LAUSR

This paper presents a novel framework combining spectral analysis and machine learning for green-tide detection. The framework incorporates a squeeze and excitation (SE) attention module into a U-shaped generator of a generative adversarial network (GAN), and is referred to as squeeze and excitation oriented generative adversarial network (SE-GAN). In the SE-GAN, the normalized differential vegetation index (NDVI) is used as the preprocessing filter, which enhances the information associated with green-tide. The SE attention module recalibrates the feature maps so as to enhance the useful features conveyed from the generator’s convolution layer while suppressing less useful ones. Overall, the generator attempts to render images that contain green-tide in a way that highly approximates the reference images, while the discriminator characterizes the difference between the generated images and the reference images. The training scheme, which is adversarial, minimizes the f- divergence between the generator and discriminator. Consequently, compared to other green-tide detection algorithms only applicable in small-area scenes, SE-GAN can automatically detect green-tide in MODIS images of any size. Experiments with both large- and small-format MODIS imagery confirm that SE-GAN’s detection results are superior to those of five other commonly used methods.