Green Tide Extraction Method Based On Multi-channel Convolutional Neural Network And Super-resolution

The Green Tide has always been a major concern of China’s coastal regions,which not only caused a lot of pollution to the national environment,but also caused a certain threat to coastal tourism,marine fisheries and aquaculture.At present,the extraction of green tides in the traditional method faces two major problems.First,in the green tide extraction method based on a vegetation index,the threshold is difficult to determine automatically.Even with the help of the current advanced GIS information development system,it requires manual Threshold selection is carried out interactively,and in order to avoid the mistake of terrestrial vegetation as the impact of green tide,it is necessary to mask the remote sensing image in advance,so it is difficult to achieve true green tide extraction and detection automation;second,the green The monitoring of tide is real-time.Due to the wide shooting range of low-Earth orbit satellites,it is often used to monitor the entire drifting dynamics of the green tide eruption,but due to the low resolution,the resolution of commonly used MODIS data is between 250m and 1000m Each pixel contains a large amount of feature information,and the monitored green tide coverage area and distribution area will be greater than the true value,which will bring some problems to the fine extraction of green tide.And so far,for the problem of green tide information extraction and detection,there has not been a set of processing flow based on deep learning methods.In order to overcome the above problems,this paper innovatively proposes a green tide extraction method based on multi-channel convolutional neural network and super-resolution.Firstly,a green tide extraction model(Green TideNet)is proposed based on a multi-channel convolutional neural network,which is mainly divided into a three-layer network structure.The first two layers of two-way UNet structure fully excavate image feature extraction,including low-level features and high-level features.Integrate global information through global average pooling,and finally further convolution integration.The experimental results show that,by designing the multi-path convolution structure,it not only successfully overcomes the difficult problem of threshold selection,but also solves the problem that the image contains inland vegetation and needs to be masked in advance.However,due to the low image resolution,there is still room for improvement in the extraction accuracy of green tides,and the influence of vegetation on islands in the sea cannot be avoided.Therefore,based on the Green TideNet model,a super-resolution reconstruction and attention mechanism module is added.The super-resolution module is responsible for improving the resolution of the image.The semantic segmentation model combined with the attention module further extracts the semantic features of the image.The technology combined with the semantic segmentation model further improves the accuracy of extracting green tide at low and medium scores.This paper proves the effectiveness of the proposed method for green tide extraction through experiments,and reduces the recognition error rate caused by insufficient image resolution and the impact of marine vegetation.