Quantitative assessment of urban wetland dynamics using high spatial resolution satellite imagery between 2000 and 2013

Accurate and timely information describing urban wetland resources and their changes over time, especially in rapidly urbanizing areas, is becoming more important. We applied an object-based image analysis and nearest neighbour classifier to map and monitor changes in land use/cover using multi-temporal high spatial resolution satellite imagery in an urban wetland area (Hangzhou Xixi Wetland) from 2000, 2005, 2007, 2009 and 2013. The overall eight-class classification accuracies averaged 84.47% for the five years. The maps showed that between 2000 and 2013 the amount of non-wetland (urban) area increased by approximately 100%. Herbaceous (32.22%), forest (29.57%) and pond (23.85%) are the main land-cover types that changed to non-wetland, followed by cropland (6.97%), marsh (4.04%) and river (3.35%). In addition, the maps of change patterns showed that urban wetland loss is mainly distributed west and southeast of the study area due to real estate development, and the greatest loss of urban wetlands occurred from 2007 to 2013. The results demonstrate the advantages of using multi-temporal high spatial resolution satellite imagery to provide an accurate, economical means to map and analyse changes in land use/cover over time and the ability to use the results as inputs to urban wetland management and policy decisions.

Wetlands are the most productive ecosystems in nature and important environments for humans 1 . Wetlands are known as the "kidneys of the earth" because of their important hydrological features and element cycles, and they can also be considered as "biological supermarkets" because of their large food webs and rich biological diversity 2,3 . Wetlands have far-reaching influences on the urban environment, including on the hydrological cycle, flood control, shoreline protection, climatic regulation, landscape building, natural species protection, and ecosystems service functions [4][5][6][7] . However, as human activity has continuously increased over the past hundred years, many wetlands have been lost, and the remaining wetland areas have substantially decreased, which has caused a large loss of natural habitats 8,9 . Recently, due to the progress of society and the rapid development of modern industry, land reclamation, water pollution and excessive deforestation have become increasingly serious issues, especially for urban wetlands. With the acceleration of urbanization, a large number of wetlands disappeared 10 . Monitoring urban wetlands and detecting their changes over specified time periods are necessary to ensure wetlands protection and optimal use 11 .
Satellite remote sensing provides a considerable opportunity to monitor land use and land cover (LULC) and detect changes because of the rapid, synoptic and repetitive capabilities of remote sensing 12 . Multi-source satellite images provide efficient information on the temporal trends and spatial distributions of urban areas that is needed to understand, model, and project land changes 13 . Specifically, moderate-resolution multispectral sensors (e.g., Landsat, SPOT, and ASTER) have been successfully used for studying the extent of flooding and affected land cover 14 , detecting the presence and composition of wetlands in heterogeneous landscapes 15 , and monitoring invasive plant species 16 . However, few studies have focused on the detection of small wetlands and ponds as small as 0.2 ha that are often important critical habitats. Sensors with 10 to 30-m spatial resolution typically require a minimum of 9 pure pixels (0.9 ha) to identify a feature 17 . Because of mixed pixels, the smaller wetlands are often missed 18 . At a local scale, the small wetlands are crucial to the stability of ecosystems, particularly for the maintenance of biodiversity 19 . Very high spatial resolution multispectral sensors (e.g., QuickBird and IKONOS) provide fine spatial resolution to capture smaller wetlands 20 , but the within-class spectral variance will be increased using these sensors. Therefore, in order to solve this problem, an appropriate classification algorithm must be employed 21,22 .
Recently, object-based image analysis (OBIA) has been applied more frequently for image classification and change detection in wetland systems 23,24 . Object-based approaches consider landscapes as aggregations of meaningful objects corresponding to ground entities and patches of surface cover 25 . In addition to being applied to extract features in urban and rural landscapes, OBIA techniques help provide detailed classifications of natural plant communities that are represented in wetland areas 26,27 . Previous studies have found that the OBIA approach performed better than the pixel-based approach for classifying urban areas and mapping LULC and LULC changes 20,28 . Fournier et al. reviewed mapping methods for wetland application and identified the OBIA approach as the most appropriate due to its ability to solve the spatial heterogeneity of wetlands 19,29 . OBIA approaches have also been successfully used in wetlands research to classify major wetland cover types 25 wetland vegetation species 27 and map wetlands in both natural and human landscapes 19 .
There are various methods for using satellites to determine wetland change in urban environments [30][31][32][33] . These methods can be broadly grouped into two general types, change enhancement methods and 'from-to' change information extraction methods. The former method (for example, image differencing) does not explicitly identify the types of land that have changed and only provides change/no change information and perhaps the relative magnitude of the change. The latter method, the post-classification approach, is one of the most commonly used and effective techniques 11 . Post-classification comparison methods use separate classifications for images acquired at different times to produce difference maps from which 'from-to' change information can be generated 34 . In addition, 'from-to' quantitative information about the type of LULC changes can be gained from a cross-tabulated change matrix 33 .
The Xixi Wetland is located in downtown Hangzhou, a city in China known for its rapid economic development and tourism, and it is a Ramsar Site 11 . However, with economic development and the expansion of the city, this area of wetlands has decreased rapidly. Its wetland functions have gradually become degraded, environmental pollution has substantially increased, and wetlands protection has become more difficult. The Xixi Wetland has faced combined pressures from natural and human factors. For protection purposes, quantitative wetland change detection and a driving factors analysis are essential to better understanding the relationship between anthropic activities and natural wetland systems. This

Materials
Study area. The Xixi Wetland, which is known as the "kidneys of the earth", was selected as the study area ( Fig. 1). The wetland complex is located west of Hangzhou City between 30°15′-30°17′N and 120°1′-120°6′E. The wetland previously covered an area as large as 60 km 2 , but it has shrunk to 11 km 2 because of urbanization in Hangzhou over the past 30 years. The wetland mainly contains fish ponds, rivers, and vegetation communities. In addition, the wetland also contains centralized villages and a few scattered areas of houses and trees 35 . The circled squares in the map of the Xixi Wetland show the rapidly developing economic zone. The area of wetlands has gradually been reduced due to industrialization and urbanization, including building construction, reclamation, and pollution. The resources in the Xixi Wetland face combined pressures from natural and socio-economic factors that must be resolved. Therefore, the Hangzhou Xixi National Wetland Park was constructed in 2005 to protect the Xixi Wetland. In 2009, it was successfully included in the "List of Wetlands of International Importance (the Ramsar List)" 36 .
Data acquisition and processing. High spatial resolution satellite imagery. To detect LULC changes, images that represent several stages within the same season are preferable. However, it was difficult to find time-series of high spatial resolution satellite images that were well matched within the same season for several reasons but primarily due to the weather conditions in this area. Five cloud-free images were acquired in January 2000, December 2005, January 2007, December 2009, and December 2013, covering the whole Xixi Wetland area, and these images were selected under the constraints of limited suitable images in the archives ( Table 1). All images were acquired from the Siwei Worldview Technology (Beijing) Co., Ltd. (www.siweidg.com). All images were provided in the Universal Transvers Mercator projection (UTM Zone 50 N) and WGS84 datum. These were careful georeferenced to a root mean squared error within 2 pixels using a 1 st -order polynomial transformation method. Finally, they were clipped to the same extent.
Reference data. Reference data were developed for each of the five years and then randomly divided for classifier training and accuracy assessment. The reference data for 2013 were a field verified set of reference sites collected in the early spring of 2012. This data set was created by collecting cover type information for a stratified random sample of 160 points with 20 points per level III class. The strata were from a previous classification year; approximately 70% of the data were used for training and 30% for the accuracy assessment.

Methods
Classification system. Currently, many international organizations and scientists have developed wetland classification systems based on different research perspectives 37 . The Land cover classification system developed by Anderson et al. 34 and the Ramsar Convention database 10 are used to define classification system used in this. The Ramsar definition of wetlands is broad, including not only vegetated wetlands but also lakes, coral reefs, and even underground caves. Therefore, the classification system is an open-ended, three-level hierarchical system organized by classes and types to provide robust information on wetlands (Table 2). At the first level, all land types are defined as urban wetlands or non-wetland areas according to the Ramsar definition. At the next level, urban wetland is divided into vegetation and water, and non-wetland mainly contains urban areas. At the third level, vegetation is classified by the dominant vegetation type: forest, herbaceous, cropland, or marsh vegetation; water is classified as river or pond; urban is classified as building or road.
Feature extraction. Feature extraction uses an object-based approach to classify imagery, where an object (also called a segment) is a group of pixels with similar spectral, spatial, and/or texture attributes. Traditional classification methods are pixel-based, meaning that the spectral information in each pixel is used to classify the imagery. With high-resolution panchromatic or multispectral imagery, an object-based method offers more flexibility in terms of the types of features that can be extracted. In this study, the object-based approach in the information extraction. This approach was mainly a two-step process, segmentation and classification 38 . In addition, to improve the accuracy of classification results, a visual interpretation based on high resolution images from Google Earth imagery was used for post-classification.
Feature development. Six input layers were used for the image segmentation process, including four multispectral layers (blue, green, red, and near-infrared), a normalized difference vegetation index (NDVI) layer, and a standard deviation texture layer 19 . The NDVI is a simple graphical indicator that can be used to assess whether the pixel contains live green vegetation or not, and can also be used to separate water from dry land and delineate wetland boundaries 19 . The NDVI layer was calculated from the red and near-infrared bands of the satellite images. Texture refers to the spatial variation of image tone as a function of scale and can reveal differences between classes in a digital image. Texture can also be created from satellite imagery without additional data. According to the method proposed by Mui et al., we created a standard deviation texture layer using a 3 × 3 pixel moving filter window 19 . Finally, all input layers were weighted equally in the segmentation process 19 .
Image segmentation. Segmentation is the process of partitioning an image into objects by grouping neighbouring pixels with common values. The objects in the image ideally correspond to real-world features. Effective segmentation ensures that classification results are more accurate. An edge-based segmentation algorithm, which computes a gradient map based on a user defined scale level, was used in this study. The scale level was chosen so that adjacent pixels with similar characteristics could be grouped. Since the aim of the study was to classify the wetland into several levels of cover types, a scale level of 25% was chosen after several iterations. We decided not to merge the images after segmentation to avoid any loss of information. Texture attributes were computed for each kernel, and 3 was selected as the default value. Finally, watershed transformation was applied to the modified gradient map to produce the final segmentation results 39 .
Image classification. Training samples were selected using high spatial resolution satellite imagery and ground truth data obtained through field surveys. A minimum of 30 training samples were chosen for each class 19 . Approximately 20 samples were directly imported from field datasets and approximately 10 samples were added by interpretation. Different classification methods have their own merits and can achieve different classification  results. While the nearest neighbour, or k-NN approach, is among the simplest of all machine learning algorithms, it performs as well as more complicated classifiers such as support vector machines (SVM) or neural net. In this study a nearest neighbour classifier (non-parametric classifier) was selected to classify images according to the defined classification system. The k-NN approach is a very intuitive method that classifies unlabelled examples based on their similarity to examples in the training set. The separation distance between classes was calculated based on the mean feature values of the pixels in each object (calculated from the input layers). The k-NN process involved selecting training samples, comparing sample attributes, and refining training samples until a satisfactory result was achieved 19 . After the k-NN method runs, each segment is assigned the class with the highest-class confidence value. This process is the most common approach for change detection and has been successfully used in several land-use changes studies 34,40 . The post-classification approach provides "from-to" change information, and the type of landscape transformations that have occurred can be easily calculated and mapped. A change detection map of "from-to" change information was derived for each of the five classification maps.
Accuracy assessment. An evaluation of the quality of a map derived from remote sensing data is important not only for verifying the quality of the map and its fitness for a specific purpose but also for understanding The most popular method includes the confusion matrix, used for maps derived from most types of classification methods 10 . In this study, overall accuracy and kappa coefficient were used for accuracy assessment 19 . The kappa coefficient is a measure of chance-corrected agreement between the actual land-cover classes and the classified land-cover classes in remote sensing. In this study, the kappa coefficient was calculated using the following equation:

Results
Accuracy assessment of the map. A quantitative accuracy assessment was conducted to evaluate the accuracies of the maps. A total of 60 random samples were selected for the accuracy assessment. Sample selection was based upon very high resolution aerial photographs and Google Earth imagery. The samples were divided into verification data and training data. Classification and change maps. Classification maps were generated for all five years (Fig. 2), and the individual class areas and change statistics for the five years are summarized in Fig. 3. From 2000 to 2013, the building land-cover type increased the most in terms of area (by approximately 4.84 km 2 ), and the herbaceous land-cover type increased by approximately 2.25 km 2 , while the pond land-cover type decreased the most in terms of area (by approximately 4.3 km 2 ), followed by forest (decreased by approximately 1.42 km 2 ), river (decreased by approximately 0.66 km 2 ), and marsh (decreased by approximately 0.61 km 2 ). In terms of the percent change in area, the non-wetland (urban) land-cover type increased by approximately 100% from 2000 to 2013, while wetlands (vegetation and water) decreased by approximately 20%, and pond decreased the most (by approximately 50%).
To    (approximately 0.77 km 2 ) were the main land-cover types that changed to non-wetland areas. The greatest increase in non-wetland area occurred from 2007 to 2013. Based on the matrices of the LULC changes, herbaceous, forest, and pond are the areas most likely to be converted to new buildings and roads. In addition, we found that the main change areas are distributed around the Xixi National Wetland Park because the areas in the park are strictly protected by law 36 . To analyse how many wetlands were converted to the non-wetland land-cover type, we calculated the specific change types shown in Table 5. The results show that herbaceous (32.22%), forest (29.57%) and pond (23.85%) are the main land-cover types that changed to non-wetland, followed by cropland (6.97%), marsh (4.04%) and river (3.35%).
Analysis of change patterns. Although similar statistics could be generated for other units such as county, township, or census tract, etc., the above change statistics shed minimal light on the question of where LULC changes are occurring. However, by constructing a change detection map (Fig. 4), the advantages of satellite remote sensing in spatially disaggregating change statistics can be more fully appreciated. Figure 4 Fig. 4, the increase in urban areas is mainly distributed west and southeast of the study area, and the greatest increase occurred from 2007 to 2013. In summary, information from satellite remote sensing can play a significant role in quantifying and understanding the nature of changes in LULC and where the changes are occurring. This information is essential to planning for urban growth and development.

Discussion and Conclusions
The LULC changes in the Xixi Wetland are affected by many factors including natural and human factors. Natural factors, which often operate on large spatial and temporal scales, combine with landscape effects to control changes in wetlands at a large environmental scale. However, human-related factors operate at short time scales and are important driving forces that affect changes in wetlands, especially in urban wetlands. In order to protect urban wetlands, the Xixi National Wetland Park was constructed in 2005 and the local government established several laws to protect wetlands in the park 35 . However, these laws only protect the wetlands inside the park, and the wetlands outside the park are not protected. The local economy developed rapidly because of tourist Cropland to non-wetland (urban) 0.50 6.97 Table 5. Specific change types from 2000 to 2013. attractions, especially after 2007 41 . Therefore, while the region inside of the protected area remains stable, the adjacent area declined rapidly. As shown in Fig. 4, most of the wetland changes occurred west and southeast of the study area, and patterns emerged that highlight urbanization activity (e.g., real estate exploitation and population growth). The Xixi Wetland is a primary example of the nature-based tourism attractions in the southern part of China, and it is the first national urban wetland area in China to receive five A grades for tourist attractions. Because of the comfortable environment, many real estate enterprises are investing in the area and developing and building houses around the Xixi National Wetland Park 41 . Many houses have even been built directly on herbaceous, forest, and pond land-cover types. Many rivers and ponds have been filled, and landscape connections have been altered by the buildings and roads found throughout the wetlands. Therefore, large areas of urban wetlands have been replaced by built-up land due to urbanization 42 . High spatial resolution satellite images covering several years and an object-based classification method were used to interpret the study area, classify different types of wetlands and analyse spatial and subordinate relationships based on the reference data. This approach guarantees the consistency of the wetlands classification system and allows for a rapid acquisition of historical data for the wetlands. Therefore, this method plays an important role in analysing the variation rules and driving mechanisms of urban wetlands. However, there are some uncertainties and errors in the change detection results, which are mainly due to variations in the data and the classification accuracy. For example, the remote sensing images were obtained over five years, from 2000 to 2013, and between January and December. Next, some of the land-cover types such as forest and herbaceous are easily confused in terms of their spectral characteristics, which can affect the classification. Moreover, training samples of every LULC type will affect the classification accuracy.