You are on page 1of 12

Annals of GIS

ISSN: 1947-5683 (Print) 1947-5691 (Online) Journal homepage: http://www.tandfonline.com/loi/tagi20

Spatial statistics in geographical information


science: from interpolation to probabilistic
robotics

Alfred Stein

To cite this article: Alfred Stein (2010) Spatial statistics in geographical information
science: from interpolation to probabilistic robotics, Annals of GIS, 16:4, 211-221, DOI:
10.1080/19475683.2010.539986

To link to this article: https://doi.org/10.1080/19475683.2010.539986

Published online: 16 Dec 2010.

Submit your article to this journal

Article views: 381

View related articles

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=tagi20
Annals of GIS
Vol. 16, No. 4, December 2010, 211–221

REVIEW ARTICLE
Spatial statistics in geographical information science: from interpolation to probabilistic robotics
Alfred Stein*
Department of Geoinformation Science and Earth Observation (ITC), Twente University, Enschede, The Netherlands
(Received 2 September 2010; final version received 10 November 2010)

This review article presents an overview of spatial statistical procedures that are of use and value in geographical information
science. Spatial statistics considers the handling of spatial data, with an emphasis on modeling and on dealing with uncertainty.
Spatial statistics includes issues of interpolation, point statistics, and sampling. Within a robotics context we distinguish
localization, mapping, and decision control and support. Localization is based on advanced image analysis in which the idea of
image mining plays an important role. Image mining considers the chain from object identification on natural or man-made
processes from remote sensing images through modeling, tracking on a series of images and prediction, toward communication
to stakeholders. Geographical information science serves as a scientific field, for example, to generate a platform for storing
observed and collected data, analyzing them, and displaying the results. This article shows how within the domain of robotics
there are novel and intriguing possibilities to use spatial statistics and image mining. The article gives a concise overview of
these procedures and several ways ahead that are of value within a robotics context. It is illustrated with a field study on robotics
in agriculture.
Keywords: spatial statistics; image mining; interpolation; robotics; spatial data quality

1. Introduction dedicated simulation studies. Finally, with increasing com-


Geographical information science (GIS) has traditionally puting facilities it has become more obvious than ever that
maintained close links with spatial statistics (Goodchild spatial objects could be uncertain and vague, and showing
1999, Burrough 2001). Goodchild (1999), when addressing internal spatial variation. Spatial statistics may be able to
uncertainty in geographic analysis and GIS-based analysis, quantify this uncertainty and assist in analyzing conse-
identifies the need to make spatial statistical methods more quences of uncertainty during spatial data processing. All
available to the general users. There are many reasons for these topics require a skillful use of spatial statistics as was
the close and necessary link. Traditionally, information has identified explicitly as early as in 1986 (Burrough 1986).
been present in the form of point data, whereas an informa- Spatial statistics, as a science on its own, may be traced
tion layer in an geographic information system requires a back to the early 1900s, when Mercer and Hall (1911) had
full coverage of the area to be able to interpret or to overlay offered the landmark starting study. It has developed rapidly
with other layers. Creating those layers thus requires the in the past 50 years. A main development was the need to
interpolation of points to areas. interpolate data by making predictions at unsampled loca-
Spatial statistics is fundamental during a quantitative tions, in particular when analyzing agricultural, geological,
spatial analysis. When analyzing spatial data, for example, hydrological, and soils data. Related issues on spatial mod-
to build relations between soil and crop data, ordinary eling, sampling, and simulation have been addressed. These
regression has been used throughout. As ordinary regres- procedures are well documented and generally applicable
sion ignores the presence of spatial dependence, more (Gelfand et al. 2010). A specific focus on remote sensing
advanced spatial regression methods have been developed. images started with the work of Curran (1988) in the 1980s
Similarly, image analysis methods currently offer advanced and 1990s (Stein et al. 1999). Somehow, the set of problems
methods of carrying out a (semi)automatic object identifica- initially appeared to be limited; as remote sensing images
tion and classification, whereas fuzzy classifiers allow a already display much information, pixels have a nonpoint
user to include some issues of uncertainty as well. From a support size and the interpolation below clouds was only
more advanced point of view, we notice an increasing realistic for relatively small clouds (Curran and Atkinson
attention for point pattern and stochastic geometry analysis. 1998, Addink and Stein 1999). To extract information quan-
At the user interface, one increasingly needs uncertainty titatively, a disciplinary approach was preferred, and more
information in decision making, for example, leading to classical statistical procedures such as nearest neighbor

*Email: stein@itc.nl

ISSN 1947-5683 print/ISSN 1947-5691 online


© 2010 Taylor & Francis
DOI: 10.1080/19475683.2010.539986
http://www.informaworld.com
212 A. Stein

clustering and principal component analysis were useful to summarize the data in novel ways that are both under-
and fruitful to address the major questions that a user standable and useful to stakeholders’ (Stein 2008). Objects
would put to images, for example, segmentation and classi- discernable on those images can be either crisp or fuzzy and
fication. Moreover, uncertainty generating disturbances to vague. We distinguish five important steps in image mining:
images have been addressed by developing models to han- identification, modeling, tracking, prediction, and communi-
dle atmospheric conditions. cation with stakeholders (Figure 1). All these processes will
This century has seen a further increase in the develop- be briefly discussed below. On top of this, we notice aspects
ment, the applications, and the use of spatial statistical of quality control in each of these steps.
procedures within geographical information science. We Identification of objects corresponds to combining an
notice the advance of lattice type approaches in advanced observed pattern in an image with a set of signals. This
classification, in super pixel resolution mapping, and in would mean that in an area a homogenous set of pixel values
drawing inferences from images. The geostatistical occurs, pointing to an object of interest. The next step
approach increasingly focuses on prediction of missing concerns the extraction of the object. This is usually done
pixels and simulation of values for use in GIS-based mod- by means of a segmentation, followed by a classification.
els. A point pattern analysis is more recently coming of age Extraction of objects makes the step from raster to objects.
when realizing that many features derived from remote Typically, extraction is done by applying a segmentation
sensing images essentially have a point shape. Finally, routine in which both the object and the uncertainty are
statistical methods for spatial sampling have found their modeled. Various procedures for image segmentation are
place when ground control points can be freely located. well documented, and include procedures based on mathe-
The analysis is largely determined by the availability of matical morphology, edge detection and identifying homo-
reliable data and of the availability of solid software geneity in one band or in a set of bands (Glasbey and
packages. Modern ArcGIS facilities now incorporate Horgan 1995). Modeling of uncertainty has been done in
much of the spatial statistical procedures. More recently, the past by using, for example, a confusion index, whereas
image mining has been introduced as an integrated proce- traditionally a discriminant analysis could be applied, hon-
dure to extract relevant information from large sets of oring the presence of different spectral bands and yielding
images. It thus extends the general GIS notion of data posterior probabilities to the non-selected classes. The result
collection and analysis. of this operation for an image is a series of n objects
A modern development that we will consider at some Ωi;t ; i ¼ 1; : : : ; nt , that are characterized by similar pixel
level of detail in this article concerns machine learning and values, which are different from pixel values in the vicinity.
robotics. Robotics concerns the optimal programming of, A segmentation and classification step in image mining
for example, a field robot that maneuvers in a well-defined delineates objects and assigns a label to it from a class of
area (Thrun et al. 2005). Such maneuvering requires collec- possible labels. Segmentation and classification lead tradi-
tion and analysis of spatial information on an isolated plat- tionally to crisp objects. At each moment t, the centroid μ(t)
form. Such information may be available when the robot may be given as a pair of coordinates, and the boundary b(t)
starts operating, whereas updating of spatial information as a set of points or as a parametric curve. We increasingly
may be convenient during its processing. A robot therefore realize, however, that many real-world objects are not crisp
may have onboard a GIS-type information center, contain- by nature, but that they are fuzzy or vague, either because of
ing images and maps of the area of operation and may their definition, their transition zones to neighbors or inher-
benefit from image-taking and analyzing capacities during ent variability. They thus have an uncertain boundary and/or
its operation. In particular, issues of machine learning are an uncertain extent. Modeling of identified objects as
important for this purpose (Bishop 2006).
The aim of this article is to show how existing methods
within GIS, and with some emphasis on spatial statistics, are
Object
suited at this moment to enter a new stage in GIS, namely identification
Modeling
their implementation within a robot environment. The arti- External factors
cle is organized as follows. It starts with a short description
of image mining, next it describes spatial statistical proce-
dures and image analysis, and in the end it addresses the Images
Prediction
combination of the two lines within a robotic environment.
Tracking
in time
t0
2. Image mining t1,...,tn

We start the overview by considering image mining. Image Quality control Stakeholders
mining is defined as ‘the analysis of (often large sets of)
observational images to find (un)suspected relationships and Figure 1. Image mining of uncertain objects.
Annals of GIS 213

uncertain objects Ω̃i;t requires either specification of inten- t + 1. Prediction results in the values of the membership
sity values, or as parameters of a membership function vi to function, of the centroid, and of the associated uncertainties.
a class Ωi of interest (Dilo et al. 2006). Predicted values are typically of interest to stakeholders,
The next stage is to model their behavior in space and who may thus have a tool to support their decision making.
time (Dubois and Jaulent 1987). As long as the object is Communication to stakeholders can be manifold. It may
characterized by a limited set of parameters, such tracking range from simple visualization tools toward assessments of
may be relatively straightforward. For a crisp object the costs and benefits. Issues from decision support typically
centroid μ(t) may be given at each moment t as two coordi- are required here. Recent developments have focused on the
nates, and the boundary b(t) as a set of points or as a use of Bayesian methods in particular in a fuzzy set context.
parametric curve. These could be followed in an interval T Stein and Van de Vlag (2006) built a fuzzy decision tree by
as a function of t 2 T. Object tracking becomes more considering a range of objects, which when combined pro-
complicated for uncertain objects splitting and merging of duced the objects of interest. In their case it was beach
objects require some special attention. The splitting of objects, whereas in the flooding study described above
object Ωt,1 at moment t leads to two new objects Ωtþ1;1 this would be flooded objects. Decision making applied in
and Ωtþ1;2 at moment t + 1. Both objects require member- this way provided essential information in a spatial context,
ship functions, defined on the basis of the membership whereas we found the combination with a probabilistic
function at t as well as the characteristics of the new objects. approach with fuzzy methods to be interesting as a relation
Inversely, merging of two objects at moment t toward one between real-world objects and remote sensing images
single object at moment t + 1 requires first that the member- leads to improved object delineation that could be commu-
ship functions are being combined into one new member- nicated to stakeholders, including the uncertainties.
ship function, and also that the two existing centroids result
into one new centroid.
After tracking successfully, the next stage of image
mining is to predict the object at the moment t0. For a single 3. Spatial statistics
object, one way to proceed is to define a parametric curve as In the overview below we shall make a distinction between
a function of time for μðtÞ; t 2 T and parameters identify- different ways of analyzing remote sensing images. We will
ing b(t). Similarly, for fuzzy objects a parametric curve may use the following notation throughout: the image is denoted
specify the parameters describing the membership function by I(.), a pixel location in the image is denoted by s 2 I, and
νðtÞ; t 2 T . These curves should then be used to predict the a pixel value at location s is denoted by p(s). The general
most likely parameters at t0 beyond the moments of obser- approach will be that pixel value show a spatial relation. We
vation so far. Rajasekar et al. (2006) showed the use of a explored this on a spot image from the Enschede area. These
linear statistical model, whereas particle filtering methods relations can be shown by comparing pixel values with
are of a possible use here as well. In prediction, one may neighbors, and calculating correlations. Of some particular
consider a future event, that is, a real prediction, or one may be that these relations need not be isotropic: relations
moment prior to image availability, like predicting the with left and right neighbors may be different from relations
moment that the object is born. Also, prediction is some- with their upper and lower neighbors (Figure 2b).
times required of the object between two moments t and Correlations are 0.911 and 0.883, respectively.

(a) (b)
100

100
80

80
60

60
40

40

40 60 80 100 40 60 80 100

Figure 2. Scatter plots between pixels and their right neighbor (a) and their lower neighbor (b). Correlations are 0.911 and 0.883,
respectively.
214 A. Stein

3.1. Interpolation Geostatistics has recently become of an increased inter-


Interpolation has been addressed in remote sensing image est in the link between deterministic model output and
mining. A common interpolation procedure requires the use images for remote sensing images. In fact, images may
of geostatistics, that is, that part of spatial statistics that deals display information reflecting patterns that are represented
with continuous data in space, collected at arbitrary points by specific bands or their combinations. Such models range
(Cressie 1991; Gelfand et al. 2010). At each pixel location from statistical models toward complicated deterministic
the observation is considered as a realization of a random models. (Van de Kassteele et al. 2006) applied geostatistics
variable. We may use a basic notation as P(s) with s 2 I, to interpolate concentrations of ground-based particulate
where we now use the capital P to indicate that the pixel matter (PM10) measurements over Western Europe. The
value is a random variable. Pixels show spatial dependence, study uses uncertain secondary information of aerosol opti-
as neighboring pixels are more similar than pixels at a larger cal thickness from MODIS satellite observations and from a
distance. The spatial dependence is usually modeled by the chemical transport model. An inherent limitation exists to
covariance function, defined as the covariance between the spatial representativeness of ground-based measure-
pixel values given by the distance h, and denoted by C(h). ments. Validation showed that adding secondary informa-
Geostatistics is of use in a GIS and remote sensing image tion from either the chemical transport model or the satellite
analysis context for various reasons. First, it is a tool to model observations improved the PM10 mapping. The RMSE
the spatial variation of point data collected in a GIS by means decreased with 17% and 11%, respectively, whereas a com-
of variograms and covariance functions (Curran 1988). bination of both sources of secondary information resulted
Second, it may be used to interpolate missing or suspicious in a reduction in the RMSE value with 30%.
values (Curran and Atkinson1998; Addink and Stein 1999).
Typical examples are pixel values below clouds, anomalies
on an image generated by sensor characteristics, or where an 3.2. Image analysis
isolated object (e.g., a tree) is disturbing a general pattern of Image analysis can be largely based on the seminal work of
an image. Third, it could be applied to downscale the image to Geman and Geman (1984) and Besag (1986) in the 1980s.
a finer resolution (Atkinson et al. 2008). Finally, geostatistics Those publications assume the existence of a multispectral
provides a tool for spatial simulation, which is important image x, that consists of the same spectral bands as I, and is
when simulating and visualizing spatial model calculations. acquired at the same spatial resolution as a classified
Geostatistics has plaid a specific role in image fusion. In image c. Image x is not observed directly, whereas image I
image fusion several images have to be overlaid, and these is observed at a coarser resolution and is assumed to be a
images may be shifted or rotated (or both) with respect to degraded observation of the scene. Furthermore, it is
each other, or they may have slightly different resolutions. assumed that each pixel in x can be assigned to a unique
In particular when an image at a course resolution has to be class: cðxj Þ ¼ α, where α 2 f1; 2; : : : ; Lg. The relation-
overlaid with an image of a low resolution, spatial interpo- ship between x and I is established by means of a degrada-
lation is playing a major role. The principle is relatively tion model. The degradation model for I at pixel p(s) given
straightforward: the coarse resolution pixel values are inter- image x is then given by
polated toward the grid nodes of the low resolution image.
Also, multitemporal image analysis may benefit from geos-
1X
2
S
tatistics using its facilities either to tune the images to each IðpðsÞÞ ¼ xðajjpðsÞ Þ (1)
2
S j¼1
other or to predict values at intermediate times where no
images were available but where somehow interest in a
physical process exists. where S equals the ratio of the resolutions in x and I
As an example, we have interpolated the pixel values at (Tolpekin and Stein 2009). For simplicity we exclude partial
points that were apparently covered by a cloud on a spot overlapping between coarse and fine resolution pixels. For
image (see Figure 3). Figure 3a shows the cloud, whereas each pixel xjjpðsÞ , we define a symmetric neighborhood
Figures 3b–d show interpolation using the mean, using an N ðxjjpðsÞ Þ on x as the set of all pixels inside a square window
inverse distance interpolator, and using ordinary kriging, with the center on the pixel xjjpðsÞ , except for the pixel itself.
respectively. Apparently, use of the mean gives too crude an The neighborhood N ðxjjpðsÞ Þ is defined completely when the
approximation, whereas both inverse distance interpolation window size S is specified. Such a neighborhood system is
and kriging show more variation. Inverse distance interpo- symmetric in the image except at the edges. Popular choices
lation has as a disadvantage that it usually shows isolated are the first-order neighborhood, that consists of the four
island type features around incidentally high pixel values, closest connected pixels, and the second-order neighbor-
whereas kriging has the major advantage that its choice of hood, that consists of the eight closest connected pixels.
weights is largely determined by the properties of the spe- Note that the second-order neighborhood corresponds to
cific band. our neighborhood system with S ¼ 3. Markov random
Annals of GIS 215

(a) (b)

1.0

1.0
0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

(c) (d)
1.0

1.0
0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Figure 3. Spatial interpolation below a (simulated) cloud (a). (b) shows interpolation by the mean of the data, (c) by inverse distance, and (d)
by ordinary kriging.

field image analysis models the classified image c as a emerge because of stochastic influences. Statistical infer-
Markov random field with neighborhood system N ðxjjpðsÞ Þ. ence for X is based on distances, either those between a
Because of the equivalence of Markov random fields and fixed reference point pðs0 Þ 2 I and the points of the process
Gibbs random fields, the posterior probability for each Xi, or between the points of Xi themselves.
configuration c given observed image y can be specified Thus, for an arbitrary pixel pðsÞ 2 I, let ds denote the
by means of a posterior energy function (Geman and Geman distance from p(s) to the nearest point of X. Then the empty
1984). space function of X for r > 0 equals

FðrÞ ¼ PrðdðpðsÞ; X Þ  rÞ (2)


3.3. Point patterns analysis
A statistical point pattern analysis is becoming of an The empty space function equals the probability of obser-
increasing importance in image analysis. It is commonly ving at least one point closer than r to the arbitrary point
assumed that a spatial process X generates a point-like marked as p(s) in the area. Under the assumption of statio-
pattern. Typical examples in remote sensing images include narity, F(r) does not depend upon p(s). The heuristic expla-
landslides within a mountainous area, lakes in a lake abun- nation of 1 – F (r) is the probability that a circle with radius
dant area, and freestanding houses on a high-resolution r placed at random in the area does not contain a point of X,
image. To describe a spatial point pattern, we consider X as p(s) can be chosen arbitrarily. A completely spatially
as a stationary point process on I. We take stationary pro- random (CSR) pattern of points with density λ shows an
cesses as the starting point whereas for non-homogeneous F-function equal to Fref ðrÞ ¼ 1  expðπλr2 Þ. A distribu-
populations we refer to Gelfand et al. (2010). Stationarity tion of clustered points has an F - function below Fref (r), as
means that the position of the points is independent from a on the average for short distances fewer points are encoun-
shift in the image, although differences in densities may tered than for a random pattern, whereas a regular pattern
216 A. Stein

has an F - function above it, as on the average more arbitrary The basic idea of design-based sampling is that each point
points encounter a point of the process X. Turning to inter- in an area A has the same probability of being sampled.
point distances, the nearest neighbor distance function G(r) Sampling points are assigned to locations within A. We dis-
is the distribution function of the distance from a point of the tinguish here three examples: complete random sampling
process X to its nearest neighbor, (Figure 4a), cluster sampling (Figure 4b), and aggregate
sampling (Figure 4c). We suppose that the edge of A is
available as a polygon and that the number n of sampling
GðrÞ ¼ PrðdðXi ; Xj Þ  rÞ (3)
points is decided beforehand. Then a random sampling
scheme is generated by applying a random number generator
for r > 0. It is the conditional probability distribution that
to generate x- and y-coordinates until n points are generated
the distance from the point Xi to the closest point Xj of the
that all fall within A. Cluster sampling is generated by first
process is less than or equal to r. A heuristic description of 1
distributing at random a low number of points, say n0 within
– G(r) is the probability that within a circle with radius r
A, in Figure 4b indicated by open symbols. This is followed
centered on a randomly selected point no further point
by generating within A n/n0 points around each of the pre-
occurs. Again, G(r) does not depend on the choice of Xi
vious, all at a close distance. Notice that the first set of points
because of stationarity. The empirical distribution function
is not actually sampled. A grid sampling can be implemented
(EDF) for G(r) is obtained for each distance r by counting
by overlaying a regular grid over A such that n points fall
the number of points at distances less than or equal to r. A
within A, and then moving the entire grid over a random
random point pattern with density λ shows a G(r)-function
vector such that all points are still located within A.
equal to Gref ðrÞ ¼ 1  expðπλr2 Þ. A distribution of clus-
On top of each of the random generation mechanisms an
tered points has a G(r)-function higher than this function, as
area may be stratified on the basis of available information into k
on short distances more points are encountered than for a
strata Aj ; j ¼ 1; : : : ; k. Available boundaries are determined
random pattern, whereas a regular pattern has a G
and n1 points are assigned to A1, n2 points to A2, and so on. The
(r)-function below it. To combine the two, the J (r) function
distribution could follow any leading principle, such as propor-
has been developed Van Lieshout and Baddeley (1999). For
tionality to the size of the stratum (Figure 4d), defining a mini-
inter-point distances to distances with respect to a reference
mum number of points to each stratum, followed by random
point, say p(s0), the J (r)-function is defined
allocation of the remaining points, equality of points in each
stratum (Figure 4e), and so on. Such segments typically would
1  GðrÞ
J ðrÞ ¼ (4) follow from a classification of a remote sensing image, for
1  FðrÞ example, the land cover. Other examples could include, depend-
ing upon the interest of the study, a soil map, a land-use map, or a
for all r > 0 satisfying F (r) < 1. For point processes without geological map. Often a pragmatic solution is necessary to deal
any aggregation or regularity, J (r) ¼ 1, whereas J (r) > 1 with irregular boundaries.
indicates inhibition between the points, and aggregated Model-based sampling is based on the assumption of a
patterns have a J (r)-function values smaller than one. random model for the data. This assumption is similar to that
Moreover, the J (r) function is constant beyond the effective made for geostatistical interpolation (Van Groenigen and Stein
range of interaction (Van Lieshout and Baddeley 1999). 1998). Central in model-based sampling is a quantitative opti-
mization criterion: for an infinite area, without any boundaries,
a triangular scheme is well-known to be optimal. Moreover,
3.4. Sampling each criterion leads to a unique optimal scheme. Examples of
The next component of spatial statistics that may find a such criteria are: equal coverage of the study area, making a
place within a robotics context concerns spatial sampling. map with the lowest average kriging variance, and estimating
Spatial sampling is crucial when collecting reliable, that is, the variogram with an equal number of pairs of points in each
statistically sound, information. It basically answers the distance class. In practical studies, however, already collected
question where the optimal locations for information collec- data and irregular boundaries should be taken into account. As
tion should be placed. Such could be important, for exam- an example we may consider the first criterion. The aim will be
ple, for collecting validation points in remote sensing to identify the scheme S consisting of the points {xj}. To
studies or to collect optimal information to calibrate deter- evaluate any scheme ne points x̃i ; i ¼ 1; . . . ; ne , are assigned
ministic models. Below we will sketch two different to the area A. Then the aim is to minimize the sum of the
approaches toward spatial sampling: design-based sampling smallest distances of any of the ne evaluation points to a point
and model-based sampling. One major difference is that in S. This can be formulated mathematically as
design-based sampling provides different schemes at each
instance and thus provides statistically sound information,
1X ne
whereas model-based sampling provides the uniquely best fðSÞ ¼ min dðx̃i ; xj Þ (5)
xj 2S ne i¼1
sampling scheme for a particular purpose.
Annals of GIS 217

(a) (b)
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(c) (d)

1 1

0.8 0.8

0.6 0.
06

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(e)
1

0.8

06
0.

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Figure 4. Different spatial sampling schemes: (a) completely random sampling, (b) cluster sampling, (c) grid sampling, (d) stratified random
sampling with number of points proportional to the size of the strata, and (e) stratified sampling with equal number of points in each stratum.

During implementation one could start with a random scheme S(1) that thus emerges has a lower f value than the
scheme S(0) and then follow a simulated annealing schedule, original scheme. Moreover, S(1) is also accepted with a
S ð0Þ S ð0Þ
that is, a point is selected at random and moved over a probability e c if S(1) is larger than S(0). A new scheme
random vector such that it finds a new location within A. S(2) can be generated out of S(1) and this procedure continues
The new position is accepted with probability 1 if the new until convergence is reached. The parameter c reduces in
218 A. Stein

value during this optimization process. An illustration of the robot’s existing position. A Gaussian assumption is
optimization is given in Figure 5. The process of optimiza- commonly made here. Grid and Monte Carlo localization
tion is shown for n ¼ 100 points within a rectangular area A are different in the sense that they are based on grid locali-
in Figure 5a–c after 100, 500, and 1000 steps. Clearly, zation and that the resolution plays an essential role. Monte
optimization is not fully reached after 500 steps, although Carlo localization is a Monte Carlo method to determine the
the fitness function f(S) only marginally reduces in size position of a robot given a map of its environment. It is
(Figure 5d). We may notice that the fitness function shows essentially an implementation of the particle filter applied to
increase during early iterations, mainly due to accepting robot localization. The second major component in robotics
inferior schemes. This does not occur at higher iterations, concerns mapping of the environment. The mapping com-
because of the reduction in the parameter s. Figure 5e shows ponent in robotics distinguishes occupancy grid mapping,
that a procedure can easily be adapted to also include exist- and in particular acquiring maps. To combine localization
ing observation points. and mapping, combined ‘Simultaneous Localization and
Mapping’ (SLAM) algorithms have been developed.
Many of the SLAM algorithms show a strong similarity
4. Probabilistic robotics with the kriging equations that are being used in geostatis-
Robotics has evolved as a field of scientific interest (Thrun tics. Planning and control issues are based on Markov
et al. 2005). In robotics, an image has to be taken and decision processes with various extensions.
interpreted rapidly, and based on the image a decision has In agricultural field studies, robotics has developed as a
to be made. Probabilistic robotics concerns a subfield of focal activity during recent years (Brooks 1985, Polder
robotics, dealing with perception and control (Thrun et al. 2007). For a field robot, spatial statistics plays a critical
2005). Three major components are distinguished: localiza- role in the following components. In the first place a robot
tion, mapping, and planning and control. Basically, the has to orient itself between the plants, ploughing rows, and
robot has to position itself within the field, it has to map other visible components. With the current development of
its environment, and it has to make either a plan for a next GPS, this is a rather standard activity. Second, it has to move
stage or to control (or explore) its environment. Localization at the right speed into the right direction. Because of field
procedures include Markov, Gaussian, grid, and Monte irregularities, an almost continuous feedback and correction
Carlo localization. Markov localization includes the is required. Third, based on acquired information a specific
extended Kalman filter as a special case. The differences action has to be taken. In field conditions, this could be the
between them can be characterized as follows. Markov application of specific amounts of manure, of a limited
localization determines a prediction step on the basis of amount of pesticides, or of digging out weeds. Finally,

(a) (d)
6.00 3
4.00

2.00
2
0.00
0.00 2.00 4.00 6.00 8.00 10.00 12.00 14.00 16.00 18.00 20.00 22.00 24.00

(b) 1

6.00
0
4.00 0 100 500 1000
2.00

0.00
0.00 2.00 4.00 6.00 8.00 10.00 12.00 14.00 16.00 18.00 20.00 22.00 24.00

(c) (e)
7
6.00 6
5
4.00 4
3
2.00 2
1
0.00 0
0.00 2.00 4.00 6.00 8.00 10.00 12.00 14.00 16.00 18.00 20.00 22.00 24.00 0 5 10 15 20 25

Figure 5. Optimizing a spatial scheme consisting of n ¼ 10 points in a rectangular area A: (a) after 100 iteration steps, (b) after 500 iteration
steps, (c) after 1000 iteration steps. The function f(S) during optimization is given in (d). In (e) three points have been kept fixed and seven
points have been optimized.
Annals of GIS 219

rapid and reliable spatial interpolation has to be done if the provides a large set of statistical and mathematical methods
field of vision is incidentally obscured, for example, by the to automatically analyze sets of data. Typical examples
presence of a bird. include the graph-cut algorithm (Karimov 2010), Bayesian
GIS issues come up as usually, several layers of infor- networks (Mustafa et al. 2010), boosting (Xiao 2008), sup-
mation are available. There will be a DEM of the field, the port vector machines (Cristianini and Shawe-Taylor 2000),
edges and positions of the corners will be present and can be artificial neural networks (Skidmore et al. 1997), and random
used for localization, and there will be information on the forests (Breiman 2001). Currently, we notice an increasing
position of plants and ridges, in particular, if it concerns interest in machine learning applications.
plants in rows such as corn or potatoes. Based on this Aspects of machine learning seem to be fundamental in
information the robot has to maneuver and most likely to robotics. We expect a robot to learn during operation. The
cover the locations in the field within a relatively small information grasped should be stored and used subse-
period of time. quently. In the localization stage feature extraction based
An integrated approach can take care of this, for example, on segmentation and classification or on object-based clas-
as a layered structure of maps within the robot. After posi- sification could take place, and issues like boosting and
tioning itself, the robot may be starting to move in a pre- support vector machines could be considered and included.
defined direction. On the basis of images and on the basis of In the mapping stage, Bayesian interpolation procedures
collected information, it adapts its direction and speed. This would be beneficial. But in particular when considering an
requires a skillful and dedicated image analysis where the integrated SLAM procedure machine learning aspects seem
most advanced techniques should be implemented such that to be a good way ahead. However, it all depends on the
real-time information and corrections can be done. The posi- quality of the data.
tion of the robot itself requires a careful handling of posi-
tional accuracy. When deciding upon an action that is to be
taken, the attribute accuracy of an observed object in the field 5.2. Spatial data quality
(like the presence of a weed species) should be done to avoid Spatial data quality refers to various aspects of data quality
spoiling the environment with overuse of pesticides or to as can be identified for in geographical objects, where
miss crop affecting weeds. Advanced sampling may be ben- quality is defined as ‘the totality of characteristics of a
eficial, in particular, when weeds take preferential positions product that bear on its ability to satisfy stated and implied
(Heijting et al. 2007). needs’ (ISO 2002). Spatial data quality is thus close to
From a GIS perspective, this leads to a multilayer, multi- fitness for use, that is, ‘the ability to satisfy stated and
scale use of spatial information. At the smallest scale (cor- implied needs’. So far, issues of spatial data quality that
responding with the largest area), we have the field with are prominent in modern GIS have barely been addressed
existing map and farmer’s information. A soil map may be within the field of robotics. The aspects of usage like costs
present and topographical maps may show the presence of and accuracy (Aronoff 1991, Devillers et al. 2005) are
tramlines and drainage pipes. As the more detailed scale, the important for assessing the fitness for use of any data and
current position of the robot is investigated and corrected of any classification (Stein et al. 2008). From the domain of
using local information. Clearly, an integration of these two spatial data quality we can easily imagine that the following
levels of information is required to help the robot function in topics are important.
an optimal way. Positional accuracy is the accuracy of address and coor-
dinate values. A distinction can be made into absolute
positional accuracy, that is, the accuracy relative to a given
5. The way ahead coordinate reference system, and relative positional accu-
racy, that is, the accuracy relative to other data in a test data
5.1. Machine learning set. Relative positional accuracy is sufficient to calculate the
Machine learning is a field that has been developed rapidly variance in area and variance in perimeter or diameter by
in the recent years (Bishop 2006). It is the study of computer means of error propagation analysis (Chrisman 1991). More
algorithms that improve automatically through experience. generally, relative positional accuracy is sufficient in an
It has its roots in artificial intelligence. In spatial studies, it error propagation analysis on a single spatial data set. If
has been linked with pattern recognition (Ripley 1996), data sets are to be combined and an error propagation
pattern classification (Duda et al. 2000), and with spatial analysis is needed, however, then absolute positional accu-
sampling (Van Groenigen and Stein 1998). Machine learn- racy needs to be known. In robotics with the various loca-
ing refers to a system capable of the autonomous acquisition lization approaches, and within a field contents of the
and integration of knowledge. This capacity to learn from importance of doing the right thing at the right place, posi-
experience, analytical observation, and other means, results tional accuracy is important. Both the orientation in the
in a system that can improve its own speed or performance, field, that is, the precision of the GPS, and the direction to
that is, its efficiency and/or effectiveness. Machine learning go for a robot have to be addressed.
220 A. Stein

Attribute accuracy is the accuracy of all attributes other There is much to be gained in the domain of advanced
than the positional attributes of a spatial data set. Attributes image analysis. Current research aims at developing and
can be measured either at the numeric or at the nominal implementing high level image analysis routines. In parti-
measurement scale. Nominal attributes are unordered, like cular, super-resolution mapping and texture-based segmen-
for example geological or soil units. The accuracy of nom- tation and classification routines are shown to become
inal attributes can be described by means of the fraction that valuable in the robotics context.
observations correspond to the hypothesized truth. Spatial sampling is another issue that could easily find a
Combined for different nominal units this leads to an error place within robotics. It is and remains important to collect
matrix (Congalton and Green 1999). The accuracy of information from various places. Within an agricultural
numerical attributes can be described with the root mean field setting, we may imagine different scales of variation,
square error (RMSE) and similar other measures. In Van de for example, determined by soil conditions, previous man-
Vlag and Stein (2006), an attempt was made for at least one agement, and natural development. Such information
type of objects (land use forms) to model and quantify should be usable within a robot. In particular, I see a good
uncertain boundaries. We notice in passing that attribute possibility to carry out a model-based sampling procedure
and positional accuracy are difficult to separate in observa- when general information on a specific variable has to be
tional studies. In a robotic context, the attribute accuracy collected, whereas a design-based sampling procedure
comes back when a robot is observing its environment. should be applicable when more general information is to
Such forms of image analysis are subtle and can have a be collected. Adaptive or sequential sampling procedures
large effect. Research at present is going on in this area, are clearly advantageous when the spatial distribution of a
becoming rather specific for various applications. variable (like the presence of a weed) is much more
uncertain.
The examples that I draw on in the above are all more or
6. Vision for the future less related to agricultural field conditions. Also outside this
We notice at present several developments with conse- domain, however, there are good opportunities to integrate
quences for the development of spatial (and temporal) sta- spatial statistics with robotics. A similar example could be
tistics in the nearby future. A solid integration of layers of drawn for application of robotics within a mining environ-
information as available and that are collected on the go by a ment, within a water environment, or within other environ-
field robot is important. Such an integration is not trivial. mental or ecological applications.
Various levels of scale have to be combined, and various In summary, one may be surprised that still so few
issues of data quality are to be matched. Even the use of a interactions exist between these two scientific domains.
relatively simple yield map may already be a complicated Some is hidden behind equations as argued above, where
activity. particular routines are available in both fields of science, but
Mapping is an important element in probabilistic still there is a long way to go. Without any doubt, we will
robotics: on the one hand, chapters 9 and 10 in Thrun see major advances in the near future.
et al. (2005) are entirely devoted to precisely this. Spatial
statistics on the other hand has a long history in kriging and
other spatial interpolation routines. A further integration of Acknowledgments
the geostatistical routines within a robotics environment This article was written during my sabbatical stay at the University
should be beneficial to make the maps as precisely as of California at Santa Barbara. I am grateful to the program
possible, using all the available knowledge in the best Spatial@UCSB, to Prof. Michael Goodchild, and to the ITC direc-
possible way. torate for making this stay possible.
As discussed in Section 5, aspects of spatial data quality
will have a large effect on the quality of a robot under field
conditions. At this stage, it is difficult to (a) collect and References
quantify data quality when a device is on the field, and (b) to Addink, E. and Stein, A., 1999. A comparison of conventional and
use this to the advantage. So far, data analysis concentrates geostatistical methods to replace clouded pixels in NOAA-
on doing the right thing. But it is not difficult to imagine that AVHRR images. International Journal of Remote Sensing, 20,
961–977.
dealing with some uncertain aspects may be beneficial. If, Aronoff, S., 1991. Geographic information systems: a manage-
for example, it is well known that a particular weed species ment perspective. Ottawa: WDL Publications.
is more abundant under rather wet field conditions, then Atkinson, P.M. Pardo-Iguzquiza, E. and Chica-Olmo, M., 2008.
such uncertain information be used by defining a preferen- Downscaling cokriging for super-resolution mapping of con-
tial trajectory for the robot to proceed. It may then start at a tinua in remotely sensed images. IEEE Transactions on
Geoscience and Remote Sensing, 46 (2), 573–580.
low part in the field, albeit this being only vaguely defined, Besag, J.E., 1986. On the statistical analysis of dirty pictures (with
and use such knowledge to detect weed occurrences to discussion). Journal of the Royal Statistical Society, series B,
delete. 48, 259–302.
Annals of GIS 221

Bishop, C.M., 2006. Pattern recognition and machine learning. Karimov, A., 2010. Graph cuts for fast optimization in Markov
New York: Springer. random field based remote sensing image analysis. Thesis
Breiman, L., 2001. Random forests. Machine Learning, 45 (1), 5–32. (M.Sc), Twente University, Enschede, The Netherlands.
Brooks, R.A., 1985. A robust layer control system for a mobile Mercer, W.B. and Hall, A.D, 1911. Experimental errors of
robotics.MIT Artifical Intelligence Laboratory, Internal field trials. Journal of Agricultural Sciences, Cambridge,
Report 864. Cambridge, MA: MIT. 4, 107–132.
Burrough, P.A., 1986. Principles of geographic information systems Mustafa, Y., Van Laake, P., and Stein, A., 2010. Bayesian
for land resource assessment. Monographs on Soil and Network Modeling for Improving forest growth estimates.
Resources Survey No. 12. Oxford: Oxford Science Publications. IEEE Transactions on Geoscience and Remote Sensing, in
Burrough, P.A., 2001. GIS and geostatistics: essential partners for press. DOI: 10.1109/TGRS.2010.2058581.
spatial analysis. Environmental and Ecological Statistics, 8 Polder, G., et al., 2007. Robot system for weed detection using
(4), 361–377. textural image analysis. In: 6th Biennial Conference of the
Chrisman, N.R., 1991. The error component in spatial data. In: D.J. European Federation of IT in Agriculture (EFITA), , Glasgow.
Maguire, M.F. Goodchild and D.W. Rhind, eds., Geographical Rajasekar, U., Stein, A., and Bijker, W., 2006. Image mining for
information systems: principles and applications, Vol. 1, modeling of forest fires from Meteosat images. IEEE
165–174. Harlow: Longman Scientific & Technical. Transactions on Geoscience and Remote Sensing, 45 (1),
Congalton, R.G. and Green, K., 1999. Assessing the accuracy of 246–253.
remotely sensed data: principles and practices. Boca Raton: Ripley, B.D., 1996. Pattern recognition and neural networks.
Lewis publishers. Cambrige: Cambridge University Press.
Cressie, N.A.C., 1991. Statistics for spatial data. New York: Skidmore, A.K., Turner, B.J., Brinkhof, W., and Knowles, E.,
Springer. 1997. Performance of a neural network: mapping forests
Cristianini, N. and Shawe-Taylor, J., 2000. An Introduction to support using GIS and remotely sensed data. Photogrammetric
vector machines. Cambridge: Cambridge University Press. Engineering & Remote Sensing, 63 (5), 501–514.
Curran, P.J., 1988. The semivariogram in remote sensing: An intro- Stein, A., 2008. Modern developments in image mining. Science in
duction. Remote Sensing of Environment, 24 (3), 493–507. China Series E: Technological Sciences, 51 (Suppl. 1), 13–25.
Curran, P.J. and Atkinson, P.M., 1998. Geostatistics and remote Stein, A., Shi, W., and Bijker, W., eds., 2008. Quality aspects in
sensing. Progress in Physical Geography, 22 (1), 61–78. spatial data mining. Boca Raton: CRC Press.
Devillers, R., Bedard, Y., and Jeansoulin, R., 2005. Stein, A., Van der Meer, F., and Gorte, B., eds., 1999. Spatial
Multidimensional management of geospatial data quality infor- statistics for remote sensing. Dordrecht: Kluwer Academic
mation for its dynamic use within GIS. Photogrammetric Publishers.
Engineering & Remote Sensing, 71 (2), 205–215. Thrun, S., Burgard, W., and Fox, D., 2005. Probabilistic robotics.
Dilo, A., De By, R., and Stein, A., 2006. A system of types and Cambrige, MA: MIT Press.
operators for handling vague spatial objects. International Tolpekin, V. A. and Stein, A., 2009. Scale effects and para-
Journal of Geographical Information Science, 21 (4), 397–426. meter optimization for super resolution mapping from
Dubois, D. and Jaulent, M.C., 1987. A general approach to para- satellite images with Markov Random Fields. IEEE
meter evaluation in fuzzy digital pictures. Pattern recognition Transactions in Geoscience and Remote Sensing 47 (9),
letters, 6, 251–259. 3283–3297.
Duda, R.O., Hart, P.E., and Stork, D.G., 2000. Pattern classifica- Van de Kassteele, J., et al., 2006. Statistical mapping of PM10
tion. 2nd ed. New York: Wiley. concentrations over Western Europe using secondary informa-
Gelfand, A.E., Diggle, P.J., Guttorp, P., and Fuentes, M., 2010. tion from dispersion modeling and MODIS satellite observa-
Handbook of spatial statistics. Boca Raton: Chapman & Hall tions. Stochastic Environmental Research and Risk Assessment
CRC Press. (SERRA), 21 (2), 183–194.
Geman, S. and Geman, D., 1984. Stochastic relaxation, Gibbs Van de Vlag, D. and Stein, A., 2006. Uncertainty propagation in
distribution, and the Bayesian restoration of images. IEEE hierarchical classification using fuzzy decision trees. IEEE
Transactions in Pattern Analysis and Machine Intelligence, Transactions on Geoscience and Remote Sensing, 45 (1),
6, 721–741. 237–245.
Glasbey, G.A. and Horgan, G.W., 1995. Image analysis for the Van Groenigen, J.W. and Stein, A., 1998. Constrained Optimization
biological sciences. Chichester: Wiley. of Spatial Sampling using Continuous Simulated Annealing.
Goodchild, M.F., 1999. Future directions in Geographic Journal of Environmental Quality, 27, 1078–1086.
Information Science. Annals of GIS, 5 (1), 1–8. Van Lieshout, M.J. and Baddeley, A.A., 1999. Indices of dependence
Heijting, S., Van der Werf, W., Kruijer, W., and Stein, A., 2007. between types in multivariate point patterns. Scandinavian
Testing the spatial significance of weed patterns in arable land Journal of Statistics, 26, 511–532.
using Mead’s test. Weed Research 47, 396–405. Xiao, S., 2008. Use of Boosting methods to improve object-
ISO (2002). ISO 19113:2002 Geographic information – quality oriented classification for vegetation mapping in Kyabobo
principles, 27 pp. Available at: http://www.iso.org/ (accessed National Park, Ghana. Thesis (MSc). Twente University,
23 November 2010). Enschede, the Netherlands.

You might also like