Remote-Sensing Monitoring Of Vegetation Change In Coastal Wetland Of Jiangsu Province Based On Machine Learning

Coastal wetland is an ecological buffer zone between the ocean and the land.Its abundant water resources,land resources and biological resources play an important role in stabilizing regional biodiversity and adjusting regional ecological balance,which is an important prerequisite for regional sustainable development.As an important part of wetland ecosystem,wetland vegetation has guiding significance to its ecological environment.Revealing the change of spatial pattern of typical vegetation in different years of coastal wetland plays an important role in understanding the succession mechanism and development trend of coastal wetland feature types.Based on multi-temporal Landsat Image data and GF-2 data,this paper uses VGG16,BP neural network,SVM and other machine learning algorithms,combining previous research results and field data sample points,research on the core area of yancheng coastal wetland reserve classification,after comparing the classification results accuracy analysis to explore in the study area under different spatial resolution information extraction and the optimal algorithm,optimal algorithm classification results using nearly 15 years in the study area typical vegetation regional spatial distribution and its change rule,Finally,the ca-markov model based on the improved transfer probability matrix is used to predict the distribution of wetland types in the study area in the next five years,which provides scientific reference for the establishment of reasonable management and control strategies for the vegetation system in the study area and the sustainable development of the ecosystem.The main conclusions are as follows:(1)Among the classification algorithms based on high spatial resolution remote sensing image(GF-2),the deep learning VGG16 model has the highest classification accuracy,with the overall classification accuracy of 98.71% and kappa of 0.99.It has the highest classification accuracy for three typical wetland vegetation,and the best classification effect for road,river and other structural features,which can effectively eliminate the traditional pixel based classification“ Salt and pepper phenomenon ".The optimal classification algorithm for Landsat is BP neural network,with an average overall accuracy of 90.05% and an average kappa coefficient of 0.88.(2)In the past 15 years,the total vegetation area of the study area has gradually increased,specifically,the area of Spartina alterniflora and Phragmites increased,and the area of Suaeda salsa continued to decline.The growth rate of reed is 167.66hm2/a and 1653.39hm2 in recent 15 years;the expansion rate of Spartina alterniflora is 68.69hm2/a and 1020.96hm2 in recent 15 years;The reduction rate of Suaeda salsa is 73.57hm2/a and 804.24hm2 in recent15 years.(3)In terms of spatial distribution pattern,Phragmites australis shows a trend of continuous expansion around the central and western part of the study area;Spartina alterniflora shows a trend of continuous extension to both sides of the inland and the outer sea,with the expansion of Spartina alterniflora,the coastline also moves to the outer sea;the spatial pattern of Suaeda salsa shows a trend of stable reduction,with alkali in the middle The canopy was encroached by the boundary vegetation,and the Suaeda salsa was replaced by alterniflora.(4)It is predicted that in the future,the spatial pattern distribution of all types of ground features will be stable,the area of Spartina alterniflora,reed and fish pond will continue to increase,while the area of Suaeda salsa will decrease.Among them,the growth range of Spartina alterniflora extends to the inland slowly,mainly to the open sea;the growth area of reed continues to expand from the original growth area to the surrounding area,and the growth area of Suaeda salsa becomes more and more fragmented,showing a trend of gradually narrowing from the surrounding to the center.