On Scalable Data Mining Techniques for Earth Science

Abstract

One of the observations made in earth data science is the massive increase of data volume (e.g, higher resolution measurements) and dimensionality (e.g. hyper-spectral bands). Traditional data mining tools (Matlab, R, etc.) are becoming redundant in the analysis of these datasets, as they are unable to process or even load the data. Parallel and scalable techniques, though, bear the potential to overcome these limitations. In this contribution we therefore evaluate said techniques in a High Performance Computing (HPC) environment on the basis of two earth science case studies: (a) Density-based Spatial Clustering of Applications with Noise (DBSCAN) for automated outlier detection and noise reduction in a 3D point cloud and (b) land cover type classification using multi-class Support Vector Machines (SVMs) in multi- spectral satellite images. The paper compares implementations of the algorithms in traditional data mining tools with HPC realizations and ’big data’ technology stacks. Our analysis reveals that a wide variety of them are not yet suited to deal with the coming challenges of data mining tasks in earth sciences.

Full text

doi: 10.1016/j.procs.2015.05.494
On Scalable Data Mining Techniques for Earth Science
Markus G¨otz
1,2
, Matthias Richerzhagen
1
, Christian Bodenstein
1,2
,Gabriele
Cavallaro
2
, Philipp Glock
1
, Morris Riedel
1,2
,andJ´on Atli Benediktsson
2
1
J¨ulich Supercomputing Center (JSC), Forschungszentrum J¨ulich, J¨ulich, Germany
{m.goetz,
m.richerzhagen, c.bodenstein, p.glock, m.riedel}@fz-juelich.de
2
School of Engineering and Natural Sciences, University of Iceland, Reykjavik, Iceland
{cavallaro,
benedikt}@hi.is
Abstract
One of the observations made in earth data science is the massive increase of data volume (e.g,
higher resolution measurements) and dimensionality (e.g. hyper-spectral bands). Traditional
data mining tools (Matlab, R, etc.) are becoming redundant in the analysis of these datasets,
as they are unable to process or even load the data. Parallel and scalable techniques, though,
bear the potential to overcome these limitations. In this contribution we therefore evaluate
said techniques in a High Performance Computing (HPC) environment on the basis of two
earth science case studies: (a) Density-based Spatial Clustering of Applications with Noise
(DBSCAN) for automated outlier detection and noise reduction in a 3D point cloud and (b)
land cover type classification using multi-class Support Vector Machines (SVMs) in multi-
spectral satellite images. The paper compares implementations of the algorithms in traditional
data mining tools with HPC realizations and ’big data’ technology stacks. Our analysis reveals
that a wide variety of them are not yet suited to deal with the coming challenges of data mining
tasks in earth sciences.
Keywords: Data Mining, Machine Learning, HPC, DBSCAN, SVM, MPI
1 Introduction
The enormous increase in variety, velocity, and volume of spatio-temporal earth science datasets
raises data management concerns for organizations storing and preserving the data. There are
a number of concrete examples in view of earth science data repositories. One of them is the
PANGAEA earth science data collection [7] that has reached, at the time of writing, approx-
imately nine billion data items within around 350,000 datasets. These consist of increasingly
large and complex data which often represent long lasting time series of measurement devices
from a multitude of different sensors. Another example are the large data collection of remote
sensing images [2], taken by airborne sensors or satellites observing, measuring, and recording
the radiation reflected or emitted by the Earth and its environment. The ’big data’ challenges
in this case concern the improvements of remote sensing capabilities and the availability of
Procedia Computer Science
Volume 51, 2015, Pages 2188–2197
ICCS 2015 International Conference On Computational Science
2188 Selection and peer-review under responsibility of the Scientific Programme Committee of ICCS 2015
c
 The Authors. Published by Elsevier B.V.

remotely sensed images with high geometrical resolution (e.g. WorldView-2 0.5m) or more de-
tailed spectral information in terms of precision, frequency, or complexity (e.g. AVIRIS 224
spectral channels).
Researchers, however, do not only face issues with the management of the data, but also
when attempting to analyze, interpret, or understand them. Due to their ’big data’ properties,
they are difficult to process, making it a problem worth investigating. One of the known quotes
in that context is: ’Big data is data that becomes large enough that it cannot be processed
using conventional methods [19]. While this definition seems vague, we would like to take the
viewpoint in this contribution that these conventional methods span the wide variety of serial
tools existing to analyze earth science datasets (e.g. R, Matlab, Weka, etc.). While trying to use
some of the already mentioned databases, we have been unable to fully analyze or even just load
them into these systems because of memory restriction or highly inefficient implementation.
Handling processing intensive, large datasets, though, is not an entirely new problem, es-
p
ecially when considering computation environments driven by High Performance Computing
(HPC). We would like to find out what benefits of these environments are applicable to the
analysis of large quantities of earth science datasets. In the foreseeable future, and in order
to better understand many of the climate issues we face today, we even require a systematic
approach of combining and studying earth science datasets derived both from observations and
model simulations often running in HPC environments. Therefore, this contribution aims to
provide a deeper understanding about two concrete case studies that are able to take advan-
tage of parallel and scalable data mining techniques of large quantities of earth science data.
Throughout the paper, and as part of these studies, we provide pieces of information about
strengths and weaknesses of various ’big data’ technologies and comparing approaches to HPC
environments and their strong capabilities that are also applicable to statistical earth science
data mining tasks.
The first case study covers the known territory of finding outliers in data using clustering
algorithms. While many different clustering algorithms are potentially applicable [16], we focus
here on the Density-based Spatial Clustering of Applications with Noise [8] known as DBSCAN.
One example of a relevant data source is a marine observatory (i.e. Koljoefjord observatory)
that delivers continuous measurements over a long period of time. We study the use of DBSCAN
in order to find out if we are able to detect events or outliers that can also support the data
quality management. A scalable and robust automated outlier detection system that is able
to cope with large volumes of data within PANGAEA is therefore required. During our study
we were unable to find a suitable implementation of the algorithm allowing us to perform the
described analysis. For this reason we had to develop our own scalable solution of a parallel
DBSCAN delaying the actual analysis of the PANGAEA Koljoefjord observations. We will
demonstrate the capabilities of the software on the basis of a simpler, but visually tangible,
earth science outlier use case, the denoising of a 3D point cloud.
Our second case study deals with land cover mass classification in remote sensing image
datasets. The classification problem aims to categorize all pixels in a digital image into mean-
ingful classes of land cover types in a particular scene. In order to obtain a satisfactory level
of detection accuracy, we perform a detailed physical analysis by exploiting the availability
of high spatial resolution images. Hence, we consider attribute filters, flexible operators that
can transform an image according to many different attributes (e.g. geometrical, textural, and
spectral) as further optimization technique [3]. One often used classification technique in the
field of remote sensing are Support Vector Machines (SVMs) [6] and their kernel methods (e.g.
radial basis function). The contribution of this paper in the context of classification is the
exploration of according SVM frameworks.
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2189

This paper is structured as follows. After the introduction into the problem domain in
Section 1, key requirements for evaluating parallel and scalable tools for the analysis of scientific
and engineering problems are given in Section 2. Section 3 offers a thorough survey of related
work in the field of parallel and scalable data mining tools with a particular focus on parallel
DBSCAN and SVM implementations. Section 4 highlights selected technical details on two
parallel implementations we use to process the aforementioned earth science datasets. The
paper ends with some concluding remarks.
2 Requirements
Technical and algorithmic solutions for the aforementioned case studies need to satisfy a number
of key requirements listed in this section. Satisfying those requirements and overcoming scal-
ability constraints given by the available datasets thus represent one of our major motivation.
There is a wide variety of traditional tools available and an increasing number of more recent
’big data stacks’ that claim to support parallel and scalable data mining in one way or the
other. We picked the following (without considering commercial tools) for our deeper analysis:
(i) Weka, (ii) R, (iii) Matlab, (iv) Octave, (v) Apache Mahout, (vi) MLlib/Apache Spark, (vii)
scikit-learn, and individual implementations for the specific algorithm in question. Those tools
are analysed in terms of three different criteria ’(a) open and free availablility’, ’(b) technical
fe
asibility’,and’(c) suitability of algorithms’ in the light of the motivating case studies. Despite
the fact that there is an ever increasing number of rather new ’big data stacks’,ourworkis
also motivated by exploring whether traditional HPC environments play a role in mining ’big
data’. Although the HPC environments are typically driven by demands of the ’simulation
sciences’, based on efficient numerical methods and known physical laws, some computational
science applications raise similar requirements to the processing environments as it is the case
for data mining tasks. In this contribution we therefore focus on comparing solutions based on
the following five key capabilities we consider important when analyzing technical solutions:
R1 Parallel file systems and large storage capacity
R2 Scalable standard data formats that take advantage of parallel I/O
R3 Standard communication protocols
R4 Fair scheduling tools and policies to enable the work on clusters for several users at a time
R5 Open source tools and open referencable data to enable reproducability of findings
3 Related Work
This section is structured along the key requirements and framework collection introduced in
section 2.
3.1 Survey of DBSCAN Clustering Tools
Clustering is an established data mining method that can be best explained by dividing data
into subgroups of similar items, whereby each member within such a found group is similar to
all the others and different from the members of other groups. The similarity is defined by a
problem-specific metric like for instance the euclidean distance between data items. Cluster
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2190

analysis belongs to the so-called unsupervised learning methods since the input data items
X
i
=<x
1
,...,x
d
>∈ R
d
with i =1...N are given, but no desired results or ground truth to
be learned from
1
.
In this survey we focus driven by our outlier use case on the DBSCAN clustering algorithm.
The
literature offers several publications describing parallelization approaches, e.g. based on
the map-reduce paradigm [12], using distribute dR-Trees [5], or more traditional variants in the
context of databases [13]. However, except of one, none of these offer concrete implementations
or point to respective source code repositories. We could therefore only include PDSDBSCAN-
D [15] in our survey despite the list of frameworks introduced in the method section.
Technology Platform Approach Supports DBSCAN Parallelization Stable
Weka Java yes no yes
R R yes no yes
Matlab Matlab no no yes
Octave Octave no no yes
Apache Mahout Java, Hadoop no yes yes
MLlib/Apache Spark Java, Spark, Hadoop no yes yes
scikit-learn Python yes no yes
PDSDBSCAN-D C++, MPI, OpenMP yes yes no
HPDBSCAN C++, MPI, OpenMP yes yes yes
Table 1: Overview of open and freely available DBSCAN clustering tools
Table 1 shows the results of our survey. The new ’big data’ technology stacks Apache
Mahout and Spark do not offer an implementation of DBSCAN to begin with. Interestingly
enough the same is true for the well-known data mining tools Matlab and Octave that do not
support this kind of analysis. In contrast to that Weka, R and scikit-learn have packages that
allow to cluster data using DBSCAN. However, all of them are not parallelized and therefore do
not have the ability to scale to modern HPC systems. Only the one remaining implementation
satisfies the ’(a) open and free availability’ criteria as well as the ’(c) suitability of algorithms’
criteria – PDSDBSCAN-D. A deeper analysis of the C++ code based on MPI or OpenMP
reveals drawbacks in terms of ’(b) technical feasibility’ in terms of scalability and speedup
while performing tests with real world data [9]. In order to satisfy our demanding earth science
outlier case study we have implemented a more efficient parallel version of DBSCAN that we
named HPDBSCAN. Our approach differs from the PDSDBSCAN-D implementation in several
ways such as a smart preprocessing into spatial cells as well as density-based chunking to load
balance the local computation.
3.2 Survey of SVM Classification Tools
Classification is the problem of identifying the membership of data items into a set of subgroups
called classes. In contrast to clustering, the statistical model for this is learned from known
observations or samples of class memberships. Classification therefore belongs to the supervised
learning methods since input data items X
i
=<x
1
,...,x
d
>∈ R
d
with i =1...N are given
in combination ’with supervising output’ data y
i
. In our survey we focus on the support vector
machine (SVM) classification algorithm, including its kernel methods allowing to classify non-
linearly seperable data.
1
The variable d equals the number of dimensions or features and N the number of different samples
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2191

Technology Platform Approach Multiclass Supported Kernels Parallelization Stable
Weka Java yes linear, rbf, polynomial, sigmoid no yes
R(kernlab) R yes linear, rbf, polynomial, sigmoid no yes
Matlab Matlab yes linear, rbf, polynomial, sigmoid no yes
Octave Octave yes linear, rbf, polynomial, sigmoid no yes
Apache Mahout Java, Hadoop - - - -
MLlib/Apache Spark Java, Spark, Hadoop no linear yes yes
scikit-learn Python yes linear, rbf, polynomial, sigmoid no yes
libSVM C, Java yes linear, rbf, polynomial, sigmoid no yes
Twister/ParallelSVM Java, Twister, Hadoop no linear, rbf, polynomial, sigmoid yes no
pSVM C, MPI no linear, rbf, polynomial yes no
GPU LibSVM CUDA yes linear, rbf, polynomial, sigmoid yes (rbf) yes
πSVM C, MPI yes linear, rbf, polynomial, sigmoid yes yes
Table 2: Overview of open and freely available SVM classification tools
As shown in Table 2, there is a wide variety of candidate solutions tools available that vary in
terms of their platform approach. In terms of ’(c) suitability of algorithms’,wehavefoundonly
four of the different implementations to be useful. The reasoning is as follows. Weka, R, Matlab,
Octave and scikit-learn, as high-level languages also attractive to non technically savy scientists,
wrap libSVM as a base implementation. Due to the lack of parallelization of the latter, however,
it is not suitable for our earth science data mining problem. MLlib of Apache Spark seemingly
overcomes this by reimplementing SVMs using the scalable and parallel Spark core, but a deeper
analysis show that it only supports linear SVMs. Again, this is not applicable for our earth
science data mining use case as we have non linearly separable classes. The remaining four
satisfy basically the ’(c) suitability of algorithms’ requirement, but a deeper analysis reveals
further drawbacks in terms of (b) technical feasibility. The implementation of pSVM [20] as
well as Twister [18] are unstable beta releases that are not meant to be used in production
yet. While using Twister for example we found dependencies to messaging systems, scheduling
issues, and other problems related to the lack of features that have made an analysis of our
data challenging. Finally, there are only two parallel implementations left. The GPU LibSVM
bears lots of potential for future use, but due to its dependencies to the proprietary Compute
Unified Device Architecture [14] technology stack, we have concerns regarding our ’(a) openly
and free available’ criteria. For this reason, we have decided to test πSVM 1.2 (and indirectly
the
recent πSVM1.3) with our data. Overall the performance has been acceptable, but some
scalabilit
y issues, have lead us to optimize it further, which we will present in section 4.2.
4 Parallel and Scalable Methods
This section gives insights about our parallel and scalable methods and their optimization
structured alongside the approaches raised from the two scientific case studies introduced in
Section 1. Beside technical details, we also discuss in this section the HPC environment benefits
that satisfy the majority of our key requirements defined in Section 2 (i.e. R1-R5).
4.1 Parallel and Scalable Clustering with HPDBSCAN
DBSCAN is a density-based clustering algorithm and its principal idea is to find cluster cores
in a data set and subsequently expand these recursively. Thereby, a cluster core is defined
as region that contains at least a parametric number minPoints of neighboring point within
a given search radius epsilon (ε) and with respect to a distance function dist. For each of
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2192

these neighboring points the same criterion is reapplied in order to extend the found cluster.
Highly parallelizable DBSCAN (HPDBSCAN)
is our parallel implementation of DBSCAN.
It overcomes the inherent sequential processing step of the recursive expansion through the
follo
wing three major techniques.
Spatial Indexing Structure Common techniques for indexing data points are dR- or kd-
trees [17] as they have been used in other parallel DBSCAN research work. These significantly
speeds up region queries, especially also for dynamic search radius. In DBSCAN, however,
we have a constant search radius equaling to epsilon. Based on this fact, we have chosen a
different spatial indexing approach for HPDBSCAN, which uses a sorted, regular n-dimensional
hypergrid indexed by hash tables that overlays the n-dimensional data space (cf. Figure 1).
Similarily to the ’big data’ array databases rasdaman or SciDB [1], we can answer queries based
on this approach faster, in armortized O(1) complexity and can cash and preload queries.
Quadratic
Split Heuristic Instead of distributing the data items in equal-sized chunks
to all processor, like introduced in previous DBSCAN research works, we use a heuristic to
achieve better load balancing, especially for spatially skewed data. Since the computation time
of DBSCAN scales in a quadratic manner with respect to the point density, we calculate a
score value for each cell of the hypergrid, which take advantage of this scaling behavior. The
scores allows us to find exact splittings of the hypergrid, such that no communication is required
during the parallel computation step. At the same time, we able to balance the workload evenly
among processors.
Differential Merging Scheme HPDBSCAN is able to combine the sub-results of the par-
allel clustering into one common view. Therefore, it exchanges the direct bordering region
halos of the hypergrid splits, finds deviating cluster labels, and rewrites them accordingly. In
the merging step, we store conflicting labels in dictionaries, so that we can apply relabeling of
points in parallel afterwards. In contrast, to other parallel DBSCAN implementations such as
PDSDBSCAN-D [15] there is no additional communication or clustering needed during the
re-labeling, which increases the computational performance tremendously.
We have realized HPDBSCAN as a MPI and OpenMP hybrid in C++ that is usable as a
standalone command line interface (CLI) tool as well as a shared library that can be wrapped
or linked to bigger applications if needed. In case of the CLI program, data points are provided
in form of Hierarchical Data Format (HDF5) files [11], in which the clustering results are also
written back. These clustering results per point are either the identity of a specific cluster or
an outlier mark that in particular is useful for our earth science case study.
In order to demonstrate its scalability and its outlier detection potential with large volumes
of data, we are filtering a 3D point cloud of the old town of Bremen for outliers, such as
false readings or points which capture the inside of a building. The points cloud contains
over 81 million individual points [10], clustering these sequentially would require days. Using
HPDBSCANand the supercomputer JUDGE at the research center Jlich, we can cut this
time down to minutes. Figure 2 shows an example of the point cloud in a version that has
been clustered using HPDBSCANwith the noise points still in, but colored in red. These can
be automatically removed, allowing us to start our next analysis steps. Our implementation
takes advantage of (R1) parallel file systems and large storage capacity existing in many HPC
environments. We a are able to fully exploit (R2) scalable standard data formats that take
advantage of parallel/IO due to the adoption of HDF5. The code is us sing (R3) standard
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2193



1 2
3
4
56
7
8
910
11 12
2 8 32 128 512
1
2
4
8
16
32
64
128
256
512
number of cores
speed up
Hybrid b
Hybrid t
MPI b
Linear
Figure 1: Left: Schematic illustration of the HPDBSCAN overlayed hypergrid in a two di-
mensional problem. The points are indexed by hash table pointing to each grid cell that has a
side length of ε; Right: Speed-up and scaling of HPDBSCAN.
Figure 2: Left: Result ts of HPDBSCAN including noise reduction; Right: Identified clusters
that are illustrated using different colors.
communication protocols and enables the use of (R4) fair and stable scheduling tools and policies
to enable the work on clusters for several users at the same time. Given that HPDBSCAN
is open source [9] it belongs to the (R5) open source tools and open referencable data in
PANGAEA [7] and Bremen [10], respectively, enables reproducibility of findings.
4.2 Parallel and Scalable Classification with πSVM
In this section we present our earth science classification use case in remote sensing preceded
by some technical details of the used SVM software stack. As introduced in section 3.2 we
have based our parallel SVM implementation on πSVM version 1.2. As of late 2014 version 1.2
is outdated, though, as its successor version 1.3 has been released. For time reasons we were
unable to port our changes to the new version, but our outlined performance advancements are
still valid, due to the fact that πSVM 1.3 does not optimized the parallel code sections.
As shown in Figure 3, our optimized πSVM implementation scales better in constant sized
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2194

124816 32 64 128
200
400
600
memory access fault
number of cores
execution time in s
optimized πSVM
πSVM
124816 32 64 128
2
8
32
128
memory access fault
number of cores
speedup
optimized πSVM
πSVM
Linear
Figure 3: Comparison between the original πSVM version and its optimized version (including
memory
access fault). Left: Execution time as a function of doubling processing cores count.
Right: Speedup values of constant-sized problem.
P1
P2
P3
P4
P1
P2
P3
P4
P1
P2
P3
P4
P1
P2
P3
P4
P1
P2
P3
P4
P1
P2
P3
Figure 4: Parallel memory access patterns in a triangular matrix using four processing cores.
From left to right: initially implemented in πSVM, naive equal chunking, implemented in our
optimized πSVM tool, theoretically ideal pattern.
problems
compared to the original πSVM implementation. While we still work on achieving
b
etter scalability (e.g. consideration of a hybrid implementation using OpenMP), the code has
proven to work with our earth science classification case study. One of the major performance
gains has been achieved through the usage of more suitable MPI collective operations. For
example,
in some source code sections a simple for loop was used to perform an MPI
Bcast()
numerous times to inform other processors about training updates. Instead we replaced them
with
singular MPI
Allgather() calls reducing it from linear to constant time and memory com-
plexity. A couple of other optimization strategies have been implemented. One of them is an
improved matrix distribution pattern during the SVM training stage to balance the computa-
tional load, schematically illustrated in figure 4. The unoptimized πSVM code distributes them
as
depicted in the left picture, leaving processing cores idle. We have improved this distribu-
tion pattern step-wise from a naive equal-sized chunking, to an skip-row approach and finally
to a theoretically ideal on the right. In tests, however, we have found that the skip-row ap-
proach outperforms all other patterns, due its simple implementation using one MPI collective
operation, in contrast to the complex pattern required for the theoretically best.
We have used our parallel and scalable πSVM implementation in different classification
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2195

studies in the field of remote sensing, e.g. on remote sensing image classification of the city
Rome that we published at a remote sensing domain-specific conference IGARSS 2014 [4]. Since
then we have shown of the scalability of the code by investigating more complex problems,
such asthe Indian Pines dataset with over 50 classes. The dataset is openly available through
EUDATs B2SHARE services [4] which enables reproducibility of the papers findings.
Our optimized πSVM implementation takes advantage of (R1) parallel file systems and large
storage
capacity existing in many HPC environments and is based on MPI thus using (R3)
standard communication protocols. The code can be used with typical HPC batch schedulers
enabling the use of (R4) fair and stable scheduling tools and policies to enable the work on
clusters for several users at the same time. Given that πSVMisopensourceaswellasour
optimized
version the (R5) open source tools and open referencable data enables reproducibility
of findings. Potential for future work is to enable the use of parallel I/O that is currently not
supported (cf. R2).
5 Conclusion
One of the major findings of this paper is the need for parallel and scalable data mining tools
in earth sciences, as has been demonstrated by the means of the two case studies. While we
face these challenges in machine learning already now, the expected increase of ’big data’ in
science in the coming years, will amplify this need even more. With our two parallel data mining
tools, HPDBSCAN and πSVM, we start to tackle this problem, as both of them show significant
impro
vements compared to their serial counterparts. However, they still only utilize a moderate
number (several dozens to hundreds) of cores and do not yet fully exploit the capabilities of a
large HPC supercomputer.
Given the page restriction, our study also did not describe all potential parallel optimization
points entirely in detail. One straightforward example is the use of cross-validation for model
selection as it is implemented in πSVM. Another example is the so-called grid search technique,
whic
h is a highly computational intensive, and that typically aims to explore the right combina-
tion of machine learning algorithm parameters (i.e. here for the SVM and its rbf kernel). Hence,
easy to implement parallelizations of this processes should enable additional massive speedups
compared to serial implementations making thus the use of parallel and scalable methods even
more feasible.
Future work beyond using the ever increasing amounts of datasets faced in earth sciences,
is the exploration of in situ analytics towards exascale computing. In other words the exascale
simulation (e.g. of a climate model) is running while another part of the machine (e.g. using
GPGPUs or accelerators) is performing analytics in situ in order to validate the simulation
with real measurement data and to perform statistics on the fly on the created simulated data.
This bears the potential to abort costly computation runs early based on the statistical data
mining that will take place concurrently to the particular numerical simulation.
Acknowledgments
We would like to express our thanks to the EUDAT European data infrastructure in general
and the B2SHARE service support team in particular for providing us with storage space and
handles to properly link research data.
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2196

References
[1] Peter et al. Baumann. The array database that is not a database: file based array query answering
in rasdaman. In Advances in Spatial and Temporal Databases, pages 478–483. Springer, 2013.
[2] James B Campbell. Introduction to remote sensing. CRC Press, 2002.
[3] Gabriele et al. Cavallaro. A comparison of self-dual attribute profiles based on different filter
rules for classification. In Geoscience and Remote Sensing Symposium (IGARSS), 2014 IEEE
International, pages 1265–1268. IEEE, 2014.
[4] Gabriele et al. Cavallaro. Smart data analytics methods for remote sensing applications. In
Geoscience and Remote Sensing Symposium (IGARSS), 2014 IEEE International, pages 1405–
1408. IEEE, 2014.
[5] Min et al. Chen. Parallel dbscan with priority r-tree. In Information Management and Engineering
(ICIME), 2010 The 2nd IEEE International Conference on, pages 508–511. IEEE, 2010.
[6] Corinna et al. Cortes. Support-vector networks. Machine learning, 20(3):273–297, 1995.
[7] Michael et al. Diepenbroek. PANGAEAan information system for environmental sciences. Com-
puters & Geosciences, 28(10):1201–1210, 2002.
[8] Martin et al. Ester. A density-based algorithm for discovering clusters in large spatial databases
with noise. In Kdd, number 34 in 96, pages 226–231, 1996.
[9] Markus G¨otz. Clustering - HPDBSCAN. http://www.fz-juelich.de/ias/jsc/EN/Research/
DistributedComputing/DataAnalytics/Clustering/Clustering_node.html, 2015. [Online; ac-
cessed
Feburary, 10th 2015; 11:14 CET].
[10] Markus et al. G¨otz. HPDBSCAN Benchmark test files. URL:https://b2share.eudat.eu/record/
178, 2015. [Online; Accessed Febuary, 9th 2015; 12:09 CET].
[11]
HDF Group. Hierarchical data format version 5. http://www.hdfgroup.org/HDF5, 2015. [Online;
Accessed
Febuary, 10th 2015; 14:17 CET].
[12] Yaobin et al. He. MR-DBSCAN: an efficient parallel density-based clustering algorithm using
mapreduce. In Parallel and Distributed Systems (ICPADS), 2011 IEEE 17th International Con-
ference on, pages 473–480. IEEE, 2011.
[13] Eshref et al. Januzaj. Scalable density-based distributed clustering. In Knowledge Discovery in
Databases: PKDD 2004, pages 231–244. Springer, 2004.
[14] CUDA Nvidia. Compute unified device architecture programming guide. 2007.
[15] Md Mostofa Ali et al. Patwary. A new scalable parallel DBSCAN algorithm using the disjoint-set
data structure. In High Performance Computing, Networking, Storage and Analysis (SC), 2012
International Conference for, pages 1–11. IEEE, 2012.
[16] Sangeeta et al. Rani. Recent techniques of clustering of time series data: A Survey. Int. J. Comput.
Appl, 52(15):1–9, 2012.
[17] Hanan Samet. The design and analysis of spatial data structures, volume 85. Addison-Wesley
Reading, MA, 1990.
[18] Zhanquan Sun and Geoffrey Fox. Study on parallel SVM based on MapReduce. In International
Conference on Parallel and Distributed Processing Techniques and Applications, pages 16–19. Cite-
seer, 2012.
[19] O Team. Big Data Now: Current Perspectives from OReilly Radar, 2011.
[20] Edward Y Changf Kaihua et al. Zhu. PSVM: Parallelizing Support Vector Machines on Distributed
Computers.
On Scalable Data Mining Techniques for Earth Science G¨otz et. al.
2197