The Detection And Trend On Mangrove Based On Deep Learning

Mangrove is a wetland woody plant community.It is one of the marine ecosystems with the highest productivity and the richest biodiversity in the sea land ecotone of tropical and subtropical areas.Its ecological value is very important,which can consolidate the bank,prevent wind and waves and purify sea water and air.In addition,thanks to a large variety of fish and birds in mangroves,local fishery production and natural tourism can also be greatly promoted.However,due to the joint action of natural and human factors,the marine ecology has been seriously damaged,and mangroves have not been spared,and many damages have been suffered.It is reported that mangroves have decreased by 35% worldwide and are still disappearing at a rate of 1? 2% per year.Due to the special growth position of mangroves,they are often in the intertidal zone,the tide will periodically submerge the mangroves,and the mangrove forest is deep and dense,so it is difficult to carry out field investigation and sampling.At present,remote sensing technology has become one of the main technologies of mangrove segmentation.In traditional geography,in order to complete the task of mangrove segmentation,the band information of remote sensing image and the derived remote sensing spectral index are usually used to supervised learning.However,this method does not take the information of surrounding pixels into account in modeling,and the accuracy is lower than other methods.Although deep learning is widely used in remote sensing images,mangrove segmentation has not been mature due to the lack of appropriate data sets.Based on the project of "unbalanced transformation of geomorphology,environment and ecology between the middle and lower reaches of the Yangtze River driven by reservoir",this study aims to preprocess Sentinel-2 satellite remote sensing images based on the method of deep learning,and propose a mangrove segmentation model based on adversarial learning,The domain adaptive technology is used to make the neural network detect the mangrove area without labeling a large number of landsat8 satellite remote sensing images.At the same time,ARIMA model is used to model the time series of mangrove area in China from 1986 to 2020,so as to predict the change trend of mangrove area in China.The specific work is as follows:1.A mangrove segmentation model based on adversarial learning is proposed.The model introduces adversarial learning on the basis of U-Net,adds self attention mechanism to generate more sensitive feature maps for mangroves,and uses Focal Loss to focus on samples that are difficult to classify,so as to segment mangrove masks in sentinel-2 remote sensing images,which are verified by experiments.The models proposed in this chapter reach 90.5% in non-mangrove Io U,79.2% in mangrove Io U and 84.9% in m Io U,which are better than the other modelz,providing source domain data for later experiments;2.A transfer learning model based on domain adaptive technology is proposed and transfers the information learned by neural network from sentinel-2 satellite remote sensing image to Landsat 8 satellite remote sensing image,so that this paper can segment mangroves in Landsat 8 images without massive label of Landsat 8 satellite remote sensing image.Combined with remote sensing spectral index,the performance is further improved.After domain adaptation,the non-mangrove Io U,mangrove Io U and final m Io U in landsat8 data set reached 85.0%,69.5% and 77.2% respectively,which verified the effectiveness of the model;3.We collect landsat8 remote sensing images in China from 1986 to 2020 in batches,use the domain adaptive semantic segmentation model proposed in this paper to calculate the area of mangroves in China,use the time series prediction model to model the mangrove area in China from 1986 to 2020,and predict the change trend of Mangrove area in China in the next five years.Based on the deep learning method,this paper preprocesses the satellite remote sensing image,makes its own data set and trains the mangrove segmentation model.Finally,m Io U in Landsat 8 data set reaches 77.2%,which is better than the other methods;At the same time,according to the task and based on the model research results,the time series modeling of mangrove area in China from 1986 to 2020 is completed.