Coastal Wetland Classification Based On Time Series Of Remote Sensing Image And Vegetation Phenology

Coastal wetlands are located in the transition zone between sea and land,and are special ecosystems with important functions such as flood protection,bank protection,nutrient cycling,biodiversity protection,and carbon sequestration.In recent years,with the rapid development of China's coastal zone economy,coastal wetlands have faced severe degradation risks due to the profound impact of human activities such as reclamation of tidal flats and the introduction of S.alterniflora.Strengthening coastal wetland monitoring and protection is of great significance to meet the urgent needs of the national coastal wetland ecological red line demarcation and adjustment,and to build a marine ecological environment governance system.The rapid development of remote sensing observation technology provides technical methods for the monitoring of large-scale coastal wetlands.However,due to the outstanding difficulties of image data sources,image interpretation methods and technologies,such as missing or insufficient data,and the indistinguishability of wetland features spectrum,it cannot provide mature technical support for coastal wetland monitoring.Based on Sentinel-2 image data with medium-high spatiotemporal resolution,this study proposed a coastal wetland classification method of pixel-level time series reconstruction and vegetation phenology,and applied it to Yancheng wetland.This study judged the phenological parameters and thresholds of different vegetations,discussed the applicability of the method in coastal wetland classification by using random forest algorithm,then explored the possibility of compressed classification of vegetation phenological variables,in order to provide a method for solving the lack or insufficiency of image data and achieving accurate classification of coastal wetland.The main conclusions were as follows:(1)Six remote sensing indexes of NDVI,SAVI,EVI,WAVI,GNDVI,and MCARI2 were selected for pixel-level time series reconstruction.Analysis showed that NDVI and SAVI were better remote sensing image feature indexes.The NDVI time series showed prominent vegetation phenology information,while the SAVI reconstructed time series curve was more stable and could fully reflect vegetation phenology information.(2)Four kinds of vegetation phenological feature fitting models,including asymmetric Gaussian function fitting(AG),double logistic function fitting(DL),harmonic function fitting(HA)and Savizky-Golay filtering(SG),were selected to reconstruct the time series.It was found that the function fitting model had obvious advantages over the smooth filtering model,and was more suitable for coastal wetland classification.Among the function fitting models,DL and HA could highlight the important vegetation phenological features,and AG could smoothly reflect the vegetation phenology change process.SG could retain the detailed phenological characteristics,but it was easy affected by outliers,which led to the deviation in phenological information.(3)Based on the evaluation of vegetation separability and research needs,the study selected the SAVI and double logistic function fitting model as the optimal time series reconstruction method,and determined the phenological parameters and thresholds for distinguishing different coastal wetland vegetations.The start time of growing season(May to June)could identify Phragmites,the end time of growing season(late December)could identify S.alterniflora,the decline rate at the end of growing season(>0.025)could identify Imperata cylindrica,in which amplitude(>0.35)was auxiliary,and amplitude(<0.25)and the growth rate at the beginning of growing season(<0.015)could identify Suaeda salsa.Vegetation discrimination phenological parameters and thresholds could provide the basis for distinguishing different vegetations for coastal wetland classification when image data was missing or insufficient.(4)The time series reconstruction and vegetation phenology feature classification methods had good applicability to coastal wetland classification,effectively improving the classification accuracy of Yancheng wetland,with an overall accuracy of 87.61%,a Kappa coefficient of 0.8358,and an overall improvement of 19.57% compared with the mono-temporal classification.This method could realize the accurate classification of mixed vegetation zone and the effective differentiation of the“foreign body with the same spectrum” by combining the phenological features,which effectively improved the classification effect of coastal wetlands.(5)Compressing the phenological variables of wetland vegetation by different numbers,it was found that compressed variables still have the ability to distinguish vegetation types,ensuring high classification accuracy.Based on this,the study proposed a compressed classification method based on four vegetation phenological variables,including the start time / end time of growing season,the growth rate at the beginning of season / the decline rate at the end of season.By avoiding the missing periods of image data and extracting characteristic variables that were not related to the vigorous growth period for classification,it could partially solve the problem of insufficient or missing image data,effectively expand the applicability of time series reconstruction methods,and provide other sensors new ideas for wetland classification.