Research On Classification And Change Prediction Of Yellow River Delta Wetlands Based On Artificial Intelligence And LUCC Mode

Wetlands are extremely biologically and environmentally diverse ecosystems in the world,as well as one of the most fragile ecosystems in the world,and their conservation of both biology and environment is of great importance.As one of the most characteristic wetlands in China,the Yellow River Delta is subject to multiple negative impacts such as marine tidal action,Yellow River sediment deposits,salinization and anthropogenic factors,resulting in a complex and unstable composition of its land cover types.Therefore,it is important to clarify the spatial distribution pattern of wetlands in the Yellow River Delta to maintain the ecological security of wetlands and protect the wetland environment.The study of spatial pattern simulation and analysis of future land use has been a key topic in the LUCC(Land Use and Land Cover Change)research program.The purpose of this study is to predict the trends of land cover types and infer the causes of land cover type changes based on the spatial and temporal distribution of ecosystems and their spatial and temporal drivers.After clarifying the distribution of wetland types in the Yellow River Delta,multiple well-established LUCC models were compared to simulate and predict wetland changes in the Yellow River Delta.The main factors influencing the future changes of each wetland type in the Yellow River Delta are further analyzed by the weight ranking of the drivers generated from the prediction results,and this analysis can provide valuable information to urban planning and decision makers.To address the above problems,the following work is conducted in this paper.(1)In this study,a machine learning classification method based on superpixel segmentation and optimal feature subset is proposed.A random forest classifier combined with object-oriented and feature selection methods was used to classify a large range of wetlands in the Yellow River Delta,and the distribution of wetlands in the Yellow River Delta in 2020 was mapped.In this method,the initial segmentation of Landsat 8 OLI remote sensing images is performed by the superpixel segmentation method based on the watershed transform of H-minima marker,which solves the "pretzel phenomenon" after the classification,while geometric features such as compactness,aspect radio and shape index in object-oriented are introduced to increase the differentiation of each category.Based on the initial segmentation,a bilevel scale-sets model is used to record the complete process of merging each level,which improves the efficiency of segmentation.The SEA method is used to calculate the changes of LV and MI indices under the scale change,while the OGF and Bayesian methods are combined to automatically obtain the optimal segmentation scale.In this study,153 features in remote sensing images were extracted,and the optimal feature subset consisting of 78 features was selected using the RFECV method,which reduced the redundancy among features.The above method reduces the operations such as manual intervention thresholding and post-processing in the traditional method,and at the same time the method has the universality to deal with the classification of a large range of images.This method effectively improves the classification accuracy compared with the traditional pixel-oriented machine learning methods,and its overall accuracy is 91.74% with a Kappa coefficient of 0.9078.(2)After obtaining the wetland distribution map of Yellow River Delta,this study compared the prediction accuracy of CA Markov,FLUS and PLUS models in the historical image of wetland distribution in combination with a variety of driving factors affecting land use type change.The experimental results show that the PLUS model has a stronger ability to mine the drivers and a higher prediction accuracy,with an overall accuracy of 83.45% and a Kappa coefficient of 0.8178.Therefore,the PLUS model is used to simulate and predict the wetland pattern of the Yellow River Delta in 2032,starting with the image classification data in 2020.And by the ability of the PLUS model to mine the drivers in the land expansion simulation,the ranking of the corresponding driver contribution weights of each wetland category generated from the simulation results was further analyzed.