IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO.

10, OCTOBER 2015 4665

A Scalable Geospatial Web Service for Near

Real-Time, High-Resolution Land Cover Mapping
Konstantinos Karantzalos, Senior Member, IEEE, Dimitris Bliziotis, and Athanasios Karmas, Member, IEEE

Abstract—A land cover classification service is introduced regions twice per week. These huge imaging streams, which
toward addressing current challenges on the handling and online are received through satellite downlink channels at gigabit
processing of big remote sensing data. The geospatial web service rates, increase with tremendous velocity, reaching currently
has been designed, developed, and evaluated toward the efficient
and automated classification of satellite imagery and the produc- several petabytes in many satellite archives [4]–[7]. However,
tion of high-resolution land cover maps. The core of our platform it is estimated that most of the datasets in existing archives
consists of the Rasdaman array database management system for have never been accessed and processed [8]. Harvesting valu-
raster data storage and the open geospatial consortium web cov- able knowledge and information from big EO data is not a
erage processing service for data querying. Currently, the system trivial task.
is fully covering Greece with Landsat 8 multispectral imagery,
from the beginning of its operational orbit. Datasets are stored and Big EO data processing turns out to be extremely challenging
preprocessed automatically. A two-stage automated classification [7], [9]–[11], while the increasing data volumes are not the only
procedure was developed which is based on a statistical learning consideration. As the wealth of data increases, the challenge of
model and a multiclass support vector machine classifier, integrat- indexing, searching, and transferring increases exponentially as
ing advanced remote sensing and computer vision tools like Orfeo well. Open issues include the efficient data storage, handling,
Toolbox and OpenCV. The framework has been trained to clas-
sify pansharpened images at 15-m ground resolution toward the management, and delivery, the processing of multimodal and
initial detection of 31 spectral classes. The final product of our sys- high-dimensional datasets as well as the increasing demands for
tem is delivering, after a postclassification and merging procedure, real-, or near real-time processing for many critical geospatial
multitemporal land cover maps with 10 land cover classes. The applications [2], [4], [12]. Among them, land cover informa-
performed intensive quantitative evaluation has indicated an over- tion and timely observations over the (bio)physical cover of the
all classification accuracy above 80%. The system in its current
alpha release, once receiving a request from the client, can process earth’s surface are of significant importance.
and deliver land cover maps, for a 500-km2 region, in about 20 s, Moreover, the development of novel geospatial web services
allowing near real-time applications. [11], [13]–[15] for on-demand remote sensing analysis is a
Index Terms—Automation, classification, earth observation key issue. Geospatial web services enable users to leverage
(EO), land cover, land use, machine learning, multispectral. distributed geospatial data, products, and computing resources
over the network and from third-party geospatial applications
I. I NTRODUCTION and to automate geospatial data integration and analysis pro-
cedures. These services should be interoperable and allow
T HE PROCESSING of more than 20 000 LANDSAT and
HJ-1 satellite images and the development of the first
global 30-m resolution land cover dataset for 2 baseline years
for collaborative processing of geospatial data for information
and knowledge discovery. The aforementioned features can be
(i.e., GlobeLand30) took almost 4 years [1]. The initial assess- accomplished through the utilization of the service computing
ment of GlobeLand30’s water layer in Northern Europe indi- and workflow technologies [16], [17].
cated high accuracy rates; however, a comprehensive validation To this end, a scalable geospatial web service has been
has yet to be completed. designed, developed, and evaluated toward the efficient and
Although this can be correctly considered as a milestone automated processing of high-resolution satellite data for the
in the history of obtaining global geospatial information from production of land cover maps. The core of our platform
satellite imagery, it indicates the urgent need for efficient consists of the Rasdaman array database management system
and automated classification tools able to process and deliver (DBMS) [18], [19] for big raster data storage and the web cov-
rapidly accurate land cover maps [2], [3]. erage processing service (WCPS) interface standard which is
Indeed, the current generation of space-borne sensors are maintained by the open geospatial consortium (OGC) for data
generating nearly continuous streams of massive earth obser- querying [20]. The WebGIS client is based on the OpenLayers
vation (EO) datasets. Shortly, high-resolution multispectral and GeoExt javascript libraries. Currently, the system is fully
images will be available almost once a week and in some covering Greece with LANDSAT 8 multispectral data, from
the beginning of the satellite’s mission. Datasets are stored
Manuscript received November 02, 2014; revised May 13, 2015; accepted
July 10, 2015. Date of publication September 08, 2015; date of current version
and preprocessed automatically. The automated classification
January 08, 2016. framework is based on a comprehensive statistical training
The authors are with the Remote Sensing Laboratory, National Technical model and on a multiclass support vector machine (SVM) clas-
University of Athens, Athens 15780, Greece (e-mail: [email protected]). sifier, integrating advanced remote sensing and computer vision
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. libraries like GDAL, Orfeo Toolbox, and OpenCV. The frame-
Digital Object Identifier 10.1109/JSTARS.2015.2461556 work has been trained to classify pansharpened images at 15-m
1939-1404 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
4666 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

ground resolution toward the detection of 31 spectral classes raster data handling through DBMSs, OGC services, and
which can address at this particular scale the diversity of the advanced data processing algorithms. The developed geospatial
terrain objects that can be found in Greece. The final product web service is scalable and has been designed to execute on the
delivers, after a postclassification and merging procedure, mul- fly, over the web, on the server-side, image classification tasks
titemporal land cover maps with 10 land cover classes. The for the production of land cover maps. To our knowledge, this
performed quantitative evaluation has indicated a classifica- is the first web-based system that is addressing these challenges
tion accuracy above 80%. Toward the operational use of the offering at a national scale (currently, fully covering Greece)
service, a comprehensive validation on numerous scenes and geospatial products.
dates along with the migration to a cloud environment has been It should be noted that the main contribution is not the
scheduled. classification framework. The system has been designed to
operate with any classifier or training procedure. However,
A. Related Work we have experimented with various advanced training models
and classifiers, with different (numbers of) land cover classes
We have been inspired by the PlanetServer system [5] and hierarchies. We propose an automated two-stage classifi-
which is a service component of the EU-funded EarthServer cation framework which can deliver in near real-time 10 land
project aimed at serving and analyzing planetary data online. cover classes at 15-m spatial resolution based on LANDSAT 8
EarthServer project1 is creating an on-demand online open multispectral, multitemporal data.
access and ad hoc analytics infrastructure for massive (100+
TB) Earth Science data based on cutting-edge array database
platform and OGC WCPS standard. The WCPS has already II. L AND C OVER G EOSPATIAL S ERVICE
proven its scalable and efficient capabilities on quickly process The main objective of this work was to design and implement
large amounts of data and deliver finished products to the end- a framework for the online analysis of multispectral, multi-
user at an extremely low cost by integrating a full hyperspectral temporal satellite imagery for the production of land cover
unmixing chain as part of the NASA Sensor Web suite of web maps. Various components and processing steps are involved
services [6] and by integrating standard processing and vegeta- in setting, running, and utilizing the developed service.
tion analysis procedures for agricultural applications [21], [22]. The overall system architecture is shown in Fig. 1. The
Such frameworks, cloud-based platforms [23]–[25], and high core functionality of the developed framework, consists of the
performance computing [26], [27] form the single way for- Rasdaman Array DBMS for storage of remote sensing data and
ward for applying, on the fly, over the web, on the server-side, OGC WCPS interface standard for querying them. Rasdaman
image analysis and data analytics tasks for the production of was selected as the core system of the implementation due to
valuable geo-information. To this end, the Global Forest Watch its proven robustness, novelty, and efficiency in handling big
Service [28], which is a milestone online platform powered by imaging data2 [5], [11], [21], [22], [37].
Google Earth Engine, forms a dynamic online forest monitoring
and alert system which exploits satellite technology, open data,
and crowdsourcing to guarantee access to timely and reliable A. Data Acquisition and Preprocessing
information about forest status globally. Currently, the open satellite imaging data that system-
Moreover, along with the importance of employing efficient atically are stored in our database are derived from the
tools and platforms for big EO data storing, handling, and US LANDSAT Data Continuity Mission (LDCM).3 The
delivery, advanced algorithms able to learn, retrieve, and clas- LANDSAT 8 Operational Land Imager (OLI) and Thermal
sify information from large datasets are required [29]–[32]. Infrared Sensor (TIRS) instruments acquire multitemporal,
Retrieval from satellite image datasets has been proposed based multispectral data with a spatial resolution of 30 m. LANDSAT
on a semi-supervised method for the annotation of images [33] 8 raw data are downloaded, stored, and preprocessed automati-
and on the enrichment of metadata, of the semantic annotations, cally through our system.
and the image content [34]. A multistage active-learning proce- As far as the download, storage, and preprocessing stages are
dure has been also employed for pattern retrieval in large image concerned, a number of Python scripts were developed which
databases [35], while GPU-based implementations have been control, facilitate, and automate the entire operation. First, any
employed for addressing the computational cost of extracting newly acquired LANDSAT 8 dataset over the Greek territory
spectral information from hyperspectral data [36]. Last but not has been detected and downloaded in an automated manner.
least, a web-based system for the unsupervised classification of Then, another script archives the data, i.e., uncompresses them
satellite and airborne images has been developed [31], which and performs all necessary image preprocessing radiometric
allows the user to perform classification tasks at different zoom correction steps that are required. It should be mentioned that
levels allowing control and supervision over the final result. a cloud-screening process does not let any scene with a cloud
coverage above 70% to be stored in our system.
B. Contribution Regarding the radiometric correction steps, the Unites States
Geological Survey’s (USGS’s) instructions were followed in
Building upon similar efforts [5], [6], [21], we introduce
2 [Online]. Available: http://www.copernicus-masters.com/index.php?kat=wi
a framework that integrates cutting edge geospatial tools for
nners.html&anzeige=winner_t-systems2014.html
1 [Online]. Available: www.earthserver.eu 3 [Online]. Available: http://LANDSAT.usgs.gov/
KARANTZALOS et al.: SCALABLE GEOSPATIAL WEB SERVICE FOR NEAR REAL-TIME, HIGH-RESOLUTION LAND COVER MAPPING 4667

Fig. 1. System architecture of the developed land cover geospatial service.

using the LANDSAT 8 product.4 Digital Numbers are con- support multidimensional arrays with unlimited size of dimen-
verted to top of atmosphere (TOA) reflectance using radio- sions while offering all the classical databases’ advantages.
metric rescaling coefficients provided in the product metadata Rasdaman supports multidimensional arrays of very large
file (MTL file). The MTL file also contains the thermal con- sizes and thus can handle inherently big remote sensing data.
stants needed to convert TIRS data to the at-satellite brightness Rasdaman’s architecture is based on a transparent array par-
temperature. titioning, called tiling. Conceptually, there is no size limi-
After the radiometric corrections, the multispectral bands tation for Rasdaman as a central DBMS of raster datasets.
(MBs) are fused with the panchromatic (PAN) one, in order Additionally, Rasdaman features parallel server architecture
to increase their spatial resolution to 15-m ground resolution. that offers a scalable, distributed environment to efficiently pro-
The pansharpening is performed based on a standard pro- cess very large numbers of concurrent client requests and serve
cess (HRM B /P ANsm ) ∗ P AN through the Orfeo ToolBox.5 distributed datasets across the web.
HRM B is the multispectral image interpolated into the same The Rasdaman database of the developed system currently
size as the PAN, while P ANsm is an isotropically smoothed contains multitemporal imagery from the LANDSAT 8 satel-
version of the raw panchromatic band. The resulting pansharp- lite that covers the entire Greek territory (approximately 45
ened multispectral images have a 15-m ground resolution and paths and rows) from the beginning of the mission. Prior to
are the data which are stored and processed by the service. The the insertion of data in Rasdaman, an appropriate data type
pansharpened images are then reprojected through GDAL6 to for the LANDSAT 8 data needs to be defined. This type def-
WGS 84 (EPSG:4326). inition initially defines the “pixel type” by setting the amount
After the preprocessing stages, data are ready to be inserted of bands and the value type for each band. Then, a raster type
into the Rasdaman database. Thereupon, a final script reads the (multidimensional array) is created which is specified as being
delivered metadata and inserts the datasets and their metadata two-dimensional (2-D), with completely open bounds in all
into the Rasdaman database. directions; thus, the Rasdaman server will allow for images at
any coordinate and with a dynamically growing extent. This
feature allows to serve all LANDSAT 8 images through the
B. Rasdaman and WCPS definition of only one data type regardless of their spatial extent.
Multidimensional arrays of large size are not supported Datasets to be processed are extracted from Rasdaman
by traditional DBMSs. As a consequence, these data are through the execution of retrieval queries written in a query lan-
served through custom-made ad hoc servers which support guage defined by OGC’s WCPS standard. This language allows
arrays, but, on the other hand, lack database features such as for retrieval, filtering, processing, and fast subsetting of multidi-
query languages, query optimization and parallelization, and mensional raster coverages, such as sensor, simulation, image,
access-efficient storage architectures. Array DBMSs, however, and statistics data. For WCPS query language, Rasdaman is the
4 [Online]. reference implementation.
Available: http://LANDSAT.usgs.gov/LANDSAT8_Using_Produ
ct.php WCPS queries are submitted to the Rasdaman database
5 [Online]. Available: www.orfeo-toolbox.org/otb/ server through the PetaScope component [38]. PetaScope
6 [Online]. Available: www.gdal.org
is a java servlet package which implements OGC standard
4668 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

TABLE I
T EN L AND C OVER C LASSES , THE N UMBER OF S UBCLASSES , AND THE T RAINING DATA (N UMBER OF P OLYGONS , T HEIR A REA , AND P ERCENTAGE )

interfaces, thus allowing on-demand submission of queries that

search, retrieve, subset, and process multidimensional arrays of
large size. Moreover, it adds geographic and temporal coordi-
nate system support. The result of the execution of the WCPS
query produced by the service on the Rasdaman server is the
multispectral image to be classified by the service.

C. Land Cover Map Production

In this section, the core process of the land cover mapping
service and, in particular, the “server-side scripts” component
is described (Fig. 1).
A main server-side script (task manager) written in php pro-
gramming language orchestrates the individual scripts, system
components, and procedures that are involved in the production
of land cover maps. It acts as the intermediate connection point
between the service’s web client (Section II-D), the Rasdaman
server, and the various programming scripts that power the
land cover service. When invoked from the Web Client, the
task manager parses the user input (i.e., dataset metadata and
task to fulfill) and retrieves through WCPS queries (data sub-
setting and extraction) the necessary dataset from Rasdaman,
which is stored as a temporary file in the server. Once the
dataset has become available, the image classification functions Fig. 2. Flowchart with the developed land cover classification procedure.
are executed. The resulting classified image is forwarded to
a C program that merges the detected classes and creates the
final land cover product. The product is then georeferenced and the given AOI. It should be noted that the AOI can take any
stored temporarily on the server. arbitrary shape and there is not any limitation regarding the
number of polygon nodes.
The coordinates of the bounding box that define an AOI are
D. Client recorded on completion of the definition of the AOI by the user.
The Web Client of the system is heavily based on GeoExt Then, a server-side script utilizes them so as to find the col-
and OpenLayers javascript libraries. It utilizes them for man- lection that contains the AOI. The returned metadata are then
aging service’s graphical interface and the interaction with the employed to form the WCPS query that will be sent to the task
user, as well as, for placing the produced land cover maps as manager. Once all necessary information has been determined,
layers inside a map. In order to achieve this, metadata need the web client sends it asynchronously to the task manager and
to be determined. These metadata are the coordinates of the waits for the response. In order to handle request timeout on the
vertices of the bounding box that defines an area of interest client side, a request timeout rule was also implemented. The
(AOI) that the user has specified and the name of the collection implementation takes into account the areal extent of the AOI
in the Rasdaman database that hosts the image which contains defined by the user and sets the timeout parameter accordingly.
KARANTZALOS et al.: SCALABLE GEOSPATIAL WEB SERVICE FOR NEAR REAL-TIME, HIGH-RESOLUTION LAND COVER MAPPING 4669

Fig. 3. Land cover classification results from the developed geospatial service. The large multispectral mosaic, from central Greece, with 15-m ground resolution
of 8684 × 8676 pixels, covering 16 952 km2 , was classified, after approximately 10 min. Images were acquired on June 25, 2013. The quantitative evaluation
indicated an OA rate of 82.8% with a Kappa coefficient of 0.801 (Table II).

The timeout parameter is set between 1 and 3 min depending on A. Land Cover Classes and Hierarchy
the exact areal extent. As far as very large AOIs are concerned, These subclasses belong to the 10 major land cover types
longer processing time is required. For such cases, the user
that are shown in Table I, i.e., water bodies, wetland, artifi-
receives an e-mail with a url address pointing to the result.
cial surfaces, forest, shrubland, grassland, bareland, cultivated
Upon receipt of the response (land cover map), the coordi- land, cloud/shadows, and snow/ice. For each class, the number
nates that define the areal extent of the AOI are again used
of the corresponding subclasses and the number of the man-
through the image layer capability offered by OpenLayers, in ually annotated training polygons are also shown in Table I.
order to place the land cover layer at its correct location in the The cultivated land class includes the larger number of both
Client’s OpenLayers map. This allows for spectral data analysis
subclasses and training samples due to the diverse spectral
in a web environment. The resulting land cover map can there- properties of the corresponding subclasses (e.g., row crop-
fore be compared with overlapping visual imagery and geodata lands, orchards, small grain croplands, fallow croplands, and
from various sources.
harvested croplands). Artificial surfaces, forest, water bod-
ies, and bareland also possess a number of subclasses and
III. C LASSIFICATION F RAMEWORK polygons.
In terms of the area that the training samples are covering
The classification framework was developed to reflect the (in square kilometer), the cultivated land and the water bodies
major land types in Greece with reference to major land cover are holding the larger share, i.e., 28% of all samples. Forest
products and mapping practices [1], [39], [40]. Focusing on and shrubland follow with approximately 15% and then the
delivering land cover classification maps with 15-m ground rest of the classes with less than 5% each. In total, the training
resolution, a two-stage classification procedure was employed, data covered approximately 11 km2 , i.e., approximately 49 000
which discriminates initially 31 subclasses and then after a pixels (Table I).
postprocessing and a grouping procedure delivers 10 aggre-
gated classes. A flowchart with the developed land cover
classification procedure is shown in Fig. 2. B. Training
The 31 subclasses, which were derived through a comprehen-
sive, time-consuming manual selection of numerous training All manually annotated samples, which were digitized in var-
data all over Greece in various images from different dates, ious images of different dates, were employed for the training
represent the land cover types, terrain objects, and surfaces of the SVM classifier. Our implementation was based on the
that can be found in Greece at this particular spatial scale. It Orfeo Toolbox7 [42], OpenCV.8 and LibSVM [43] software
should be noted that experiments started with multiple unsu- libraries, which provide an efficient framework for the inte-
pervised classification in various images, all over Greece, gration of existing machine learning algorithms. Briefly, Orfeo
with 60 classes. After a demanding, thorough, trial and error Toolbox (OTB) is an open source library which is focus-
procedure, these 31 subclasses were selected as representative. ing on the analysis of large remote sensing images, OpenCV
This procedure included study of the spectral properties, sea- is a computer vision library which is focusing on computa-
sonal behaviors, and a cross validation from higher resolution tional efficiency and real-time applications and LibSVM is
imaging (e.g., google earth) and vector geospatial open data 7 [Online]. Available: http://www.orfeo-toolbox.org/SoftwareGuide/
(e.g., geodata.gov.gr) [41]. 8 [Online]. Available: http://opencv.org/documentation.html
4670 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

TABLE II
C ONFUSION M ATRIX FOR THE Q UANTITATIVE E VALUATION OF THE D ELIVERED M ULTITEMPORAL L AND C OVER C LASSIFICATION M AP OF F IG . 3

an integrated software tool for support vector classification, IV. E VALUATION

multiclass classification, regression, and distribution estimation A. Experimental Results
(Fig. 2).
The training/prediction model was built based on image The experimental validation was performed by applying
statistics including all multispectral bands, the normalized dif- the developed service on numerous regions of different sizes,
ference vegetation index (NDVI), and the normalized differ- on different dates, of various land cover types and complex
ence water index (NDWI). The Kappa coefficient, which was terrain objects. A demo with the overall functionality and per-
employed for validating the training model, indicated a Kappa formance of the developed system can be found here.9 In
coefficient above 80%. particular, the alpha version of our system is currently hosted
on an eight-core machine with 32-GB RAM running Debian
GNU/Linux (Release 7.5), Apache Tomcat 6, and the open
source Rasdaman community version 8.5. Under this demon-
C. Two-Stage Classification
stration environment, the stored and preprocessed remote sens-
The developed geospatial service was designed to deliver in a ing data are fully covering the Greek territory providing every
systematic and operational manner, validated land cover maps approximately 16-day satellite imagery from the beginning of
in near real-time. Toward addressing such challenges, a two- the LANDSAT 8 operational orbit (April 2013). Obviously, the
stage classification strategy was implemented in Fig. 2. First, current hardware cannot support numerous concurrent users
experimenting on a national scale with LANDSAT 8 datasets and a migration to a cloud environment has already been
from the beginning of its operational orbit (i.e., 4/2013) and scheduled.
validating land cover maps at 15-m resolution for 31 land cover Along with the collection of the training samples for the 31
subclasses was neither feasible nor within the perspectives of subclasses in Table I, the collection of the ground truth data
our work. Moreover, in all experiments, on all datasets, on all was performed with another independent procedure. In partic-
dates, and by employing different classifiers (e.g., SVMs and ular, ground truth/ reference polygons were manually collected
decision trees), the overall classification accuracy for the 31 and annotated from various images and acquisition dates. About
subclasses was above 70% reaching approximately 80%. 50 000 pixels were used for the training process and about
Therefore, by employing a standard postprocessing (i.e., 50 000 for the quantitative evaluation, all referring to the 31
majority voting filter) technique and grouping the subclasses land cover classes. For the quantitative evaluation, the reference
into 10 land cover classes (Table I), the framework managed to polygons were merged, respectively, to match the 10 land cover
systematically deliver land cover maps with an overall accuracy classes of the end product. These manually annotated polygons
(OA) of above 80% across different datasets and dates. This was were verified by intensive comparisons from very-high reso-
acceptable for the current alpha system release. lution imaging data (e.g., orthophotos and google earth) and
Moreover, in all experiments, the SVM classifier [43] using vector geospatial open data (e.g., geodata.gov.gr).
the one-against-one strategy delivered the highest accuracy High-resolution imaging data and vector information were
rates with an acceptable computational cost, allowing near real- also employed for the qualitative evaluation of the produced
time computations on the server side. In order to slightly refine land cover maps. Such systematic qualitative examinations
the initial classification results and enhance its classification of the delivered maps and comparisons (both qualitative and
accuracy, a standard majority voting filtering (in a 3 × 3 win- quantitative) with other maps and data sources indicated high
dow) was employed, which ameliorated the OA in all experi-
ments by 1%–5%. 9 [Online]. Available: http://users.ntua.gr/karank/Demos/WebLC.html
KARANTZALOS et al.: SCALABLE GEOSPATIAL WEB SERVICE FOR NEAR REAL-TIME, HIGH-RESOLUTION LAND COVER MAPPING 4671

Fig. 4. Land cover classification results from the developed geospatial service over a multitemporal dataset. Each multispectral image of 1446 × 1682 pixels,
covering 547 km2 was processed and delivered to the client after approximately 22 s. The quantitative evaluation indicated an OA rate above 87% in all cases
(Table III).

multitemporal and spatially distributed confidence values for should be noted that the initial reported classification accuracy
the developed service. In all experiments, the qualitative assess- for the 31 subclasses was 72.1% with a Kappa coefficient of
ment was generally in accordance with the quantitative assess- 0.713.
ment. During the evaluation of the first classification step The performance of the land cover service was demonstrated
(detection of 31 land cover classes), the performed assessment over this particular extended scene because it was representative
was mainly based on the quantitative results of various con- of all land cover classes, subclasses, and terrain objects in
fusion matrices, the reported Kappa coefficient, and OA rates. Greece, covering thousands of square kilometer. The results
In particular, the OA for this initial classification step was from its quantitative evaluation were also representative of
above 64%, including scenes and dates with cloud coverage the numerous experimental results performed over different
above 15%. The lowest accuracy rates were found in certain multitemporal scenes. In particular, the quantitative assess-
subclasses, i.e., artificial surfaces, cultivated land, bareland, ment (Table II) indicated that the water bodies class has been
and forest. In particular, areas belonging to the dense urban fab- detected with the highest rates (above 99%) based both on the
ric subclass of the Artificial surfaces class were systematically producer’s accuracy (PA), which indicates the probability of a
misclassified to the white sand or to the quarries subclasses of reference pixel being correctly classified, and the user’s accu-
the bareland class. racy (UA), which indicates the probability of a pixel in the map
In Fig. 3, the delivered land cover classification map from the actually representing this category on the ground. This is in
developed geospatial service is shown over a region in Central accordance with the literature [1], [44]. High accuracy rates
Greece. A large multispectral mosaic of LANDSAT 8 data, (above 93%) have also been obtained for the cloud/shadows
with 15-m ground resolution of 8684 × 8676 pixels, covering class for both PA and UA. The shrubland class has also been
almost 17 000 km2 was processed by the system and the land systematically detected with very high PA rates (98% in this
cover map was delivered after approximately 10 min. The cor- case) and both the wetland and forest classes with very high
responding LANDSAT 8 images were acquired on June 25, UA rates (above 97% in this case).
2013. The performed quantitative evaluation indicated an OA The lowest PA has been reported for the artificial surfaces
rate of 82.8% with a Kappa coefficient of 0.801 (Table II). It and bareland classes (with 64.3% and 68.2%, respectively).
4672 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

Fig. 5. Land cover classification results from the developed geospatial service over a multitemporal dataset. Each multispectral image of 1446 × 1682 pixels,
covering 547 km2 was processed and delivered to the client after approximately 22 s. The quantitative evaluation indicated an OA rate above 86% in all cases
(Table III).

The bareland class was also calculated with the lowest UA at Qualitative comparison between the experimental results and
55.8% (Table II). In particular, in all conducted experiments, different color composites and other higher resolution images
pixels from the artificial surfaces class have been systemati- showed that the delivered multitemporal land cover maps agree
cally misclassified as bareland. This is mainly inherited by the spatially with the actual land cover types in the ground. This
initial classification step with the 31 subclasses. Bareland pix- observation is verified by the quantitative assessment that indi-
els were also misclassified as cultivated land, while cultivated cated OA rates ranging from 86% to 89% (Table III). In
land pixels were mixed with the shrubland ones. particular, there is a rather small difference between the cal-
In addition, in order to demonstrate the performance of the culated accuracy rates at all dates and this can indicate the
land cover service for multitemporal data, experimental results stable performance of the developed classification procedure.
from six different acquisition dates are shown in Figs. 4 and 5. A small difference in the OA rates during the initial classifi-
In the top row of both figures, a natural color composite is cation with the 31 subclasses is also reported in all cases, i.e.,
shown and the corresponding land cover map, which was deliv- from 75% to 81%. It should be mentioned that the lowest accu-
ered by the system, is shown in the bottom. Land cover classes’ racy rates during the initial step do not necessarily result into
colors are the same ones as in Fig. 3. This particular region, the lowest accuracy rates at the end-product. Such a correlation
covering approximately 550 km2 , contains almost all land cover has not been observed in any of our experiments. For exam-
classes, mountainous, rural and urban regions. The size of the ple, the initial 31 subclasses have been detected with a relative
images was 1446 × 1682 pixels and in all cases, the system low OA (75%) for May 10, 2014, while after the postprocess-
delivered the land cover product to the client after approxi- ing and merging procedure, the evaluation indicated an OA
mately 22 s. The time that the developed service required for of 89%.
the entire process (from the client request to the delivered map In general, the corresponding confusion matrices from all
on the client) was, in all cases, measured using the Firebug dates of Figs. 4 and 5 agree with the systematic observa-
2.0.4 application. In Fig. 6, an example of a WCPS request tions regarding the performance of the developed classification
script is shown which retrieves an image from the Rasdaman procedure as previously reported. The worth-mentioning dif-
database. ferences are the misclassified pixels from the artificial surfaces
KARANTZALOS et al.: SCALABLE GEOSPATIAL WEB SERVICE FOR NEAR REAL-TIME, HIGH-RESOLUTION LAND COVER MAPPING 4673

in its current alpha release, delivers, after a postclassification

and merging procedure, multitemporal land cover maps with
10 land cover classes. The performed quantitative evaluation
has indicated a land cover classification accuracy above 80%.
Toward the operational use of the service and the forth-
coming beta release, comprehensive validation on numerous
scenes, which will also include ground truth data from field
surveys, will be carried out. Furthermore, the migration of
the platform to a cloud environment is already scheduled.
Moreover, since the system has been designed to operate with
any processing and classification algorithm, a benchmark of
state-of-the-art approaches on training, learning, and classi-
Fig. 6. Example of a WCPS request script which retrieves from the Rasdaman fication on the continuously increasing dataset (and ground
database an image named “c” as a multiband geotiff file, in order to form the truth) is to be performed. Last but not least, toward improv-
input for the classification procedure. ing significantly current computation times, parallel computing
formulations and GPU implementations will be also employed.
TABLE III
Q UANTITATIVE E VALUATION OF THE L AND C OVER C LASSIFICATION
M APS OF F IGS . 4 AND 5
R EFERENCES
[1] J. Chen et al., “Global land cover mapping at 30m resolution: A POK-
based operational approach,” Int. J. Photogramm. Remote Sens., vol. 103,
pp. 7–27, 2014.
[2] E. Zell, A. Huff, A. Carpenter, and L. Friedl, “A user-driven approach
to determining critical earth observation priorities for societal benefit,”
IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 5, no. 6,
pp. 1594–1602, Dec. 2012.
[3] C. Giri, B. Pengra, J. Long, and T. Loveland, “Next generation of global
land cover characterization, mapping, and monitoring,” Int. J. Appl. Earth
Observ. Geoinf., vol. 25, pp. 30–37, 2013.
[4] A. J. Plaza and C. I. Chang, High Performance Computing in Remote
Sensing. Boca Raton, FL, USA: Chapman & Hall/CRC Press, 2007.
[5] J. Oosthoek et al., “Planetserver: Innovative approaches for the online
analysis of hyperspectral satellite data from Mars,” Adv. Space Res.,
class to the cultivated land and vice versa, apart from June 25, vol. 53, pp. 219–244, 2013.
2013 case. In a smaller extent, shrubland pixels have also been [6] P. Cappelaere, S. Sanchez, S. Bernabe, A. Scuri, D. Mandl, and A. Plaza,
misclassifed as grassland ones. These differences are mostly “Cloud implementation of a full hyperspectral unmixing chain within
the NASA web coverage processing service for EO-1,” IEEE J. Sel.
due to the particular spectral variation among the different land Topics Appl. Earth Observ. Remote Sens., vol. 6, no. 2, pp. 408–418, Apr.
cover classes in this specific region and available ground truth 2013.
polygons (reference pixels employed for the evaluation were [7] Y. Ma, L. Wang, P. Liu, and R. Ranjan, “Towards building a data-intensive
index for big data computing—A case study of remote sensing data
approximately 2100 in total). processing,” Inf. Sci., vol. 319, pp. 171–188, 2014.
[8] C. Nikolaou et al., “Big, linked and open data: Applications in the german
aerospace center,” in The Semantic Web: ESWC 2014 Satellite Events.
New York, NY, USA: Springer, 2014, pp. 444–449.
V. C ONCLUSION AND P ERSPECTIVES [9] Y. Ma, L. Wang, A. Zomaya, D. Chen, and R. Ranjan, “Task-tree
In this paper, a land cover classification web service has been based large-scale mosaicking for massive remote sensed imageries with
dynamic DAG scheduling,” IEEE Trans. Parallel Distrib. Syst., vol. 25,
designed, developed, and evaluated toward the efficient and no. 8, pp. 2126–2137, Aug. 2014.
automated processing of high-resolution satellite data. The core [10] B. C. Pijanowski, A. Tayyebi, J. Doucette, B. K. Pekin, D. Braun, and
of the platform consists of the Rasdaman Array DBMS for big J. Plourde, “A big data urban growth simulation at a national scale:
Configuring the GIS and neural network based land transformation model
raster data storage and the OGC WCPS interface standard for to run in a high performance computing (HPC) environment,” Environ.
data querying. The WebGIS client is based on the OpenLayers Modell. Softw., vol. 51, pp. 250–268, 2014.
and GeoExt javascript libraries. Currently, the system is fully [11] A. Karmas, A. Tzotsos, and K. Karantzalos, “Scalable geospatial web
services through efficient, online and near real-time processing of earth
covering Greece with LANDSAT 8 multispectral data, from observation data,” in Proc. IEEE Int. Conf. Big Data Comput. Serv. Appl.,
the beginning of its operational orbit. Datasets are stored and 2015, pp. 194–201.
preprocessed automatically in our hardware. The automated [12] A. J. Plaza, “Special issue on architectures and techniques for real-time
processing of remotely sensed images,” J. Real-Time Image Process.,
classification framework is based on a comprehensive statistical vol. 4, no. 3, pp. 191–193, 2009.
training model and on a multiclass SVM classifier, integrat- [13] P. Zhao, T. Foerster, and P. Yue, “The geoprocessing web,” Comput.
ing advanced remote sensing and computer vision libraries like Geosci., vol. 47, pp. 3–12, 2012.
[14] P. Yue, L. Di, Y. Wei, and W. Han, “Intelligent services for discovery
GDAL, OTB, and OpenCV. The framework has been trained of complex geospatial features from remote sensing imagery,” ISPRS J.
to classify pansharpened images at 15-m ground resolution Photogramm. Remote Sens., vol. 83, pp. 151–164, 2013.
toward the detection of 31 different land cover classes which [15] A. Karmas, A. Tzotsos, and K. Karantzalos, “Big geospatial data for envi-
ronmental and agricultural applications,” in Big Data Concepts, Theories
can address at this particular scale the diverse land cover types and Applications, S. Yu and S. Guo, Eds. New York, NY, USA: Springer,
that can be found in Greece. The final product of our system, 2016.
4674 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

[16] P. Yue et al., “GeoPW: Laying blocks for the geospatial processing web,” [39] B. Waske, M. Chi, J. Atli Benediktsson, S. van der Linden, and B. Koetz,
Trans. GIS, vol. 14, no. 6, pp. 755–772, 2010. “Algorithms and applications for land cover classification—A review,”
[17] W. Han, Z. Yang, L. Di, and P. Yue, “A geospatial web service approach in Geospatial Technology for Earth Observation, D. Li, J. Shan, and
for creating on-demand cropland data layer thematic maps,” Trans. Amer. J. Gong, Eds. New York, NY, USA: Springer, 2009, pp. 203–233.
Soc. Agric. Biol. Eng., vol. 57, no. 1, pp. 239–247, 2014. [40] M. Hansen and T. Loveland, “A review of large area monitoring of
[18] P. Baumann, “Management of multidimensional discrete data,” Int. J. land cover change using Landsat data,” Remote Sens. Environ., vol. 122,
Very Large Data Bases, vol. 4, no. 3, pp. 401–444, 1994. pp. 66–74, 2012.
[19] P. Baumann, “Array databases and raster data management,” in [41] G. Foody, A. Mathur, C. Sanchez-Hernandez, and D. Boyd, “Training set
Encyclopedia of Database Systems, T. Ozsu and L. Liu, Eds. New York, size requirements for the classification of a specific class,” Remote Sens.
NY, USA: Springer, 2009. Environ., vol. 104, p. 1–14, 2006.
[20] P. Baumann, “The OGC web coverage processing service (WCPS) [42] E. Christophe and J. Inglada, “Open source remote sensing: Increasing
standard,” GeoInformatica, vol. 14, no. 4, pp. 447–479, 2010. the usability of cutting-edge algorithms,” IEEE Geosci. Remote Sens.
[21] A. Karmas, K. Karantzalos, and S. Athanasiou, “Online analysis of Newslett., pp. 9–15, 2009.
remote sensing data for agricultural applications,” in Proc. OSGeo’s Eur. [43] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector
Conf. Free Open Source Softw. Geospat., Jul. 2014, pp. 1–9. machines,” ACM Trans. Intell. Syst. Technol., vol. 2, pp. 27:1–27:27,
[22] K. Karantzalos, A. Karmas, and A. Tzotsos, “RemoteAgri: Processing 2011.
online big earth observation data for precision agriculture,” in Proc. Eur. [44] Z. Zhu and C. E. Woodcock, “Continuous change detection and clas-
Conf. Precis. Agric., 2015, pp. 421–428. sification of land cover using all available Landsat data,” Remote Sens.
[23] K. Evangelidis, K. Ntouros, S. Makridis, and C. Papatheodorou, Environ., vol. 144, pp. 152–171, 2014.
“Geospatial services in the cloud,” Comput. Geosci., vol. 63, pp. 116–122,
2014.
[24] D. Fustes, D. Cantorna, C. Dafonte, B. Arcay, A. Iglesias, and Konstantinos Karantzalos (SM’05) received the
M. Manteiga, “A cloud-integrated web platform for marine monitoring Engineering Diploma degree in survey engineer from
using GIS and remote sensing. Application to oil spill detection through the National Technical University of Athens (NTUA),
SAR images,” Future Gener. Comput. Syst., vol. 34, pp. 155–160, 2014. Athens, Greece, and the Ph.D. degree in remote
[25] I. A. T. Hashem, I. Yaqoob, N. B. Anuar, S. Mokhtar, A. Gani, and sensing from NTUA, in collaboration with Ecole
S. U. Khan, “The rise of ‘big data’ on cloud computing: Review and open Nationale de Ponts et Chaussees (CERTIS, ENPC),
research issues,” Inf. Syst., vol. 47, pp. 98–115, 2015. Marne-la-Vallée, France, in 2007.
[26] C. Lee, S. Gasster, A. Plaza, C.-I. Chang, and B. Huang, “Recent devel- In 2007, he joined the Department of Applied
opments in high performance computing for remote sensing: A review,” Mathematics, Ecole Centrale de Paris, Paris, France,
IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 4, no. 3, as a Postdoctoral Researcher working with Prof.
pp. 508–527, Sep. 2011. Nikos Paragios (Head of the Center for Visual
[27] X. Su, J. Wu, B. Huang, and Z. Wu, “GPU-accelerated computation for Computing). He is currently an Assistant Professor with the Remote Sensing
electromagnetic scattering of a double-layer vegetation model,” IEEE J. Laboratory, National Technical University of Athens, joining the Remote
Sel. Topics Appl. Earth Observ. Remote Sens., vol. 6, no. 4, pp. 1799– Sensing Laboratory (Head: Prof. Demetre Argialas). He is also with i-SENSE
1806, Aug. 2013. Group, Institute of Communication and Computer Systems (ICCS), NTUA. He
[28] M. C. Hansen et al., “High-resolution global maps of 21st-century forest has several publications in top-rank international journals and conferences and
cover change,” Science, vol. 342, no. 6160, pp. 850–853, 2013. a number of awards and honors for his research contributions. He has more
[29] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification than 15 years of research experience, involved with more than 14 EU and
with deep convolutional neural networks,” in Proc. Neural Inf. Process. national excellence/competitive research projects. His research interests include
Syst. (NIPS’12), 2012, pp. 1097–1105. geoscience and remote sensing, big data, hyperspectral image analysis, com-
[30] B. Liu, E. Blasch, Y. Chen, D. Shen, and G. Chen, “Scalable sentiment puter vision and pattern recognition, environmental monitoring, and precision
classification for big data analysis using Naive Bayes Classifier,” in Proc. agriculture. http:// users.ntua.gr/karank
IEEE Int. Conf. Big Data, Oct. 2013, pp. 99–104.
[31] A. Ferran, S. Bernabe, P. Rodriguez, and A. Plaza, “A web-based system
for classification of remote sensing data,” IEEE J. Sel. Topics Appl. Earth Dimitris Bliziotis received the B.Sc. degree in
Observ. Remote Sens., vol. 6, no. 4, pp. 1934–1948, Aug. 2013. rural and surveying engineering from the National
[32] M. Muja and D. Lowe, “Scalable nearest neighbor algorithms for high Technical University of Athens (NTUA), Athens,
dimensional data,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 36, Greece, in 2014. He is currently pursuing the M.Sc.
no. 11, pp. 2227–2240, Nov. 2014. degree in geoinformatics at the same university.
[33] P. Blanchart and M. Datcu, “A semi-supervised algorithm for auto- Since 2013, he has been a Researcher with the
annotation and unknown structures discovery in satellite image Remote Sensing Laboratory, NTUA. He is a Skilled
databases,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 3, Programmer and Free and Supporter of Open Source
no. 4, pp. 698–717, Dec. 2010. software. His research interests include remote
[34] D. Espinoza-Molina and M. Datcu, “Earth-observation image retrieval sensing, geoinformatics, artificial intelligence, and
based on content, semantics, and metadata,” IEEE Trans. Geosci. Remote machine learning.
Sens., vol. 51, no. 11, pp. 5145–5159, Nov. 2013.
[35] P. Blanchart, M. Ferecatu, S. Cui, and M. Datcu, “Pattern retrieval in large
image databases using multiscale coarse-to-fine cascaded active learn- Athanasios Karmas (M’15) received the Diploma
ing,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 7, no. 4, degree in electrical and computer engineering from
pp. 1127–1141, Apr. 2014. the National Technical University of Athens (NTUA),
[36] J. Sevilla, S. Bernabe, and A. Plaza, “Unmixing-based content retrieval Athens, Greece, in 2014. He is currently pursu-
system for remotely sensed hyperspectral imagery on GPUS,” J. ing the M.Sc. degree in geoinformatics at the same
Supercomput., vol. 70, pp. 1–12, 2014. university.
[37] P. Baumann, “Rasdaman: Array databases boost spatio-temporal analyt- Since 2013, he has been a Researcher with the
ics,” in Proc. 5th Int. Conf. Comput. Geospat. Res. Appl. (COM. Geo), Remote Sensing Laboratory, NTUA. He is a pas-
Aug. 2014, pp. 54–54. sionate Supporter of Free and Open Source soft-
[38] A. Aiordachioaie and P. Baumann, “Petascope: An open-source imple- ware. His research interests include array DBMS,
mentation of the OGC WCS geo service standards suite,” in Scientific and big data, cloud computing, geoinformatics, as
Statistical Database Management, vol. 6187, M. Gertz and B. Ludascher, well as earth observation data management and geospatial applications.
Eds. New York, NY, USA: Springer, 2010, pp. 160–168. http://users.ntua.gr/karank