An Adaptive Mixed Pixel Decomposition Method for Macroalgae Extraction Based on Remote Sensing Images

Harmful algal blooms caused by macroalgae pose a serious threat to the global marine ecosystem. Current high-precision monitoring methods for algal blooms primarily rely on remote sensing data with high spatial resolution, but the limited coverage of such data hinders large-scale and high-frequency monitoring. On the other hand, low-resolution remote sensing data offers wider coverage and higher monitoring frequency but lacks accuracy. Therefore, accurately identifying and quantifying algal biomass in low-resolution remote sensing data is a pressing issue. This study introduces a novel approach using machine learning techniques to establish a relationship between water background and algal endmembers, and an adaptive mixed pixel decomposition method is proposed accordingly. By leveraging Sentinel-2 MSI high spatial resolution data, the study employs three machine learning models - Support Vector Regression (SVR), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN) - to establish a model for the relationship between water background and algal endmembers. Subsequently, transfer learning is utilized to unmix large-scale low spatial resolution MODIS remote sensing images. The findings highlight the superior performance of the CNN model across various water environments and algal density distributions. This method streamlines the estimation of algal biomass and enhances the accuracy of remote sensing monitoring of macroalgae, presenting a novel technical approach for multi-source satellite data fusion and high-precision marine ecological environment monitoring.