Green Tide Extraction And Drift Prediction Based On Deep Learning And GF-4

The Yellow Sea Green Tide Disaster is a marine ecological disaster that has occurred in the past decade,which has seriously affected the development of the tourism,aquaculture and transportation industries along the Yellow Sea.Remote sensing data is an important technical support for green tide monitoring and green tide drift prediction,and is of great significance for the operational monitoring of green tide disasters.In view of the above problems,this paper is based on the Convolutional Neural Network(CNN),combined with the mixed pixel linear decomposition and CFAR algorithm,and finally achieves the accurate extraction of the high-level 4(GF-4)satellite image green tide.Algorithm;through the application of green tide data and the application of Generative Adversal Network(GAN)model,the research of green tide drift prediction algorithm is completed.The main research content is divided into the following three parts:First,the CNN-based GF-4 satellite image classification algorithm is studied.The high-scoring satellite 4(GF-4)is a geostationary satellite that provides all-weather,high-temporal resolution images for green tide monitoring.Therefore,this paper focuses on the chlorophyseal extraction of GF-4 images;The characteristics of the model were selected,and the NDVI vegetation index model,which is most suitable for green tide extraction research,was selected,and the NDVI index was calculated for the GF-4 green tide image.The basic principle of the CNN model was studied,and the characteristics of the green tide data were set.The corresponding convolution kernel parameters were used to train the CNN model applied to the green tide extraction.The trained CNN model was used to classify the NDVI green tide image by seawater,green tide and mixed pixels.Second,the linear decomposition of the mixed pixels of the CNN classification is implemented to improve the accuracy of the green tide extraction.Based on the linear pixel decomposition model and the NDVI vegetation index calculation formula,the relationship between the green tide ratio and the NDVI vegetation index in the mixed pixels is derived.The function pixel is used to decompose the mixed pixels,and then the decomposition results are compared with the GF-1 green tide image extraction results in the simultaneous region to establish a linear relationship model.The CFAR algorithm is used to combine the linear decomposition results of the mixed pixels.Classification,and then complete the accurate extraction of the green tide of the entire GF-4 image.Taking the results of GF-1 green tide extraction as the standard,comparing the results of the algorithm with the traditional threshold method,the comparison results show that the accuracy of the algorithm is increased by about 1 time.Third,the Generative Adversarial Network(GAN)predicts the green tide drift image in the future.According to the requirement of GAN for large data training samples,the green tide data is processed.Based on the principle of GAN model,based on the time series of green tide images,multi-scale convolutional network is introduced as the generation and discriminant model of GAN model.The image gradient difference loss function is added to improve the stability of the model training.The parameters are trained to train the GAN model,and the trained model is used to predict the drift of the green tide.The results show that the image generated by the GAN model predicts the shape of the original image.The drift trend is basically the same,and the trained GAN model generates prediction samples much faster than the currently used numerical simulation operations,which effectively improves the drift prediction operation rate.