Research On Remote Sensing Monitoring Method Of Flooded Vegetation In Poyang Lake Wetland Based On Machine Learning

Wetland is an important ecosystem for human and nature,providing many useful ecosystem services.However,climate change and human activities have affected the hydrological system and ecological functions of wetlands in different ways,resulting in huge ecological challenges such as reduced lake wetland area and vegetation decline.The National Wetland Protection Plan(2022-2030)points out that improving the level of wetland monitoring and supervision is particularly important for protecting the quality and stability of wetland ecosystems.Flooded vegetation(FV)is widely distributed in rivers and lakes flooded wetland systems due to seasonal flooding,which has an important impact on the stability and functional integrity of wetland ecosystem structure.Although remote sensing technology is widely used in FV monitoring research,due to the complex and changeable geographical environment of FV,the traditional method is complicated and cumbersome,which limits the accurate monitoring of FV and the effective protection of wetland ecology.Machine learning is widely used in remote sensing image analysis tasks.The research on FV monitoring method based on machine learning is helpful to expand wetland monitoring method technology,improve wetland resource assessment,monitoring and management level,and provide important reference for the research and protection of wetland ecosystem.In this paper,the land cover classification system of Poyang Lake is constructed for FV.Based on multi-source remote sensing data such as Sentinel-1A,Landsat8 and DEM,spectral band characteristics,vegetation index,water body index,radar characteristics and terrain characteristics are constructed.The random forest classifier is used to evaluate and optimize the importance of different feature variables involved in the classification of Poyang Lake,and four feature variable set combination schemes are constructed.Based on three machine learning methods of support vector machine,random forest and gradient boosting tree,the training and accuracy evaluation of four feature variable set combination schemes are carried out.The optimal combination scheme and the classification results of the machine learning algorithm are selected to correct the ground objects,and the data set is augmented while making the deep learning data set.The constructed data sets are trained,evaluated and analyzed on Deep Lab V3 +,U-Net,Swin-Unet and improved Deeplabv3+ networks to explore effective machine learning monitoring methods for wetland FV,deepen and expand Poyang Lake wetland and flood monitoring research,and provide decision support for wetland ecological management and protection.The main results of this paper are as follows:(1)FV characteristic index optimization and scheme combination.In order to explore the influence of optical image,radar image and multi-source remote sensing data on FV information extraction of Poyang Lake,this paper constructs a classification system of surface coverage of Poyang Lake for FV,and constructs multiple feature variable sets such as spectral band characteristics,vegetation index,water body index,radar feature index and terrain characteristics.The feature variables are optimized and sorted by random forest bag error,and four feature variable set combination schemes are constructed(Scheme 1: spectral feature + vegetation index + water body index;scheme 2: Radar characteristics;scheme 3: spectral characteristics + vegetation index+ water index + radar characteristics;scheme 4: preferred feature variable set).In the process of Poyang Lake surface classification,the importance score of the feature variable set is ranked as vegetation index > water index > spectral band feature > radar feature index > topographic feature.The feature importance score in the classification of Poyang Lake features is related to the separability between features.The better the separability between features,the higher the score.The preferred feature variable set is NDWI,SAVI,NDVI,B4,VH,SDWI,EVI,FAI,B1,B2,RV,B3,VV.(2)FV feature combination scheme optimization and deep learning dataset construction based on machine learning algorithm.In order to select the best feature combination scheme and classifier to construct the deep learning method classification label data set and augment the original data set.Based on three machine learning models of random forest,support vector machine and gradient boosting tree,this paper evaluates the accuracy of four feature variable set combination schemes.The overall accuracy and Kappa coefficient are ranked as scheme 3 >scheme 4 > scheme 1 > scheme 2.The performance of machine learning classifiers is ranked as random forest > gradient boosting tree > support vector machine,and the F1-Score of various ground objects is ranked as water > FV > other ground objects > vegetation.The study found that machine learning methods using only spectral features or radar features are difficult to achieve efficient and accurate classification of Poyang Lake wetlands;the fusion of the two information can effectively improve the extraction accuracy of ground objects and significantly improve the efficiency of wetland FV monitoring.More importantly,the machine learning algorithm after feature index optimization not only reduces feature redundancy,but also enhances the recognition ability of ground objects.(3)Improvement and accuracy evaluation of neural network model based on Deep Lab V3 +.In order to improve the accuracy of FV extraction,this paper uses Mobile Net V2,Seg Former,convolutional attention and global attention modules to improve the Deep Lab V3 + model.In this paper,four deep learning datasets are constructed: 2-class feature variable A dataset,multi-class feature variable B dataset,2-class radar feature variable(VV,VH,SDWI,NDI,CR)C dataset,multi-class radar feature variable D dataset(VV,VH,SDWI,NDI,CR).Through U-Net,Deep Lab V3 +,Swin-Unet and improved Deep Lab V3 + models,the same data set is trained,verified and tested,and the extraction effect of different models on FV is qualitatively and quantitatively analyzed.The improved Deep Lab V3 + model in this paper is superior to U-Net,Deep Lab V3 + and Swin-Unet models in overall accuracy,accuracy,recall rate,F1-Score and MIOU on A,B and D datasets.The improved Deep Lab V3 + model has the best model accuracy on the A dataset.The overall accuracy is 93.58 %,the accuracy is 87.18 %,the recall rate is84.79 %,F1-Score and MIOU are 85.77 % and 76.96 %,respectively.Compared with U-Net,Deep Lab V3 + and Swin-Unet models,its MIOU increased by 5.36 %,12.93 % and 10.21 %respectively.In terms of model size,the improved Deep Lab V3 + model is only 75.8 M,which is135.9 M,42.7 M and 31 M lower than the original Deep Lab V3 +,U-Net and Swin-Unet models,respectively.In summary,the improved Deep Lab V3 + model not only lightens the network,but also improves the recognition ability of the model.The main innovations of this paper are as follows: Aiming at the effective monitoring of FV in Poyang Lake wetland,this paper mainly carried out three aspects: feature variable optimization,machine learning application and deep learning algorithm improvement.It is found that the random forest algorithm in common machine learning algorithms can effectively monitor wetland FV.(2)The improved Deep Lab V3 + reduces the model size on the basis of improving the efficiency of wetland FV monitoring,which can provide support for FV monitoring in Poyang Lake wetland.