2019 6th NAFOSTED Conference on Information and Computer Science (NICS) Mapping land cover types in Vientiane, Laos using multi-temporal composite Landsat images Sanya Praseuth Economic, Technology and Environment Committee, Laos National Assembly Vientiane, Laos sanyapraseuth62@gmail.com Hung Bui Quang Center of Multidisciplinary Integrated Technology for Field Monitoring, University of Engineering and Technology Hanoi, Vietnam hungbq@fimo.edu.vn Dung Pham Tuan Center of Multidisciplinary Integrated Technology for Field Monitoring, University of Engineering and Technology Hanoi, Vietnam dungpt@fimo.edu.vn Chuc Man Duc Center of Multidisciplinary Integrated Technology for Field Monitoring, University of Engineering and Technology Hanoi, Vietnam chucmd@fimo.edu.vn Thanh Nguyen Thi Nhat Center of Multidisciplinary Integrated Technology for Field Monitoring, University of Engineering and Technology Hanoi, Vietnam thanhntn@fimo.edu.vn Abstract- Land-cover mapping is now effortless due to the availability of multi-satellite imagery such as Landsat data However, the global-scale land-cover classification method performs a low accuracy at the local scale, especially in developing countries, because of lacking ground truth data This research tries to build land-cover maps in Vientiane Capital, Lao PDR using multi-temporal composite Landsat images In our study, the combination of multi-temporal cloudless data with a strong classifier (XGBoost) gave an overall accuracy of 75.13% Keywords: land cover types, multi-temporal composite image, Vientiane, Landsat As a result, up-to-date land-cover information will not be available in time, causing difficulties in the management and planning needed by management agencies Therefore, it is necessary to develop a rapid, accurate method for updating the current status of the land covers to meet with the requirements The objective of this research is to classify land cover types in Vientiane Capital, Lao PDR using multi-temporal composite Landsat images, based on cloud-free composite images for classification year and a state-of-art classifier (XGBoost) II STUDY AREA AND DATA I INTRODUCTION Nowadays, Laos is in a stage of rapid economic development Therefore, the urbanization process in Lao PDR is increasing, although a traditionally low base Towns and cities are becoming the engines of nationwide growth, like other countries in the region [1] This process led to a rapid change in land covers, i.e urban areas, agricultural areas, particularly in the Vientiane capital As these trends continue, urban planning and land management policies will become ever more important mechanisms to guide development and help protect communities, the environment and cultural resources Meanwhile, the latest land-cover maps for the country is for 2014 and backward only With the current land-cover mapping technology, the technician will determine land covers on satellite imagery by using human eyes and prior knowledge over the mapping regions The process is also aided by using some GIS software support such as ENVI, ArcGIS This makes the mapping time last long, up to 5-10 years depending on mapping areas, human resources and budget There are few studies in researching land-cover classification methodologies in Laos For example, the study on Lao’ urban land management [1], using a time series of Landsat and MODIS data and landscape metrics to delineate the dynamics of shifting cultivation landscapes in Northern Lao PDR Between 2000 and 2009 [2], investigating urbanization in Vientiane Capital to address the relationship between urbanization and land use [3], monitoring of land use and land cover change Vientiane area using Landsat images [4] A Study area In this study, Vientiane Capital was selected as the study area (Fig 1) Vientiane Capital, which was built in the 16th century, is located in the center of the country, now covers an area of 3,920 square kilometers The topography of Vientiane is mainly a mix of mountainous and small-scaled flatten areas Elevation ranges from 70 m to 950 m This area has dry and rainy seasons The dry season is from October to March, the rainy season is from April to September of the year afterward The average temperature is 25oC Rainfall is from 1,300mm – 2,100mm annual In terms of land covers, in 2014, Vientiane Capital has eight main types: agriculture, forest, road, industrial, rice, bare land, urban, and water, as shown in Table I From the actual demand and the present status of the land-cover information, in this study, eight land-cover types were identified: agriculture, forest, grass/shrub, industrial, rice, bare land/soil, urban, and water Figure Vientiane Capital, Lao PDR 978-1-7281-5163-2/19/$31.00 ©2019 IEEE 563 2019 6th NAFOSTED Conference on Information and Computer Science (NICS) SUMMARY OF LAND COVERS AND THEIR AREAS AS PER LAO GOVERNMENT STATISTICS IN 2014 TABLE I Agriculture (ha) District Chanthabouly Sikhotabong 588 Forest (ha) Road (ha) Industrial (ha) Rice (ha) Bare land (ha) Urban (ha) Water (ha) Total (ha) 56 160 98 400 272 1,080 184 2,250 1,264 377 2,103 5,767 3,782 1,785 15,666 4,253 3,534 3,620 739 14,503 93 569 294 1,610 518 3,096 Xaysettha 314 2,043 Sisatanak 12 Naxaythong 4,603 35,782 373 7,750 36,457 1,586 4,250 90,801 Xaythany 14,316 15,439 912 14,323 29,618 6,794 3,283 84,685 Hatsayfong 8,760 101 354 806 6,100 4,473 1,448 3,756 25,798 Sangthong 1,168 28,329 891 8,855 38,542 75 2,138 79,998 Pakngeum 10,211 29,176 59 11,202 1,705 1,111 2,536 56,000 Total 39,658 110,147 3,897 55,555 120,662 21,106 19,189 372,797 2,583 Figure Ground truth points (yellow marker) displayed on Google Earth Figure Landsat footprints over the study area TABLE II SUMMARY OF LANDSAT SR IMAGES Year Number of images 2015 2016 2017 Total 71 70 61 202 TABLE III B Image data Landsat Surface Reflectance (Landsat SR) data was used as the main data in producing the up-to-date land-cover map in this study Landsat SR images are produced from Landsat OLI top of the atmosphere (TOA) images by using LaSRC algorithm [5] The Landsat SR images remove the effect of the atmosphere on the reflection of the surface object, thereby providing a more accurate signal Due to the large area of the study site, there are four different Landsat path/rows (images) (Fig 2) in the study area All of the images should be collected before developing the algorithm Therefore, the number of images that need to be processed is large In addition, in order to ensure adequate data for subsequent image synthesis, all Landsat images taken in the study area from 2015 to 2017 (2017 is the target year for producing the up-to-date land-cover map) were collected Images that are fully covered by clouds will be discarded Agriculture Forest Grass/Shrub Industrial Rice Bare land/soil Urban Water Total SUMMARY OF TRAINING AND TESTING DATA Training data 750 1275 907 575 863 486 1053 418 6327 Testing data 71 151 96 141 14 49 16 547 C Training and testing data To build and validate the land-covers classifier, training data and testing data are collected To collect training data, the Lao’s government 2014 land-cover data was used to generate randomized points using a stratified random sampling method The points are then collated with highresolution imagery from Google Earth to give the corresponding label For testing data, this data was collected through a field trip in the study area (Fig 3) Table III provides information on the number of training and testing points for each type of land cover Table II provides information about the number of images each year 564 2019 6th NAFOSTED Conference on Information and Computer Science (NICS) Figure Official land-cover maps for 2014, from Lao government D Ancillary data In this study, the official land-cover map for 2014 (Fig 4) is used for two main purposes: (1) support for collection of training data and (2) for validation of proposed land-cover classifier III METHODOLOGY The proposed land-cover classification method is presented in Fig This method includes parts: (1) generation of cloud-free composite images for classification year, (2) image stacking and feature extraction, (3) classification using XGBoost classifier, (4) validation of the classifier A Generation of composite images The purpose of this step is to create a cloudless image sequence of the classification year This is to capture the major spectral variations of land-cover types in the study area Composite images are essentially a raster image which has the same specifications as original Landsat SR images But the point is that composite images have no or very little cloudy pixels thus significantly improve classification accuracy To create a composite image, the timing of the image creation (target date) needs to be specified at the very first step In this study, six target dates were selected as 15/01/2017 (Date of year - DOY 15), 16/03/2017 (DOY 75), 15/05/2017 (DOY 135), 14/07/2017 (DOY 195), 12/09/2017 (DOY 255), 11/11/2017 (DOY 315) Hereafter, composite images are called target images Ancillary images, used to create a composite image, are called candidate images For each pixel of a target image, its values will be selected among pixels of the same location from candidate images The “best” pixel among candidate pixels is selected by following a set of strict rules Here, the method for the creation of composite images is followed by in the previous study [6][7][8] There are scores to consider: year, DOY, opacity, distance to cloud/cloud shadow Those scores can be categorized into image-level scores and pixel-level scores Year and DOY are image-level scores This means that all pixels of a particular candidate image has the same value Opacity and distance to cloud/cloud shadow are pixel-level scores In this case, each pixel has its own value depending on its situation The year score is inversely proportional to the distance from the target year (2017) to the year of the candidate image (2016, 2015, 2014) Next, the DOY score measures the proximity between the target Figure The proposed land-cover classification method image and the candidate image, which is modeled using the Gaussian function A DOY score is calculated as follows: ScoreDOY = = ( ) (1) √ Where σ is the DOY standard deviation, calculated from the DOY value of all candidate images μ is DOY value of the target image (i.e 15, 75, 135, 195, 255, 315) xi is the DOY value of the ith candidate image DOYs are only concerned with the day-to-day difference between the target image and candidate images, regardless of the difference in years between the two For example, assuming two candidate images have the same DOY distance to the target image, then DOY scores of these two images will be the same, even if the images are taken in different years The higher the DOY score, the more likely the pixel of the image is selected The distance to the cloud/cloud shadow of a pixel is another score to consider Sigmoid function is used to calculate this score The cloud/cloud shadow map provided with the Landsat SR image from the manufacturer (USGS) is used to calculate this score [9] Specifically: ScoreCloud/Shadow_Distance = ( ∗( , (2) )) Where Di is the closest distance to the cloud/cloud shadow of the pixel in the candidate image under consideration Dreq is a predefined maximum distance of impact, here it is defined as 50 pixels The meaning is that the cloud/cloud shadow only has impacts on the surrounding area of within 50 pixels (~1500m) Dmin is a predefined minimum distance, here is The closer the pixel to cloud /cloud shadow pixels, the lower the score The opacity score needs an aerosol-level image as input [8] The Landsat SR provides only discrete aerosol information which has four levels including high aerosol content, average aerosol content, low aerosol content, and climatology-level aerosol content Therefore, each aerosol level will be assigned a score by employing a sigmoid function The result is points corresponding to levels of aerosol content 565 2019 6th NAFOSTED Conference on Information and Computer Science (NICS) TABLE IV Year Opacity Distance to cloud/ cloud shadow DOY Score Description 1.00 2016 0.68 2015 0.42 2014 0.22 2013 In which, Ω( ) = + λ|| || is a regularization term T is tree leaf number, w is leaf weight, and λ are predefined hyper-parameters Having defined the objective function, users can define different cost functions as needed Furthermore, users can also define base learners (i.e decision trees) SUMMARY OF FOUR SCORES USED IN COMPOSITION METHOD Score The constraint is ± 30 days from the target day Scoring by a Gaussian function 0.023 0.223 0.777 0.977 In order to optimize hyper-parameters for the classifier, a 10-fold cross-validation technique on the training data set is used The best cross-validation hyper-parameters were then used to train the classifier over the entire training set After the classification model was trained, the feature dataset was used to build a land-cover map The testing set and the official land-cover map then used to validate the final map Description High aerosol content Average aerosol content Low aerosol content Climatologylevel aerosol content The constraint is a 50-pixel radius from the considered pixel Finally, the total score of each pixel is the sum of year score, DOY score, distance to cloud/cloud shadow score and opacity score Candidate pixel owning the highest total score will be selected as a replacement in the target image Table IV summarizes the scores mention above B Feature extraction Land-cover types have different spectral characteristics Some land covers have more variable year-round fluctuations than others For example, seasonal vegetation (rice, short-term crops) is more volatile than evergreen forests Specifically, rice is a plant with a special spectral variation which is different from other land covers Spectral signal of rice varies considerably throughout the growth cycle, i.e from watering to ripening and harvesting Thus, accurate mapping of such land covers requires a significant amount of observations throughout the year In this study, after creating five composite images of the classification year, the images were stacked together into a single image Features will be extracted from this stacked image In this study, classification features are all of the spectral bands Thus, there will eventually be 35 features corresponding to 35 spectral bands C Classification method and evaluation For classification, the XGBoost classifier was used in this study [10] XGBoost is a new classifier and has been proven to work well on a variety of applications However, XGBoost has not yet been widely applied in land-cover classification XGBoost is based on the principle of Gradient Boosting Machines (GBM) with several improvements For example, it can be trained in parallel mode, scalable, less overfitted to data, and can work well on sparse data The XGBoost model can be expressed as the sum of base learners as follows: Φ(xi) = ( ), ∈ (3) In which, F is the function space of the base learners, xi is the input data vector, Φ is model function In order to build basic learners, an objective function is needed For XGBoost, the objective function is defined by the following formula: ( )=∑ ( ′, ) + ∑ Ω( ) (4) Overall accuracy (OA), precision, recall and F1 score (F1) are used as evaluation metrics in this study [11] [12] OA and kappa coefficient are computed for the classification level PA, UA, and F1 are class-specific The formulas of the metrics are presented below OA = Ncorrect / Ntotal Precision = NTcorrect / NTclassified Recall = NT_refcorrect / NT_ref F1 = ∗ (5) (6) (7) (8) In which: Ncorrect: number of correct classified points Ntotal: total number of points NTcorrect : number of correctly classified points in a given class NTclassified: number of classified points in a given class NT_refcorrect: number of correctly classified in reference data of a given class NT_ref: number of points in reference data in a given class Additionally, classification maps are validated against official land-cover data and visually examined IV EXPERIMENT RESULTS This section presents the composition and classification evaluations It consists of three main parts: results of the image composite method, classification evaluation based on ground truth points, and classification evaluation based on the official land-cover map A Result of image-composition method The RGB-false color image of the composite images is shown in Fig It can be observed that the images are not or little clouded Furthermore, each image shows surface changes at its scanned time in the study area Specifically, the composite image representing DOY 15 (15/01/2017) and DOY 195 (14/07/2017) showed that some rice-growing areas were in the watering stage The image of DOY 135 may represent the harvesting period For forest areas, urban areas, one could see some stability as compared to agricultural areas in the study area 566 2019 6th NAFOSTED Conference on Information and Computer Science (NICS) Figure Land-cover map for Vientiane capital in 2017 Figure Composite images TABLE V CONFUSION MATRIX Agriculture Forest Grass/Shrub Industrial Rice Bare land Urban Water Precision (%) Agriculture 41 16 59 Forest 120 19 1 88 Grass/Shrub 18 10 46 10 51 Industrial 0 1 38 Rice 1 129 0 90 Bare land 2 67 Urban 0 0 47 75 Water 0 0 0 16 89 Recall (%) 58 79 48 67 91 43 96 100 75.13 (OA) TABLE VI EVALUATION BASED ON OFFICIAL LAND-COVER MAP Agriculture Forest Grass/Shrub Industrial As land-cover map 241.28 1,774.38 624.40 As official land-cover map 335.95 2,034.25 120.905 Rice Bare land Urban Water 44.87 539 41.53 243.93 160.07 16.650 670.1 23.56 276.4 204.71 may be the result of the clearance of agricultural areas for urbanization and industrialization Comparative results have shown the effectiveness of the classification method and demonstrated the actual status and causes of variation of the major land covers in the study area B Validation against statistical data Fig shows the derived land-cover map for Vientiane Capital in 2017 It can be seen that forest, rice, and urban areas occupy most of the area Table VI compares land covers areas derived from the proposed method and the official land-cover map V CONCLUSION In this study, an annual land-cover classification method for the Vientiane Capital, Lao PDR is proposed This method consists of four parts: (1) generation of cloud-free composite images for classification year, (2) image stacking and feature extraction, (3) classification using XGBoost classifier, (4) validation of the classifier The initial results show some potentials of the method as compared to the traditional mapping technique used in the region The image composition method can produce a cloudless image from a It could be seen that areas of land-cover types have certain differences Agricultural, forest, rice, and water observed a decline in 2017 as compared to 2014 This is due to: (1) for the agricultural area, especially rice, it is the problem of abandoning paddy fields due to lack of human resource and some other causes in Vientiane Capital; (2) forest area is declining due to urbanization and industrialization, and even agriculturalization under the Lao’s government plan An increased area of grass/shrub 567 2019 6th NAFOSTED Conference on Information and Computer Science (NICS) set of cloudy images This facilitates the classification Combining multi-temporal cloudless data with a strong classifier (XGBoost) gave an overall accuracy of 75.13% Forest, water, and rice are best-classified land covers The image-derived land-cover map in 2017 showed time and spatial variations of land covers in the region as compared to the official land-cover map in 2014 However, classification accuracy should be improved in subsequent studies, i.e for better classification of bare land and industrial areas, in order to make a more accurate land-cover map ACKNOWLEDGMENT REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] T T Paul Rabé Vongdeuane Vongsiharath, “Study on Urban Land Management and Planning in Lao PDR Under No.10,” Urban L Rev Res., no 5, pp 327–330, 2007 K Hurni, C Hett, A Heinimann, P Messerli, and U Wiesmann, “Dynamics of Shifting Cultivation Landscapes in Northern Lao PDR Between 2000 and 2009 Based on an Analysis of MODIS Time Series and Landsat Images,” Hum Ecol., vol 41, no 1, pp 21–36, Feb 2013 K Okamoto, A Sharifi, and Y Chiba, Integrated Studies of Social and Natural Environmental Transition in Laos 2014 S W Hue, A Korom, Y W Seng, V Sihapanya, S Phimmavong, and M H Phua, “Land use and land cover change in Vientiane area, Laos PDR using object-oriented classification on multi-temporal Landsat data,” Adv Sci Lett., vol 23, no 11, pp 11340–11344, Nov 2017 U.S Geological Survey, “Landsat Surface Reflectance Code (LASRC) Product Guide (No LSDS-1368 Version 2.0) Retrieved from https://www.usgs.gov/media/files/landsat-8surface-reflectance-code-lasrc-product-guide,” no December, p 40, 2019 C D Man, T T Nguyen, H Q Bui, K Lasko, and T N T Nguyen, “Improvement of land-cover classification over frequently cloud-covered areas using Landsat time-series composites and an ensemble of supervised classifiers,” Int J Remote Sens., vol 39, no 4, pp 1243–1255, 2018 P Griffiths, S van der Linden, T Kuemmerle, and P Hostert, “A Pixel-Based Landsat Compositing Algorithm for Large Area Land Cover Mapping,” IEEE J Sel Top Appl Earth Obs Remote Sens., vol 6, no 5, pp 2088–2101, 2013 J C White et al., “Pixel-based image compositing for large-area dense time series applications and science,” Can J Remote Sens., vol 40, no 3, pp 192–212, 2014 Z Zhu, S Wang, and C E Woodcock, “Improvement and expansion of the Fmask algorithm: Cloud, cloud shadow, and snow detection for Landsats 4-7, 8, and Sentinel images,” Remote Sens Environ., vol 159, pp 269–277, 2015 T Chen and C Guestrin, “XGBoost: A Scalable Tree Boosting System,” Friuli Med., vol 19, no 6, Mar 2016 G Congalton and K Green, Assessing the Accuracy of Remotely Sensed Data: Principles and Practices, vol 25, no 130 2010 D M W Powers, “Evaluation: From Precision, Recall and FMeasure To Roc, Informedness, Markedness & Correlation,” J Mach Learn Technol ISSN, vol 2, no 1, pp 2229–3981, 2011 568 ... B Image data Landsat Surface Reflectance (Landsat SR) data was used as the main data in producing the up-to-date land- cover map in this study Landsat SR images are produced from Landsat OLI top... 1,4 48 3,756 25,7 98 Sangthong 1,1 68 28, 329 89 1 8, 855 38, 542 75 2,1 38 79,9 98 Pakngeum 10,211 29,176 59 11,202 1,705 1,111 2,536 56,000 Total 39,6 58 110,147 3 ,89 7 55,555 120,662 21,106 19, 189 372,797... 4 18 6327 Testing data 71 151 96 141 14 49 16 547 C Training and testing data To build and validate the land- covers classifier, training data and testing data are collected To collect training