COMPARISON OF QINZHOU BAY WETLAND LANDSCAPE INFORMATION EXTRACTION BY THREE METHODS

Wetland ecosystem plays an important role on the environment and sustainable socio-economic development. Based on the TM images in 2010 with a pretreament of Tasseled Cap transformation, three different methods are used to extract the Qinzhou Bay coastal wetlands using Supervised Classification (SC), Decision Trees (DT) and Object -oriented (OO) methods. Firstly coastal wetlands are picked out by artificial visual interpretation as discriminant standard. The result shows that when the same evaluation template used, the accuracy and Kappa coefficient of SC, DT and OO are 92.00%, 0.8952; 89.00%、0.8582;91.00%、0.8848 respectively. The total area of coastal wetland is 218.3km by artificial visual interpretation, and the extracted wetland area of SC, DT and OO is 219 km, 193.70km、217.40km respectively. The result indicates that SC is in the first place, followed by OO approach, and the third DT method when used to extract Qingzhou Bay coastal wetland. * Corresponding author: LUO Ming-liang. E-mail:Lolean586@163.com。 1.INTRODUCTION Wetlands is a unique ecosystem which is formed by the interaction between water and land, being one of the important natural habitats and ecological landscapes whose biodiversity is the richest (Yang, 2002a). According to preliminary statistics, the total wetlands area of China is about 6,594 million hm, about 10% of the world's wetland area (Wu et al, 2007). Using RS technology to extract the information of wetland timely and accurately is an important base for Wetland Research. There are currently a large number of research results, almost covering the most aspects of wetland research, mostly focusing on wetland definition and classification(Yang,2002), wetland formation and characteristics, resource distribution, wetland eco-tourism, landscape, tourism development and wetlands protection, etc(Yang, 2002b). In recent years, the researchers are always focusing on the following topics such as the physical, chemical, biological processes and mechanisms of wetland ecosystems; relationship between the processes and the functions (Yang, 2002b); the concept of wetland ecosystem’s health, diagnosis indexs, meantime devoting to improvement the ability of early warning(Sun et al , 2013). Due to the unique and complex nature of the wetland ecosystem, visual interpretation is used to be the main method of wetland information extraction with high precision. But this method often takes more manpower and plentiful time (Yang,2002b). The implementation of interactive wetland information extraction is a hot spot in current research based on RS software platform of Erdas. Many interactive methods have been used such as supervised classification (SC), decision tree (DT) and object-oriented (OO) method to extract wetland information (Huan et al, 2009; Shen et al, 2007). Due to different understanding and different purpose of information extraction, as well as the diversity of wetlands, a extensive distribution and variability of water environment, etc., one single method used only to extract wetland information is often difficult to The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-4, 2014 ISPRS Technical Commission IV Symposium, 14 – 16 May 2014, Suzhou, China This contribution has been peer-reviewed. doi:10.5194/isprsarchives-XL-4-21-2014 21 obtain the desired results (Yin et al, 2010). This article selects qinzhou bay as the sample area, comparing and analyzing the difference of SC, DT and OO, which are often used to extract wetlands information.

are always focusing on the following topics such as the physical, chemical, biological processes and mechanisms of wetland ecosystems; relationship between the processes and the functions (Yang, 2002b); the concept of wetland ecosystem's health, diagnosis indexs, meantime devoting to improvement the ability of early warning (Sun et al , 2013).Due to the unique and complex nature of the wetland ecosystem, visual interpretation is used to be the main method of wetland information extraction with high precision.But this method often takes more manpower and plentiful time (Yang,2002b).
The implementation of interactive wetland information extraction is a hot spot in current research based on RS software platform of Erdas.M any interactive methods have been used such as supervised classification (SC), decision tree (DT) and object-oriented (OO) method to extract wetland information (Huan et al, 2009;Shen et al, 2007).Due to different understanding and different purpose of information extraction, as well as the diversity of wetlands, a extensive distribution and variability of water environment, etc., one single method used only to extract wetland information is often difficult to obtain the desired results (Yin et al, 2010).This article selects qinzhou bay as the sample area, comparing and analyzing the difference of SC, DT and OO, which are often used to extract wetlands information.

1.1The study area
Qinzhou Bay locates in the south of Qinzhou City, the south of the Guangxi Zhuang Autonomous Region, belonging to the BeiBu Gulf, and it is one of the most important part of the Beibu Gulf.

Data S ources And Pretreatment
Landsat TM image data of Qinzhou in 2010 was used with the horizontal resolution of 30 m.In order to get the boundaries of Qinzhou Bay, digital elevation models was used to clip the study area by watershed division.
TC transformation is also called the KT transformation, and its essence is a special method of principal component analysis.
TC not only independs on a single image but also can compare soil brightness and green degree produced by different images, with the aim of image enhancement in extraction of vegetation information (Zi et al, 2011).TC transformation was used to enhance the brightness, green degree and humidity of the image information.

Object-Oriented Method
Object-oriented essentially differes to the traditional classification method based on pixels, for OO not dealing with single pixel, instead the whole image is regarded as a combination of different objects which can be interpreted   respectively.The result shows that extraction of wetland area through SC is most similar with visual interpretation with a more area of 0.7 km 2 , while OO resulting in 0.9 km2 missing and DT 28.3 km 2 missing.On the whole both classification accuracy and Kappa coefficient of supervised classification are higher than the other two methods.
In general, OO classification method should be optimal.While in this study, the coarse horizontal resolution of TM image and the particularity of coastal wetlands will be the main factors influencing the classification accuracy.As it known that horizontal resolution of TM image is only 30m, while existed research shows that OO approach will get the best results when 10m, 0.61m or 0.5m horizontal resolution image used with the classification accuracy as high as 90% (Yu and Zhan,2012;Hou et al,2010;Li etal,2008;Huang et al,2010;Xu etal,2013).
Secondly, QinZhou bay wetlands are mainly distributed in MAO tail sea -LongM en islands -inland sea area of longM en port and the coast of south sea area of longM en port.The governed types of wetland are as followed, such as beaches, Although the total classification accuracy of SC is higher than But in general the water and wetlands can be separated apart.
SC is mainly according to sample pixels.Since the differences of tidal flats, mangroves and Shrimp ponds are bigger than other, then they are easier to be identified, while oyster rows is similar with water, their accuracy is not good as the features above.

CONCLUS ION AND DIS CUS S ION
A comparison of SC, DT and OO used to extract QinZhou Bay wetland after TC transformation has been done, and the result shows that a certain accuracy of wetland information can be obtained by any one in the three methods.The suitable method should be selected according to specific data accuracy and classification requirements.
When the same classification characteristics and the premise of accuracy evaluation samples used, the accuracy and Kappa coefficient of the three methods are 92.00% and 0.8952, 89.00% and 0.8582, 91.00% and 0.8848 respectively.The results indicate that SC used to extract Qinzhou Bay wetland will achieve the best overall effect.
In the process of coastal wetland information extraction, a small number of wetland and water mixed together, and the OO method does not show its advantage when 30m resolution TM The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-4, 2014ISPRS Technical Commission IV Symposium, 14 -16 May 2014, Suzhou, China This contribution has been peer-reviewed.doi:10.5194/isprsarchives-XL-4-21-2014

Figure
Figure. 2 Analysis process 2.2 S upervised Classification SC is often applied to more understanding of the situation in study area.M aximum likelihood classification also known as bayesian classification, is supervised classification based on image statistics, which is a typical and the most widely used in supervised classification method, and is also regarded as the

Figure
Figure. 3 Classification Template

(
Wang et al, 2009).OO utilizes spectrum information, texture feature and topology relations together to segment multiscale objects joint the thematic information.Fuzzy discrimination function is established through visual interpretation to classify different objects(Sun,Tong and Qin, 2008).OO is implemented based on software platform of ENVI and the process can be divided into two steps: finding object firstly and extracting feature followed.The steps for finding objects as following: image segmentation, block merge, block refining, computing object properties.Feature extraction can be achieved by supervised classification or classification based on rules.After many experiments(Yu and Zhan,2012;Hou et al,2010;Li   etal,2008;Huang et al,2010;Xun etal,2013), the segmentation and merging threshold used are 51 and 48.3 respectively.The process of OO can be organized as figure6.

Figure. 6
Figure.6 Flow chart of OO tidal flats hygrophilous mangroves, fish (or shrimp) ponds, the big oyster beds of Oyster row.The amount of wetland types and area Inland sea are bigger than offshore area.From the viewpoint of spatial distribution, the main aggregated wetland includes farms, tidal flats and mangrove forests.Oyster row concentrated at M aowei sea -the mouth of longM en Port and the Shrimp (or fish) ponds partly distribute at the northeast shore of M AO wei sea and the big ring village.The main type which distributes discretely is shrimp ponds near the shoreline.The size of every oyster rafts ranges between 20m and 40m, while oyster rafts connected into a large scale pieces.The average size of shrimp pounds is at about 25*25m.Obviously TM 30m horizontal resolution is insufficient to express shrimp discrete distribution, thus resulting in lower classification accuracy of OO.From the viewpoint of spatial location, the water is assigned into the wetland through SC as a mistake, consequently resulting in the increasing of wetland area.The typical case is that an isolated Gold Nest Reservoir and a small part of Dafeng River Basin divided into wetlands.Some scattered wetlands (average size about 2.44* 10 -4 km 2 ) produced by OO method, are divided into water bodies, mainly distributed in M AO sea, while in the results of DT the naccuracy region scattering along rivers and estuary partly with the average area about 2.70 * 10 -4 km 2 .Some interpretation error happens in the sub-type of wetland, such as Gold Nest Reservoir is interpreted into the wetland, while the oyster rows into urban landuse by SC.The oyster rows, tidal flats and mangrove forests of longmen islands are divided into water by DT.MAO tail farms (mainly to shrimp pounds) are interpolated into farmland or water, which locates in the north bank of M aowei Sea by OO.And other inaccurate classification by OO is that tidal flats mangrove is categorized into water at the shore of M aowei Sea.The finnal classification by SC, DT and OO is shown in Figure 11.b, 11.c and 11.d, while Figure 11.a is the original TM images.
Figure 11.The extraction results contrast The connection between Damiandun (Tortoiseshell Island) at south of Rhino Feet Peninsula and Tiantang Corner of Qisha Peninsula is the south border of Qinzhou Bay.The north of Qinzhou Bay is M aoWeiHai with greatly developed aquaculture.There are hundreds hectares of M angrove Forests growing in the region.Diurnal tides is he main type in Qinzhou Bay with an average annual temperature 21.3℃.According to the status quo of China's wetland classification system, "Convention on Wetlands" and "national wetland resources investigation and monitoring of technical regulations", Qinzhou Bay wetlands in southern China is a typical coastal wetlands.
, Tidal flats, M angrove, River wetland, Natural vegetation Artificial wetland Aquaculture ponds, Coastal farms (such as Shrimp ponds), Reservoir, Paddy field, Artificial vegetation Table1.Classification systems of Qinzhou Bay Wetland Since this study focuses on the coastal wetland, the key of information needed to extracted is as followed: tidal flats, farms, vegetation, rivers and seas.Among those elements, the tidal flats include beach wetland, mangroves and Shrimp ponds farms, while vegetation includes shelter forest, trees and natural forest.Artificial wetland is mainly refers to farms, fish farms formed by the pond wetland, such as shrimp ponds (although artificial trace existed, it still maintains the characteristics of wetlands).Because the horizontal resolution of TM images is only 30m, the sea, river and lake are unified into water bodies, while Tidal flats, mangroves, Shrimp ponds farms into coastal wetlands.Urban landuse types include urban land, village land,

OF WETLAND INFORMATION EXTRACTION 2.1 Technical Process
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, VolumeXL-4, 2014   ISPRS Technical Commission IV Symposium, 14 -16 May 2014, Suzhou, ChinaThe International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, VolumeXL-4, 2014   ISPRS Technical Commission IV Symposium, 14 -16 May 2014, Suzhou, China This contribution has been peer-reviewed.doi:10.5194/isprsarchives-XL-4-21-2014 the DT methods, the definition of rules and the determination of variable are the key in wetland extraction.Water and coastal wetland choose the same band of 4, and wavelength ranges in 5-38 and 35-60 respectively.Since the spectral characteristics of artificial breeding farms, such as shrimp pounds, seems like water surface, which leads to a less obvious results, as it shows that water is confused with artificial breeding farm.But on the whole, it still reachs the classification accuracy of 91%.The comparisons of three methods used are as shown in tableThis contribution has been peer-reviewed.doi:10.5194/isprsarchives-XL-4-21-2014 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, VolumeXL-4, 2014   ISPRS Technical Commission IV Symposium, 14 -16 May 2014, Suzhou, Chinaother methods, while accuracy of the producers and the users of SC are the lowest.On the contrary, the OO methods has the highest accuracy as table 3-5 shown.The texture feature of shrimp ponds and oyster rows is obvious, so those feature can be picked out more easily by OO.Although it is difficult for DT to define the threshold of the wetland and water, since in the value of 35 to 38 water and wetlands are mixed together, which mainly reflected in the oyster rows and Shrimp ponds.