Comparison between Land Use/Land Cover Mapping Through Landsat and Google Earth Imagery

Transcription

1 American-Eurasian J. Agric. & Environ. Sci., 13 (6): , 2013 ISSN IDOSI Publications, 2013 DOI: /idosi.aejaes Comparison between Land Use/Land Cover Mapping Through Landsat and Google Earth Imagery 1 2 Shirkou Jaafari and Aliakbar Nazarisamani 1 Department of Environment, Faculty of Natural Resources, University of Tehran, Karaj, Iran 2 Natural Resource Department, University of Tehran, Karaj, Iran Abstract: It is proved that land use/land cover mapping (LULC) in scales such as urban district through high spatial resolution datasets is too expensive for many pilot projects mainly due to the cost of purchasing raw satellite images and deciphering the LULC types with remote sensing techniques. Since the launching of Google Earth (GE) on June, 2005, its potential has been approved to be used in image processing and dissemination of scientific information. Images from GE with high spatial resolution are free for public and can be used directly in LULC mapping in small geographical extend like mapping of green areas of cities. This study explores the feasibility of mapping of green areas in Karaj city through images from GE and to constitute a comparative basis the results were compared with green area map produced by means of Landsat TM image. Accordingly, both images were classified using image processing techniques and final accuracy of both classified images were identified and compared by field investigation. The accuracy assessments show overall accuracy of 0.93 and Kappa coefficient of 0.87 for GE based map and also overall accuracy of 0.77 and Kappa coefficient of 0.75 belong to map derived from Landsat TM imagery. Key words: Google Earth Land use maps Landsat images Accuracy assessment Karaj INTRODUCTION practices in the past data with medium spatial resolution like Landsat imagery were used to create LULC maps [3-5]. Humans impact the environment through various In current, researchers tend to use high spatial land uses to meet different needs of living [1]. In the past, resolution data in order to obtain more accurate and information and data related to LULC mapping were precise result. In this regard, images with high spatial obtained via field investigation in conjunction with aerial resolution from GE that are free to the public are a good photography interpretation which consume time and source of imagery including satellite images and air budget extensively. But, since the time the first satellite of photos. GE ( provided by Google Landsat series in 1972 has been launched, mapping Inc., is a virtual globe programming that maps the earth by process of LULC has become more convenient and more superimposition of high resolution satellite images. This accurate. Accordingly, in last decade with the advent of software has free version and is user-friendly and has satellites with high spatial resolution the use of satellite thus become a popular data viewer in a variety of fields imagery in fine scales such as urban planning has been including geosciences. Consequently, some barriers such significantly increased while it was not possible in the as sampling of training points in spatial and temporal unit past [2]. have been removed and have led to in improvement and LULC mapping through remote sensing and satellite development of mapping algorithms and accuracy imagery is an essential part of urban planning and assessment [6, 7]. management in order to comprehend different effects of Since the time that GE software has been released in polices and decision made by authorities. Furthermore, June, 2005 its potential in image processing and urban plans can be monitored and their consequences can dissemination of scientific information has been be predicted. Regarding to the cost of purchasing of raw recognized [8, 9]. However, GE images are capable to be satellite images with high spatial resolution, most processed in LULC mapping and change detection Corresponding Author: Shirkou Jaafari, Department of Environment, Faculty of Natural Resources, University of Tehran, Karaj, Iran. 763

2 studies. Images from GE are now available with high spatial resolution (>2.5m). Main classes of LULC maps like human dwellings, industrial units and local roads are clearly recognizable in GE [10]. Although, the number of users has been increasing, its potential as a database of high spatial resolution images is still unknown and needs more investigation in different parts of the world to reveal how much it is credible and capable. Digital images from GE are processed through a series of stages such as color balancing and warping to produce the final mosaic for the entire area. However, because the image from GE only contains information from three visible bands, its band information is rather coarse and limited compared to the commonly used Landsat and other satellite images. Images are without metadata, acquisition date and spectral transformation or spatial interpolation. Most of the digital images of GE are from Quick Bird and Digital Globe satellites. Besides, all images are registered in WGS84 coordinate system. This new source of information can be integrated with other sources like expert opinion and additional images in order to constitute a basis for classification and accuracy assessment of LULC maps in regional and global scales [6, 7, 11]. Up to now, few studies have been conducted to explore the potential and feasibility of GE images in LULC mapping. This study investigates the possibility of using the image from GE to derive urban green area and urban forest maps. Research Background: Literature review among studies in recent years has been done to explore the usage of GE images in spatial and geosciences practices. A brief overview of studies is given in which GE image takes an important part in these studies. Clarke et al. [12] has proposed a scalable approach to mapping annual land cover at 250 m using MODIS time series data in Dry Chaco ecoregion of South America. In this study, by using source data and processing techniques over more extended spatial and temporal units, a classification scheme with compatible accuracy were conducted. Digital images with high spatial resolution were derived from GE and thorough visual interpretation LULC map with 7 classes was produced. In accuracy assessment, most of the classes were assessed by images from GE (0.83 of reference sample). This method resulted in precise statistical sampling without temporal and spatial as well as inaccessibility limits over the entire area. Yang et al. [13] investigated the feasibility of using the image from GE to derive a high-resolution map on rural population distribution for a town in the Lake Tai basin of eastern China. The result showed that in spite of limited band numbers and information only from visible bands, images are capable enough to extract human dwellings with less depended techniques to spectral information such as texture analysis. In this case, since it not only utilizes the spectral information but also takes into account the spatial configuration of pixels which can clearly be seen in GE images. This study represented image from GE is an appropriate alternative for those expensive raw satellite images to purchase with equal spatial resolution. Aguirre-Gutiérrez et al. [14] used combined pixelbased and object-based approach in a mountainous area in Mexico via Landsat ETM+þ images and seven common land cover categories were recognized. Authors utilized GE images for geometric correction and accuracy assessment. In addition, the location of 50 training samples were captured using GPS device, extra training samples were verified from high resolution imagery available in GE (150 in total) dated between 2007 and This study depicted that more advanced and combined techniques of classification can be applied to more efficiently extract information from traditional remotesensing sources. Poetre [15] compared horizontal positional accuracy of GE images with Landsat images by using 436 control points in 109 cities from all around the world. It means in order to horizontal positional accuracy of GE images be identified, the location of 436 control points located in 109 cities were compared to Landsat GeoCover scenes [16, 17]. Landsat GeoCover is an orthorectified product with known absolute positional accuracy of less than 50 meters root-mean-squared error (RMSE). Relative to Landsat GeoCover, the 436 GE control points have a positional accuracy of 39.7 meters RMSE. Control points accuracy in developed countries with 24.1 meter RMSE is significantly more precise than developing countries with 44.4 meter RMSE. Dorais and Cardille [18] used MODIS and GE images to create annual maps by purpose of estimating the possibility of deforestation in hot spots of Boreneo island for time period between 2000 and GE images were used as ground truth for accuracy assessment of classified MODIS images. In this case, visual interpretation of GE images was implemented. Specified points in map were agreed with the deforestation points in reality. In the line with the study aim, the results of this practice helped for continuous monitoring as well as identification of those parts of forest with high probability 764

3 of deterioration [19, 20]. Results showed combined and Then, motioned layer was imported into GE and images compatible data derived from free sources can date pushed back to the closest year coincided with implemented more efficiently than be used separately. Landsat satellite image (26/5/2011) through show historical imagery tool bar. Despite the fact that there is MATERIALS AND MEDTHODS no red band in GE images, users can easily discern any type of LULC and other visual interpretation factors [23]. Study Area: Karaj city is the capital of Alborz province, Visual interpretation method was applied for image spanning between latitudes N and classification mainly because of high spatial resolution longitudes E and extended over total area capability. In this regard, green area polygons of three 2 of 2255 km (Fig. 1). It is characterized by Alborz Chain districts of Karaj city were digitized on images from GE. Mountains in the north. The elevation is descending After that polygons were exported in KML format and in from north to south. Average height is 1320 meter above the next step they converted into shape file using ArcGIS the sea surface. Dominant wind direction is toward North 9.3. Finally, shape files of green area of Karaj city were West and annual rainfall is mm as well as C geometrically corrected in UTM WGS 84 system (Fig. 2). annual average temperature. Total population of Karaj city In order to constitute a comparative basis to find out is 1,666,674 and almost 96.2% of the whole population how feasible are maps derived from GE, Landsat image lives in urban areas [21]. was also classified. The TM image dated to 29/5/2011 which is so close to date of GE image was implemented. Data Input Preparation: In this study, Landsat TM image Using Topographic map, TM image co-registered with an to map green area land cover of the year 2011 plus acceptable RMSE by means of nearest neighborhood topographic map which was obtained from National resampling type. Supervised classification with training Cartographic Center (NCC) of Iran for geometric sites was applied to Landsat TM image through Maximum correction were applied. Sampling points were selected by Likelihood (ML) algorithm. ML classification is means of GPS. All calculations and image processing were acknowledged as one of the most efficient parametric performed using Erdas Imagine v.9.1 and also ArcGIS 9.3 methods for image classification. The ML classifier takes as well as GE v.5. First of all, vector layer of targeted area into account the variance and the covariance of the class were geometrically corrected by the source in UTM WGS depending on its feature characteristics. In the line with 84 system and saved in KML format. It should be noted the study aim, a binary map was created in which green that not only GE shows images date, it also represents are of the study area encoded 1 and others encoded 2 quick movement among the images date [22]. respectively (Fig. 2). Fig. 1: Illustrates the study area across Alborz province, Iran. 765

4 Fig. 2: Green area maps of Karaj city, GE based map (right), Landsat imagery based map (left). Table 1: Accuracy assessment results. GE based map Landsat imagery based map Overall accuracy Kappa coeficient RESULTS In order to assess the accuracy of green area map, meter grid was overlaid on output map and those points that located on green area were determined by field investigation and using GPS. The total number of points was 70 in which 50 points belonged to green area land use and 20 points were assigned to other land uses. Finally, error matrix was constituted and overall accuracy and Kappa index were calculated using Erdas Imagine 9.1 (Table 1). According to output maps through GE imagery in comparison with Landsat images are more accurate refer to Kappa coefficient and overall accuracy. It can be concluded that in practices with fine scales like mapping of green area and other urban LULC images from GE are more capable to produce more reliable and valid data. DISCUSSION In this study, visual approach was implemented due to the existence of high spatial resolution data for image classification and results show that the out put map can be used even as ground truth. In the same practice conducted by authors [18] annual map of deforestation probability was created in which GE images were used as ground truth for classification and accuracy assessment of MODIS Images as well as visual interpretation. Finally, they concluded that GE contains acceptable level of spatial resolution and it can be implemented as an assistant in accuracy assessment. As mentioned before, GE images are geometrically corrected with acceptable RMSE (39.7 m) compared with Landsat images with greater RMSE (50 m). This benefit denotes that these image are quiet reliable and appropriate in urban LULC mapping in different study areas. Author [15] also used these kinds of digital images in his study which is in accord with our results. It should be noted that there is no enough information on GE data base validity. Because in some cases it can be seen that images are not appropriately merged in final mosaic. It can lead to confusion in geometrically correction of images. With refer to author [15] this problem can be removed through the controlling of transportation network which can be inserted in the GE image in KML format. Because of the small study area, this error was not seen over the targeted area in this practice. Moreover, there is another error which is called interpretation error and it is varied among interpreters [24, 25]. It is mainly dependent on difference in class percent cover and level of users training as well as mistakes in data recording. In the best case, locations of points for accuracy assessment are determined through GPS. Authors [12] in a same practice benefited from GE imagery for data collecting. In this case, visual interpretation was implemented with the aim of image classifying with seven categories. Final overall accuracy of 0.83 was obtained in this study which is less than the accuracy acquired in current practice and it was related to lack of possibility on distinguish between natural grass cover and agriculture fields over the digital images. Authors [12] not only considered the high spatial resolution of images but also paid attention to accessibility of study area, date and quality of image in addition to final mosaic of images. They believed that use of GE provides scalability to implemented method that its high-resolution imagery is 766

5 free to access, easy to navigate, is distributed across the 7. Helmer, E.H., M.A. Lefsky and D.A. Roberts, region and can be interpreted with a consistent set of rules. With regard to disadvantages of using GE images as a reference data, classification process should start with quite general categories which can result in better classification of LULC map. If the two class labels disagreed, then the technician interpretations were discarded and an expert interpreter estimated the sample's final percent cover and then the majority-class label came from the expert's estimation. GE imagery is a great assistant to validate different studies with remotely sensed data [26, 27]. Finally, with regard to increasing desire among researchers for applying of GE imagery plus the capability of high spatial resolution as well as the using freely, it could be considered as a reliable and valid source of data in LULC mapping and similar studies. REFERENCES 1. Lausch, A. and F. Herzog, Applicability of landscape metrics for the monitoring of landscape change: issues of scale, resolution and interpretability. Ecol. Indic., 2: Khoram, S., J. Gregory, D.F. Stalhng and H. Cakhr, High Resolution Mapping Land Cover Classification of the Homony Creek Watershed. Final Report. Available on: cranfield.ac.uk/bitstream/1826/1392/1/thesis+ report.pdf, pp: Defris, R., M. Hansen and J.R.G. Townshend, Global land cover classifications at 8 km spatial reolution: the use of training data derived from landsat imagery in decision tree classifiers. Int. J. Remote Sens., 19: Carreiras, J.M.B., J.M.C. Pereira and Y.E. Shimabukuro, Land-covermapping in the Brazilian Amazon using SPOT-4 vegetation data and machine learning classification methods. Photogramm. Eng. Rem. S., 72: Hansen, M., R. DeFries, J.R.G. Townshend and R. Sohlberg, Global land covers classification at 1km resolution using a decision tree classifier. Int. J. Remote Sens., 21: Bicheron, P., P. Defourny, C. Brockman, L. Schouten, C. Vancutsem, M. Huc, S. Bontemps, M. Leroy, F. Achard, M. Herold, F. Ranera and O. Arino, GlobCover 2005-Products description and validation report, Version 2.1, Available on the ESA IONIA website ( Biomass accumulation rates of Amazonian secondary forest and biomass of old-growth forests from Landsat time series and the Geoscience Laser Altimeter System. J. Appl. Remote Sens., 3: Butler, D., Virtual Globe: The Web-Wide World. Nature, 439: Guralnick, R.P., A.W. Hill and M. Lane, Towards a collaborative, global infrastructure for biodiversity assessment. Ecol. Lett., 10: Leachtenauer, J.C., K. Daniel and T.P. Vogl, Digitizing Corona imagery: Quality vs. cost, Corona: Between the Sun and the Earth, The First NRO Rreconnaissance Eye in Space (R.A. McDonald, editor), American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland, pp: Friedl, M.A., D. Sulla-Menashe, B. Tan, A. Schneider, N. Ramankutty and A. Sibley, MODIS Collection 5 global land cover: algorithm refinements and characterization of new datasets. Remote Sens. Environ., 114: Clark, M.L., A.T. Mitchell, G.H. Ricardo and R. George, A scalable approach to mapping annual land cover at 250 m using MODIS time series data: A case study in the Dry Chaco ecoregion of South America. Remote Sens. Environ., 114: Yang, X., G.M. Jiang, X. Luo and Z. Zheng, Preliminary mapping of high-resolution rural population distribution based on imagery from Google Earth: A case study in the Lake Tai basin, eastern China. Appl. Geogr., 32: Aguirre-Gutiérrez, J., A.C. Seijmonsbergen and J.F. Duivenvoorden, Optimizing land cover classification accuracy for change detection: a combined pixel-based and object-based approach in a mountainous area in Mexico. Appl. Geogr., 34: Potere, D., Horizontal Positional Accuracy of Google Earth s High-Resolution Imagery Archive. Sensors, 8: Tucker, C.J., D.M. Grant and J.D. Dykstra, NASA s Global Orthorectified Landsat Data Set. Photogramm. Eng. Rem. S., 70: Masek, J., M. Honzak, S.N. Goward, P. Liu and E. Pak, Landsat-7 ETM+ as an observatory for land cover: Initial radiometric and geometric comparisons with Landsat-5 Thematic Mapper. Remote Sens. Environ., 78:

6 18. Dorais, A. and J. Cardille, Strategies for 23. Panoramio, Press Release: One million Incorporating High-Resolution Google Earth registered users and five million photos uploaded. Databases to Guide and Validate Classifications: ( Understanding Deforestation in Borneo. users-and-5-million-photos-uploaded/). Remote Sens., 3: Congalton, R.G. and R.A. Mead, A quantitative 19. DeFries, R., F. Achard, S. Brown, M. Herold, method to test for consistency and correctness in D. Murdiyarso, B. Schlamadinger and C. de Souza, photo interpretation. Photogramm. Eng. Rem. S., Reducing Greenhouse Gas Emissions from 49: Deforestation in Developing Countries: 25. Powell, R.L., N. Matzke, C. de Souza, Jr, M.L. Clark, Considerations for Monitoring and Measuring. I. Numata and L.L. Hess, Sources of error in GOFC-GOLD Report Series 26, Global Terrestrial accuracy assessment of thematic land-cover maps in Observing System: Rome, Italy. the Brazilian Amazon. Remote Sens. Environ., 20. DeFries, R., F. Achard, S. Brown, M. Herold, 9: D. Murdiyarso and B. Schlamadinger, Knorn, J., A. Rabe, V.C. Radeloff, T. Kuemmerle, Earth observations for estimating greenhouse gas J. Kozak and P. Hostert, Land cover mapping of emissions from deforestation in developing large areas using chain classification of neighboring countries. Environ. Sci. Policy., 10: Landsat satellite images. Remote Sens. Environ., 21. Iranian Statistics Center, General Census of 113: Population and Housing of Karaj City. 27. Standart, G.D., K.R. Stulken, X. Zhang and Z.L. 22. Taylor, F., February Google Earth 5-Historical Zong, Geospatial visualization of global Imagery.Google Earth Blog 2. satellite images with Vis-EROS. Environ. Modell. Softw., 26: