A Novel Green Tide Detection Method Based on Superpixel Fisher Vectors in Multispectral Remote Sensing Images

Abstract

Green tide detection in multispectral remote sensing images (MSRSIs) is of great importance to protect the ecological security of marine environments. Existing green tide detection methods based on pixel-level features of MSRSIs have shortcomings to leverage the boundary and structure information of green tide areas, suffering from a mass of false alarms. To address this issue, this paper proposes a novel green tide detection method that utilizes superpixels as the basic processing units and fisher vector (FV) as feature encoding strategy, exploiting both spatial and statistical characteristics of MSRSIs. Firstly, MSRSIs are segmented via superpixels to enhance the spatial boundary characteristics of MSRSIs. Secondly, Gaussian mixture model is utilized to model the statistical distribution of superpixels, and the FV for each superpixel is calculated to extract the multi-order features therein. Finally, FVs of superpixels in MSRSIs are classified into green tide and non-green tide areas by the random forest classifier. Experimental results show that our proposed method outperforms the existing commonly used methods based on measured data.