Share Email Print
cover

Proceedings Paper • new

An unsupervised feature learning method to distinguish Sargassum
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

In recent years, the golden tide, which is caused by the explosive proliferation of Sargassum, has occurred frequently in China Seas. It has made a great negative impact on the marine ecosystem, aquaculture, and coastal tourism. Fortunately, satellite observation can monitor and track the growth of large algae such as Sargassum in a timely and effective manner, providing scientific basis for disaster prevention and mitigation in fisheries and environmental protection departments. Most of traditional extraction methods of macroalgae are pixel-oriented. Although these methods can be performed easily, they loss the rich texture information of the natural objects. The Sargassum seen from remote sensing imageries tends to aggregate in groups, like strips, covering several to dozens of pixels. Therefore, this paper considered distinguishing Sargassum from a certain area based on scene by utilizing contextual relationships among pixels and the diversity of spatial and structural features. In this paper, the image acquired by GF-1 during the golden tide disaster in the sea area near Jiangsu Province of China on December 31, 2016 were used. We adopter an unsupervised feature learning method to distinguish Sargassum. The Voting method was used to divide the original image into small image blocks guided by the corresponding saliency image. After 0-meanization and ZCA whitening, the initial weights were obtained by training the sparse autoencoder, then these weights were convolved as the convolution kernel to obtain the local features of the image, the features convoluted were passed. We pooled them to reduce the eigenvectors of the convolutional layer output so that the global statistical features of the image could be extracted. Finally, the Softmax classifier was used to distinguish the regions of Sargassum in the original image. The experimental accuracy was 77.79% and superior to the threshold extraction methods compared with the result of manual labeling.

Paper Details

Date Published: 14 October 2019
PDF: 12 pages
Proc. SPIE 11150, Remote Sensing of the Ocean, Sea Ice, Coastal Waters, and Large Water Regions 2019, 1115012 (14 October 2019); doi: 10.1117/12.2532576
Show Author Affiliations
Feng Ye, The Second Institute of Oceanography (China)
Zengzhou Hao, The Second Institute of Oceanography (China)
Yanlong Chen, National Marine Environmental Monitoring Ctr. (China)
Fang Gong, The Second Institute of Oceanography (China)
Haiqing Huang, The Second Institute of Oceanography (China)
Difeng Wang, The Second Institute of Oceanography (China)


Published in SPIE Proceedings Vol. 11150:
Remote Sensing of the Ocean, Sea Ice, Coastal Waters, and Large Water Regions 2019
Charles R. Bostater Jr.; Xavier Neyt; Françoise Viallefont-Robinet, Editor(s)

© SPIE. Terms of Use
Back to Top