Western Yellow Sea Based on Machine Learning

In recent years, large-scale macroalgae blooms with two dominant species of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Ulva prolifera</i> (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">U.prolifera</i>) and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Sargassum horneri</i> (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S.horneri</i>) have occurred frequently and concurrently in the Yellow Sea and East China Sea. Existing remote sensing models for distinguishing between <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">U.prolifera</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S.horneri</i> have achieved high accuracy in specific scenarios but exhibited weak generalizability, which stemmed primarily from training data inadequately representing diverse scenarios. We first established a more representative sample set from 2015-2023 GaoFen-1, Wide Field of View imagery. This sample set encompassed diverse scenarios, including clear water, turbid water, and areas affected by thin clouds and sun glint. Through spectral analysis, we found that: 1) influenced by multiple factors, algal spectra exhibited significant variability and approximately 30% of algal pixels even had similar spectral shapes; 2) combining difference spectra (spectral difference between algae and surrounding water) and water turbidity could effectively distinguish these fuzzy pixels. Subsequently, we developed a random forest classification model, with input features encompassing three dimensions: algal spectra, difference spectra and turbidity. The overall accuracies of our model on both validation set and test set were higher than 90%. Compared to existing methods, our model could effectively distinguish spectrally similar algal pixels and improved classification accuracy by over 10%.