The effective and rapid detection of red tide has significant research implications in China’s offshore regions, where severe seawater eutrophication leads to frequent red tide events. With the rapid development and widespread application of remote sensing and deep learning technologies, the technical means for high-performance, large-scale red tide detection are now available. In this paper, aiming at solving the problems of limited number of samples in red tide detection and the limited improvement of red tide detection accuracy based on traditional methods, we propose a red tide detection method based on improved DenseNet, which uses dense convolutional blocks and neighborhood space features to extract information at different levels and scales, makes full use of and integrates underlying boundary details and high-level semantic information, and solves the problem of limited improvement of detection accuracy caused by a small number of samples and an unbalanced sample distribution. At the same time, through the attention mechanism based on the squeeze-and-excitation (SE) module, feature weighting optimization is carried out for the bands conducive to red tide detection, which can further improve the detection accuracy. To verify the effectiveness of this method, we use Geostationary Ocean Color Imager (GOCI) data of the red tide that occurred in the Bohai Sea in 2014 in our experiment. The experimental results show that the proposed method achieves better red tide detection (overall classification accuracy: 98.03%) than state-of-the-art red tide detection methods and is more suitable for red tide detection by remote sensing.