2016 Eighth International Conference on Knowledge and Systems Engineering (KSE) Optimizing GLCNMO version method to detect Vietnam’s urban expansion Pham Tuan Dung1, Man Duc Chuc1, Nguyen Thi Nhat Thanh1, Bui Quang Hung1, Doan Minh Chung2 Center of Multidisciplinary Integrated Technology for Field Monitoring, University of Engineering and Technology, Vietnam National University, Hanoi, Vietnam Space Technology Institute, Vietnam Academy of Science and Technology, Hanoi, Vietnam dungpt@fimo.edu.vn In Vietnam, there are few researches in urban classification with limited scope, such as the relation between surface temperature and land cover types using thermal infrared remote sensing in Hochiminh city [4], study about land use change pattern in Danang city [5], optimizing spatial resolution of imagery for urban form detection [6], and assessing the impact of urbanization on urban climate by remote sensing perspective [7] Abstract—No global scale land cover classification method performs with high accuracy at local scale This study tries to develop a classification algorithm for urban area in Vietnam This is the first assessment of the Global Land Cover by National Mapping Organizations (GLCNMO) version method for producing global urban map in 2008 An improved and optimized algorithm is then developed based on the GLCNMO method for Vietnam taking into account of local natural and social conditions Improving method is then applied to produce urban maps of Vietnam for the years of 2008 and 2015 Accuracy assessment showed that the improved method can achieve up to 13% higher precision and 10% higher F1 measure as compared to the global GNCNMO method Also, an increasing trend was observed in population density in urban area in the period from 2008 to 2015 in Vietnam which may correspond to fast urbanization process in the country The cities also tend to become less green in 2015 than 2008 as indicated by comparing the Normalized Difference Vegetation Index (NDVI) between the two years This paper presents results from a research to create urban maps, which serves as the first stage in our development of a comprehensive database of urban land surface characteristics for Vietnam The primary goal of this paper is to provide urban maps at 500m spatial resolution, by combining gridded population density, nighttime lights and MODIS-NDVI data From those maps, physical extents of urban areas are detected II DATA Data used in this research are described in Table I Keywords—optimizing; GLCNMO version2; Vietnam’s urban map; urbanization; urban expansion I A High Resolution Population Distribution Maps for Vietnam in 2009 and 2015 Landsat images and land cover information were combined with other datasets to model population distributions for 2010 and 2015 for countries in the Southeast Asia region including Vietnam [8] Vietnam’s population distribution datasets for 2010 and 2015 were already generated at a fine-scale spatial resolution (100m) and projected to a geographic coordinate system and WGS 84 The data was freely downloaded from website http://www.worldpop.org.uk/ INTRODUCTION Despite of huge achievements in economic growth, Vietnam’s government has also implemented various long term policies in an attempt to boost its economy Urbanization is a necessary effect of economic and urban development, which is related to functional and spatial transformations and its form will have long-lasting consequences on the lives of urban residents [1] Because urbanization might cause many environmental problems, such as vegetation loss, air pollution, water shortage and contamination, and urban heat island, it has been recognized as an important factor affecting the functions of terrestrial ecosystems and climate change [2] This study aimed to answer questions about how urban population growth relates to urban spatial expansion; and the relationship between urbanization, energy consumption growth, and green areas reduction in Vietnam from 2008 to 2015 B Nighttime light data for Vietnam in 2008 and 2015 The Version Defense Meteorological Satellite Program Operational Linescan System (DMSP-OLS) nighttime light imagery is available in http://ngdc.noaa.gov/eog/dmsp/ For the present study, the stable night light data within 2008 DMSPOLS (F16 satellite) composite product was used Ephemeral detections of fires, gas flares, volcanoes or aurora in DMSPOLS nighttime imagery have already removed Also, the background noise has been subtracted The original spatial resolution of the products was 500m, and the DN (Digital number) values range from to 63 [9] Remote sensing is a useful source for mapping the expansion of urban land Recent coarse resolution urban mapping from satellite imagery are also unsatisfactory, because of the tedious process of training data collection and inadequacies of in classification algorithm [3] 978-1-4673-8929-7/16/$31.00 ©2016 IEEE 309 TABLE I REMOTE SENSING DATA USED IN THIS RESEARCH Abbreviation Data Description Spatial Resolution Worldpop Population density 100m 2008 2015 DMSP-OLS Stable nighttime light 1km 2008 NPP-VIIRS Suomi NPP nighttime light imagery 500m 2015 MOD13Q1 MODIS/Terra 250m Vegetation Indices 16Day L3 Global 2008 2015 EstISA Impervious surface area 1km 2010 MOD44W MODIS inland water mask 250m N/A Shuttle Radar Topography Mission Water Body Data (SWBD) in combination with MODIS 250m data to create a complete global map of surface water MOD44W data was downloaded at https://lpdaac.usgs.gov/data_access/ Time III METHODOLOGY A Urban definition Recently, in order to have an urban definition widely accepted by most of researchers is a challenge There are various ways to define an “urban area” in different studies and countries The United Nations itself recognizes the difficulty of defining urban areas globally, stating that, “because of national differences in the characteristics that distinguish urban from rural areas, the distinction between urban and rural population is not amenable to a single definition that would be applicable to all countries” [12] Urban classification is the identification and delimitation of urban areas and their assignment to classes Urban areas are heterogeneous mixtures of land cover types, and may contain any number of vegetated and man-made surfaces As defined by the Food and Agriculture Organization-United Nations Environment Programme (FAO-UNEF) Land Cover Classification System (LCCS) [13], urban areas are non-linear built up areas (i.e urban, industrial and other areas related to trade, manufacturing, distribution and commerce) covered by impervious structures adjacent to or connected by streets Linear elements like roads, railways and communication lines/pipelines occur but are not dominant features Impervious surfaces are dominated implies which coverage greater than or equal to 50% of a given landscape unit (here, the pixel) [14] Pixels that are predominantly vegetated (e.g a park) are not considered urban, even though in terms of land use, they may function as urban space Suomi National Polar-Orbiting Partnership - Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) nighttime light imagery was downloaded from the website http://ngdc.noaa.gov/eog/viirs/ Twelve month composites generated from the observations between 2015/1/1 and 2015/12/31 was used in this study C MODIS-NDVI data for Vietnam in 2008 and 2015 MODIS/Terra Vegetation Indices 16-Day L3 Global 250m SIN Grid images were downloaded from the NASA Land Processes Distributed Active Archive Center (http://earthexplorer.usgs.gov/) 23 composite images for each year in 2008 and 2015 were used in this study A Maximum Value Composition (MVC) was then applied to all NDVI images with the aim of selecting pixels less affected by clouds and other atmospheric perturbations [10] In GLCNMO version method, a definition of urban areas based on physical attributes: urban areas are places that are dominated by the built environment The “built environment” includes all non-vegetative, human-constructed elements, such as buildings, roads, runways [15] D Estimate the density of constructed Impervious Surface Area (EstISA) data for Vietnam in 2010 The estimate of ISA is derived solely from the brightness of satellite observed nighttime lights and population count The initial global ISA density grid was produced on a 30 arc second grid since that is the native grid of both the Landscan and nighttime lights This was then converted to a km equal area grid in a WGS 84 projection A threshold of 0.4 percent was applied to eliminate the salt and pepper noise present at the very low end of the ISA scale ISA values over 100 were reset to 100 Extractions of the digital values were run to tally the quantity of ISA for countries, sub-national units (states/ provinces) and major watersheds [11] The global grid of ISA at a resolution of 1km is freely available at: http://www.ngdc.noaa.gov/dmsp/ The Vietnam urban classification system, established in 2001 and updated in 2009 with the inception of Decree No 42/2009/ND-CP [1], serves as an important part of urban policy and management in Vietnam It is a hierarchical system constituted by six classes of urban centers that are defined by different levels of economic activities, physical development, population, population density, and infrastructure provision In 2009 there were special cities (Hanoi and Hochiminh city), class I cities, 12 class II cities, 40 class III towns, 47 class IV provincial towns, and 625 class V small townships The most important indicators are as follows: (i) Population of an urban center is at least 4000; (ii) The population density is at least 2000/km2 This study defines “urban area” as population density distribution, percentage of impervious surface, and nighttime light Green fields and water bodies (such as a big park or a golf course) are not considered as urban Minimum mapping unit of an urban area is km2 E Water body data for Vietnam The MODIS land-water mask at 250 meter spatial resolution (Short Name: MOD44W) is a new product using the 310 Fig Flowchart of urban mapping in 2008 boundaries were used to extract study area Clipping the study area was done by using the ArcMap’s Extract by Mask with the Vietnam’s shapefile set as an analysis mask TABLE II THRESHOLD VALUE OF POPULATION DENSITY, NIGHTTIME LIGHT, NDVI_MAX, AND ISA Processing step: for population density, DMSP-OLS, and EstISA datasets, high threshold values generate a small urban area, and vice versa Potential urban map is derived from population density data using threshold Because EstISA data is not available for 2008, the data for 2010 were used in this research DMSP-OLS, EstISA, and NDVI_MAX thresholds were used to exclude low NTL data, low ISA data, and green areas from potential urban map, respectively Finally, the MODIS land-water mask was used to exclude inland water bodies Thresholds per pixel Population density 500 Nighttime light 11 NDVI_MAX 0.69 Impervious surface areas B Sample selection methods In order to calculate thresholds, polygon samples were chosen and checked by comparing with Google Earth and Landsat ETM+ images The numbers of pixel samples of each class (except urban class) were decided by percentages of GLCNMO’s classes 1046 non-urban pixels and 540 urban pixels were chosen totally Urban class has higher priority than others in deciding thresholds Population density threshold was based on the Vietnam urban classification system Due to the lack of EstISA data for 2015, this data was not used to generate the urban map of 2015 as in Fig2 IV RESULTS A Accuracy assessment Accuracy assessment was performed independently for the urban class For each urban map, an equalized, random set of test points was selected The test points within a sample were further randomized to avoid bias in the reference labeling of those pixels This set of points was exported into a shape file and used to assess the accuracy of this method classification In this research, Google Earth and Landsat ETM+ images were used as independent sources of reference data of higher precision and known accuracy for validating the classifications GLCNMO 2008, urban maps of 2008 and 2015 (VN 2008, VN 2015) are assessed Precision and Recall measures are used to quantify map accuracy Also, F1 measure (the harmonic mean of precision and recall) is applied to provide additional information on the effectiveness of each urban map C Data processing The method divided in two main steps as shown in Fig1 Preprocessing step: the population distribution maps with 100 m spatial resolution was aggregated to generate proportional settlement values in a new data set with a pixel size of 500 m by 500 m DMSP-OLS and EstISA data at spatial resolution of km were resembled to spatial resolution of 500 m NDVI_MAX data was generated from MODIS-NDVI data of 23 periods in 2008 with the maximum algorithm and aggregated to match the same spatial resolution with MODIS and DMSP-OLS other data The Vietnam’s administrative 311 Fig Flowchart of urban mapping in 2015 Vietnam’s urban map 2008 Vietnam’s urban map 2015 Fig Vietnam’s urban maps R setaccuracies of three urban maps P1 including set In this section, P2 set Fig Testing sets 312 P3 set TABLE III MAPS’ ACCURACY ASSESSMENT GLCNMO 2008 VN 2008 VN 2015 Recall (%) 85.71 89.29 89.29 Precision (%) 57 70 86 F1 measure 68.47 78.48 87.61 Fig Population density trending Google Earth images Urban map 2008 Urban map 2015 Urban expansion 2008-2015 Hanoi Hochiminh city Fig Urban expansion in Hanoi and HCMC, 2008-2015 To assess recall, 140 urban pixels (R set) were randomly selected based on ground truth images (collected from field trips and Global Geo-Referenced Field Photo Library which allows downloading freely in http://www.eomf.ou.edu/photos/) very high resolution images from Google Earth (using historical imagery function) It should be noted that the pixels are not included in the training sets which are used to estimate algorithm’s thresholds GLCNMO 2008, VN 2008, VN 2015 maps are then compared to R set For precision, 100 urban pixels are randomly extracted from each urban map Similar observations are also applied for other indexes Also, this study adopted the water mask to assist the classification because of the natural distributions of lakes and rivers in urban areas in Vietnam By optimizing the thresholds, significant higher precision on urban map of 2008 could reach This is also reflected in 10.01% improvement of F1 measures between GLCNMO 2008 and VN 2008 It is observed that VN 2015 urban map has highest recall and precision of 89.29% and 86% respectively It should be noted that the thresholds of 2015 map are estimated from the same set of pixels as 2008 map Impervious surface data was not used in the production of 2015 urban map This change may help to explain the improved result, because NPP Nighttime light data has much higher spatial resolution than DMSP-OLS thus providing more detailed information than DMSP-OLS data This is resulted in three sets: GLCNMO 2008 (P1 set), VN 2008 (P2 set), and VN 2015 (P3 set) in which each set is independent from the other Distribution of pixels in R, P1, P2, P3 sets are provided in Fig Assessment results of urban maps are reported in Table III From the results, there is a small difference between recalls of the three urban maps, around 3.58% All three methods can reach to a recall of above 85% However, it is remarkable that the precision of GLCNMO 2008 map over Vietnam is very low, around 57%, which is 13% and 29% lower than precisions of VN 2008 and VN 2015 respectively This may be due to very rough thresholds of the indexes (NDVI, population density, nighttime light, and impervious surface) which are not suitable to social and natural conditions in Vietnam For example, Vietnam has a population of above 90 millions and a higher population density in urban regions As a result, a global threshold of population density applied on Vietnam is not suitable B Trends in population density and the greenness of urban area in Vietnam Population density Analyzing of urban pixels in the training data showed a significant improvement of population density in urban areas From Fig 5, it could be also observed that the highest increase belongs to regions that already have very high population density such as the core or center of cities Considering the population density is not normally distributed, Wilcoxon rank sum test over the data in 2008 and 2015 were performed The p-value is 0.0016 which indicates statistical 313 significant improvement of population density over urban regions in Vietnam REFERENCES [1] The greenness of urban areas in Vietnam The difference of NDVI values of pixels in the training data is not clearly seen as the population density data There is an increase of NDVI in some areas and decrease in other areas However, by considering some descriptive indicator of the data such as mean and standard deviation, small lower threshold of NDVI in 2015 (0.66) is obtained comparing to those in 2008 (0.69) This is also easy to understand that in most urban areas, tree density tend to become less dense over time especially in regions suffering from fast urbanization process [2] [3] [4] [5] C Vietnam’s urban expansion from 2008 to 2015 This study used satellite imagery and demographic data to measure expansion in urban areas of Vietnam between 2008 and 2015 Two urban maps for two time periods were calculated and analyzed as in Fig Because DMSP-OLS data is not available for 2015, NPP-VIIRS nighttime light images are alternatively used The urban mapping method for 2015 is described in Fig Threshold for NDVI_MAX, NPP-VIIRS nighttime light are 0.66, 1.1, respectively Vietnam’s urban expansion in period 2008-2015 is about 600 km2 Fig describes the urban extension of Vietnam’s two largest city: Hanoi and Hochiminh city [6] [7] [8] V CONCLUSION [9] The results presented here indicate that application of population distribution in conjunction with the yearly vegetation photosynthesis product from MODIS images, the nighttime data, and the estimate of ISA, which can be used as an effective tool to compare the expansion of human settlements The methodology is easily implemented since it relies only on remote sensing data Even though this methodology is not the substitution for ground monitoring and field analysis, its scale of applicability makes it particularly interesting for regional planning purposes and to increase public and political perception of the problems connected with unregulated development [10] [11] [12] [13] The Gridded population density data, nighttime lights and NDVI offers a useful perspective for monitoring the extent and level of urbanization accurately and regularly at normal scale Compared with GLCNMO method, new thresholds are appropriate to create Vietnam’s urban maps more accurate, simpler and less costly [14] [15] ACKNOWLEDGMENT The authors would like to thank the ARC 2015 research project “Research, transfer, develop the Global land cover datasets - GLCNMO and CERes Gaia system for collection, management, analysis, and sharing of geospatial data” for financial support 314 The World Bank (2011) Vietnam Urbanization Review: Technical Assistance Report, 263 Retrieved from http://documents.worldbank.org/curated/en/2011/11/15817674/vietnamurbanization-review-technical-assistance-report Foley, J A (2005) Global Consequences of Land Use Science, 309(5734), 570–574 http://doi.org/10.1126/science.1111772 Asimu, A K., & Ateishi, R T (n.d.) GLCNMO Global Urban Mapping, Validation and Comparison with Existing Global Urban Maps, 427–440 Van, T (2005) Relationship Between Surface Temperature and Land Cover Types Using Thermal Infrared Remote Sensing, in Case of HoChiMinh City The Sixteenth Workshop of OMISAR on the Application …, 1–3 An, T T., & Tuan, V A (2008) Application of Remote Sensing in Land Use Change Pattern in Da Nang City , Vietnam Tran, T D B., Puissant, A., Badariotti, D., & Weber, C (2011) Optimizing spatial resolution of imagery for urban form detection-the cases of France and Vietnam Remote Sensing, 3(10), 2128–2147 http://doi.org/10.3390/rs3102128 Hoang Khanh Linh, N., & Van Chuong, H (2015) Assessing the impact of urbanization on urban climate by remote satellite perspective: a case study in Danang city, Vietnam ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XL7/W3(May), 207–212 http://doi.org/10.5194/isprsarchives-XL-7-W3207-2015 Andrea E Gaughan, Forrest R Stevens, Catherine Linard, Peng Jia, A J T., Stevens, F R., Linard, C., Jia, P., & Tatem, A J (2013) High Resolution Population Distribution Maps for Southeast Asia in 2010 and 2015 PLoS ONE, 8(2) http://doi.org/10.1371/journal.pone.0055882 Xie, Y., Weng, Q., & Weng, A (2014) A comparative study of NPPVIIRS and DMSP-OLS nighttime light imagery for derivation of urban demographic metrics 3rd International Workshop on Earth Observation and Remote Sensing Applications, EORSA 2014 - Proceedings, 335– 339 http://doi.org/10.1109/EORSA.2014.6927907 Holben, B N (1986) Characteristics of maximum-value composite images from temporal AVHRR data International Journal of Remote Sensing, 7(11), 1417–1434 http://doi.org/10.1080/01431168608948945 Elvidge, C D., Tuttle, B T., Sutton, P C., Baugh, K E., Howard, A T., Milesi, C., … Nemani, R (2007) Global Distribution and Density of Constructed Impervious Surfaces Sensors, 7, 1962–1979 http://doi.org/10.3390/s7091962 United Nations (2008) Principles and recommendations for population and housing censuses - Revision Statistical Papers (Vol 3) Gregorio, A Di (2005) Land Cover Classification System (LCCS), version 2: Classification concepts and user manual FAO Schneider, A., Friedl, M a., & Potere, D (2010) Mapping global urban areas using MODIS 500-m data: New methods and datasets based on “urban ecoregions.” Remote Sensing of Environment, 114(8), 1733– 1746 http://doi.org/10.1016/j.rse.2010.03.003 Phong, D X., Hoan, N T., Kobayashi, T., Tateishi, R., & Cover, L (2012) A GLOBAL 500-M URBAN MAP FOR GLCNMO VERSION 2, (Glcnmo 2013) ... ASSESSMENT GLCNMO 20 08 VN 20 08 VN 20 15 Recall (%) 85.71 89 .29 89 .29 Precision (%) 57 70 86 F1 measure 68.47 78.48 87.61 Fig Population density trending Google Earth images Urban map 20 08 Urban map 20 15... images Urban map 20 08 Urban map 20 15 Urban expansion 20 08 -20 15 Hanoi Hochiminh city Fig Urban expansion in Hanoi and HCMC, 20 08 -20 15 To assess recall, 140 urban pixels (R set) were randomly selected... mapping in 20 15 Vietnam’s urban map 20 08 Vietnam’s urban map 20 15 Fig Vietnam’s urban maps R setaccuracies of three urban maps P1 including set In this section, P2 set Fig Testing sets 3 12 P3 set