Hyperspectral Remote Sensing Reflectamce Data Augmentation Method For Red Tide Dominant Species Based On Condition Generative Adversarial Net

Red tide is a significant global marine disaster, with the characteristics of various dominant species and varying outbreak times and locations. Different dominant species have different bio-optical properties due to differences in morphology and pigment content. Therefore, identifying dominant species is an important part of red tide monitoring. However, the lack of in-situ remote sensing reflectance data, the identification performance of red tide species is not good. Therefore, the expansion of red tide dominant specie samples is necessary for identification and to provide a basis for future studies. To tackle this problem, this paper proposes a model for augmenting hyperspectral remote sensing reflectance for different red tide dominant species, based on Conditional Generative Adversarial Net (CGAN). On the basis of the CGAN framework and the combined characteristics of hyperspectral remote sensing reflectance data, the Linear layer was taken as the basic layer of the network in the proposed model. It can improve the spectral similarity between the augmented and reference samples by linearly weighting the reflectance between different wavelengths. In order to evaluate the data correlation and spectral similarity of the model results, the Pearson correlation coefficient and spectral angle were calculated between randomly selected generated second derivative spectrum and in-situ remote sensing reflectance. Experimental results show that the generated spectra are in good consistency with the in-situ spectra, with an average Pearson correlation coefficient and an average spectral angle of 0.98 and 3.82°, respectively. Moreover, the support vector machine was employed to identify the dominant species of red tide based on the generated second derivative remote sensing reflectance spectra. The accuracy of identifying dominant red tide species is 92.5%.