Study On Wetland Information Extraction From The Minjiang River Estuarine Based On Combination Of Feature Knowledge And Decision Tree Model

Wetland is the important component of ecosystem,which is called "the kidney of the earth".It's noted that wetland can be used to protect biodiversity,purify water source and even control climate change.The largest estuary wetland in Fujian Provinces was one of the most important wetland areas in China,which located in the lower reaches of the Minjiang River.The increasing urbanization and industrialization have led to wetland losses in estuarine area of Mingjiang River over past decades.There has been increasing attention given to produce wetlands inventories using remote sensing and GIS technology.This group of studies were conducted in the lower reaches of the Minjiang River during timeperiod in March 2018 to collect the remote sensing images of different sources using Sentinel-2A data and Landsat8-OLI data,respectively.It is proposed to use advanced object-oriented classification technology combined with advanced decision tree and feature knowledge to accurately monitor wetland resources,effectively improve the limitations of traditional pixel-based image classification monitoring techniques,and greatly improve image classification accuracy.This paper firstly employed pansharp2 fusion algorithm to enhance spatial resolution while preserve image spectrum information for the Landsat imagery and different spatial resolutions of 10 m,15m and 30 m image in the study area were obtained.Further,Our results have developed an update classification system of wetland with the existing wetland system and high-resolution of field images.We performed the multi-scale segmentation procedure,taking the scale,hue,shape,compactness and smoothness of the image into account to get the appropriate parameters,using the top and down region merge algorithm from single pixel level,the optimal texture segmentation scale for different types of features was selected.According to the best segmentation object,This study established the typical sequence of various wetland types by representative wetland samples in the field survey and effectively avoided the interference of other land types during the classification process.Additionaly,by analysing the adjacent relationship of spatial objects,Based on the knowledge base of wetland classification features,the CART classifier was used to realize the automatic establishment of classification rule sets.The accuracy of wetland classification among different sources of image classification were compared using eCognition software.It can be concluded that the high-resolution of spatial images were key factor to determine the accuracte extraction of images.In this study,the total accuracy of the 10 m Sentinel-2A image was set as 89%,the Kappa coefficient was 0.87;the total accuracy of the 15 m Landsat8-OLI image was 84.08%,the Kappa coefficient was 0.80;the total accuracy of the 30 m Landsat8-OLI image was 82%.The Kappa coefficient was 0.79.The analysis of the intervening spatial feature index showed that the spatial complexity of the wetland type of the classified image objects has a close relationship with the spatial resolution of the image.The higher the spatial resolution of the image,the higher the index representing the complexity of the plaque,and the finer the wetland extraction.On the contrary,the lower resolution image is affected by more mixed pixels,and the boundary of the ground extraction was more blurred.