Design And Production Of Mangrove Remote Sensing Products And Data Sets In Guangdong-Hong Kong-Macao Greater Bay Area

Mangrove is an important part of China’s ecosystem.Remote sensing technology has become an important means to monitor and protect mangrove and promote ecological sustainable development.In this thesis,mangroves distributed in the Guangdong-Hong Kong-Macao Greater Bay Area were taken as the research object,and high-resolution remote sensing images of Goole Earth and Landsat were used as experimental data sources.Object-oriented machine learning and deep learning methods were used to extract mangroves in the study area.Different algorithms were compared and the optimal algorithm was selected to generate the mangrove data set from 2000 to 2020.Among them,object-oriented machine learning algorithm can reduce the phenomenon of "same object with different spectrum,foreign object with the same spectrum" on the one hand,and can make full use of the geometric characteristics and texture characteristics of remote sensing images to further improve classification accuracy.Deep learning algorithm can achieve end-to-end intelligent extraction,independent can efficiently extract of shallow and deep characteristics of remote sensing image,to intelligent extraction of remote sensing image characteristics of shallow and deep character,make the classification results of automatic fitting,suitable for multispectral remote sensing image information extraction,at present has become the effective way of remote sensing image information extraction.The Deeplab V3 algorithm based on deep learning finally determined in this thesis improves the extraction accuracy of mangrove and the extraction accuracy of broken mangrove area.The generated deep learning training model has certain generalization and universality,and can deduce the mangrove distribution in different years.The main research contents of this thesis are as follows:(1)Download landsat8 and Google Earth remote sensing images in the study area,perform image clipping and image fusion for Landsat8 images,and perform image enhancement and clipping for Google Earth images to improve the quality of remote sensing images and lay a solid foundation for the next experiment;(2)The deep learning sample database of remote sensing images of Guangdong-Hong Kong-Macao Greater Bay Area is made based on Arcgis Pro software.Intelligent expansion of remote sensing image samples of deep learning can be realized by selecting different slice sizes,step sizes and flip angles,saving manpower and resources,improving efficiency and saving time;(3)Use Deeplab v3 algorithm based on depth of learning,by choosing different backbone model,training are derived for different data sources of remote sensing image mangrove intelligent extraction model,implementation of the experimental area of mangroves intelligent extraction has high precision,effective extraction of mangrove characteristics of deep and shallow features of remote sensing image,improve the classification accuracy;(4)In order to verify the advantages of the proposed model in mangrove boundary extraction,compared with other deep learning methods such as un ET based on deep learning,the combined method of Deeplab V3 and RESNET-152 based on deep learning in this thesis achieved better results in mangrove extraction of whole area and mangrove extraction of broken boundary.Among them,the recall rate is0.939,and the predictive reasoning effect of the model is also the most outstanding,which has a certain universality.(5)In order to verify the applicability of the model in this thesis to other remote sensing data sources and mangrove forests in other regions,the trained model is used to extract mangrove forests in The Guangdong-Hong Kong-Macao Greater Bay Area from Google Earth at different times and phases,where the recall rate is,The experimental results show that the deep learning model selected in this thesis is universal for mangrove extraction from remote sensing images of different regions at different times,which greatly improves the work efficiency and demonstrates the practicality of the model in this thesis.(6)In order to highlight the superiority of Google Earth image and deep learning,object-oriented machine learning algorithm is used for comparison.Landsat8 remote sensing image data is used as experimental data,feature index,NDVI and NDWI are selected,and then vegetation texture and geometric information are combined.Finally,object-oriented random forest algorithm and support vector machine method were used to calculate the distribution of mangrove forest,and the classification scheme of random forest combined with spectral characteristics,vegetation index,geometric information and texture information was the best.(7)The deep learn-based Deeplab V3 algorithm combined with high-resolution remote sensing images was used to manually correct the missing and misclassification caused by the deep learning model,and finally output the 2000-2020guangdong-Hong Kong-Macao Greater Bay Area mangrove data set through data transformation.To realize the monitoring and protection of mangroves in the Guangdong-Hong Kong-Macao Greater Bay Area,data support should be improved.