Research On Red Tide Hazard Detection Based On Improved DenseNet And GOCI Data

In China’s offshore areas,with increasing human activities,water bodies are becoming more eutrophic and red tide disasters occur frequently,disrupting the ecological environment of the sea and threatening human life and health.The study of red tide hazard detection technology can help people to better understand the occurrence of red tide and reduce the economic losses caused by red tide disasters.Compared with traditional detection methods,optical remote sensing technology provides the technical means to quickly acquire a large amount of sea surface information from space,and deep learning technology provides the technical means to achieve highly accurate red tide detection.However,there are still some shortcomings in deep learning-based remote sensing detection of red tide hazards: firstly,optical remote sensing will be limited by weather factors,which will have certain impact on the acquisition of effective experimental data,and the actual measured data is small,and there are problems of small samples and unbalanced sample distribution;secondly,at the boundary where red tide occurs,the current deep learning methods cannot extract sufficient red tide boundary features,and due to insufficient feature extraction,the The detection effect at the boundary of red tide is more ambiguous,and it is difficult to distinguish the small differences at the boundary of red tide,resulting in low accuracy of boundary detection of red tide disaster.To address the above problems,the main research of this paper is as follows:(1)To address the problem of limited samples in red tide detection and the limited improvement of red tide detection accuracy based on traditional methods,this paper uses DenseNet for red tide detection,using dense convolutional blocks and neighborhood spatial features to extract more red tide image features,making full use of the underlying boundary information,and solving the problems of limited improvement of detection accuracy due to small samples and unbalanced sample distribution;this paper This paper uses GOCI data of the Bohai Sea red tide in 2014 to conduct experiments.Compared with traditional machine learning methods and traditional networks such as U-Net,the DenseNet network achieves higher red tide detection accuracy and is more suitable for remote sensing red tide detection.(2)To address the problem of limited detection accuracy at the red tide boundary,this paper proposes a dense feature optimization-based red tide detection method(Red Tide DenseNet,RT-DenseNet).By adding the SE module to the RT-DenseNet Block,the model can extract more effective image features from sensitive bands for feature fusion when using the Dense strategy to extract features at multiple levels and scales;in addition,feature optimization is performed again in the Transition Layer to remove redundant features and retain important features.In addition,the model is optimized again in the Transition Layer to remove redundant features and retain important features,so that the neural network can focus more on the bands that are more sensitive to the occurrence of red tide and better perceive the subtle changes in the occurrence of red tide hazards,further improving the accuracy of red tide detection(98.03% detection accuracy).(3)In this paper,GOCI data of the 2016 Yangtze River estuary red tide were also used for experiments.Due to the severe cloud cover at the time of the occurrence of the red tide in the Yangtze River estuary,the available experimental data were small,which led to the insignificant improvement of the red tide detection accuracy of the RTDenseNet model.In order to obtain better detection results even for small samples,this paper proposes the Red Tide Inception-DenseNet(RTI-DenseNet)network model by adding the Inception structure to the RT-DenseNet model instead of the traditional convolution layer,so that the model can extract more effective features from remote sensing images of different scales during convolution.The model is able to extract more effective features from remote sensing images at different scales during convolution for red tide calling.The experimental results show that the proposed RTI-DenseNet model is able to extract more features from the same data to a certain extent.