Nowadays, millions of landscape images are uploaded on photosharing platforms such as Flickr or Panoramio. More and more of these images are also accurately geotagged via GPS devices mounted on personal cameras. Each image results from a twofold spatial choice: the choice of the location and the choice of the picture subject. In this study, our focus is on landscape images in large touristic areas. Firstly, our goal is to learn which geographic features play a role in the choice of the location of shared images. For our analysis, we extract a series of geographic features from a Digital Elevation Model (DEM) and a Topographic Landscape Model (TLM) and model the photographic suitability as a density estimation problem in the space of the geographic features. Secondly, each combination of geographic features of a region is associated with a probability to be a location suitable for a photography. The resulting map is useful to promote tourism, to evaluate the landscape attractiveness or with a more technical objective as a prior in close-range photogrammetry. This study shows that databases of geotagged pictures can be used to understand tourists behaviour also in rural areas, even if most of current researches are adressed to cities. The application to a touristic region in the Swiss Alps shows that the proposed method suits well this Geographic OneClass Data problem and is more accurate than both standard KPCA and One-Class SVM to model the suitability for touristic photography at locations unseen during training. As expected, picture locations are mostly correlated with geographic features extracted from the digital elevation model, as well as with those related to accessibility (distance to roads, paths). However, the force of this study is the combination of the geographic features via a kernel method to model more accurately suitable picture locations.
H.2.8 [Database management]: Database Applications Spatial databases and GIS Geographic One-Class data (GOCD), image geotags, landscape images, density mapping 1.
With the advent of "Web 2.0", the number of collaborative databases has dramatically increased. Very often, the databases objects are geotagged, meaning that an attribute related to its geographic location is available.
In this paper, we focus on collective pictures databases.
Pictures uploaded on the Web via a photosharing platform
(Flickr, Panoramio, Instagram...) have a time stamp, some
textual tags describing its content and sometimes a world
coordinate representing the geographic position. We are
specially interested to the geotag which can be attributed
in several ways. Most of the images are located with a click
on a map. The accuracy of the localization is related to the
zoom level used and to the ability to recognize an area in
an aerial view. On the contrary, more and more cameras
have a GPS device able to track very accurately the latitude
and longitude coordinates. As stated in [
Such a study aims at providing a map of the suitability to
be picture location. This map can be useful in several ways.
First, it could be used to promote tourism in areas sharing
similar geographic feature with famous regions. Second, it
can be used to assess landscape attractiveness, a measure
needed in environmental planning [
However, the collaborative database are more widely used
to analyse people behaviour and general trends in the tourists
movements in urban areas [
In order to interpolate probability values in each location of the map and not only in areas where picture are found, we propose to compute density in a space generated from geographic features rather than the space of latitude and longitude. To this goal, we require precise image locations which are provided by the GPS embedded in recent cameras and appropriate geographic features. Such geographic features are extracted with a GI Software either from a Digital Elevation Model (DEM) or from a Topographic Landscape Model (TLM: roads, lakes, forests).
By modeling the density of pictures in the space of
geographic features, we estimate the probability of being a
picture location over all the territory considered. This
setting is equivalent to a One-Class problem (OC) where there
is a set of positive data but no negative data: the locations
of landscape photographies compose the positive set, the set
of \attractive places". However, for the map locations with
no pictures, we don't know if the absence of photographies
is due to the inappropriateness of the location (a \bad" or
negative place) or simply to the lack of pictures in an
\attractive" location. This type of problems is also common in
geographic classi¯cation problems, such as change detection
from satellite images [
During the last years, Kernel methods have been widely used
for OC problems. In this study, we consider the following
OC models: the One-class SVM (OCSVM) [
We will therefore focus on non-parametric methods for
distribution support estimation. Such methods do not
require knowledge about the distribution of the data and ¯t
well with our problem. Besides support vector methods,
another straightforward method is the Kernel Density
Estimation (KDE, also known as Parzen window) [
In this paper, we propose a KDE-based strategy to estimate the probability distribution function (PDF) of the data. We estimate the support of the data both in the original space of geographic features and in the feature space spanned by KPCA. The PDFs of the labeled vs the unlabeled data are used to de¯ne a Bayesian criterion, which measures the probability for a map location to be an image location (given its geographic features). To ¯t the free parameters of KDE and KPCA we de¯ne a performance measure to ensure that most of the locations with photographies are classi¯ed as positives and most of the unlabeled locations are classi¯ed as negative. We compare the proposed approaches with OCSVM and KPCA on a real dataset of touristic pictures taken in the Swiss Alps.
The geographic features are extracted using a GIS software for an ensemble of N cells on a grid with 100m resolution. The considered features are summarized in table 1 and are computed for each cell zj of the grid forming the
N unlabeled set zu = fzj gi=1 and for the n image locations zi, forming the labeled set zl = fzigin=1. Each zi and zj are thus represented by a d-dimensional vector formed by the d geographic features.
The ¯rst set of features is extracted from the DEM (Altitude, slope, curvature and visible sky). Then, for the TLM based features, the distances from the cell to the nearest forest, lake, road and cable car are computed. All the geographic features are mean centered and scaled unit variance. 2.2
The KDE function in equation (1) is used to evaluate the density of a data set from the observations of the positive class zl. Once the density has been estimated, we can evaluate the density for an unknown location zj.
f^(zj; zl) = 1 nh i=1 n X K ³ zj ¡ zi ´ h where K(x) is a local smoothing operator, or kernel function. Among the di®erent kernel functions, we used the Gaussian function: 1 2
K(x) = (2¼)¡d=2exp(¡ 2 x )
where h is the bandwidth of the kernel function. Scott's
rule is used to compute the bandwidth [
We consider the following scheme to describe our GOCD problem. Let Y 2 0; 1 be the event (or class) \is a picture location": Y = 1 for a cell being on a picture location and Y = 0 otherwise. The probability of a cell being a \picture location", given its geographic features zj , is
P (Y = 1jzj) = p(zjjY = 1)P (Y = 1) p(zj) P (Y = 1jzj) is the value we want to compute for each cell of the map. The data density p(zj) is estimated with a KDE on a random set of unlabeled cells: f^(zj; zu), the conditional probability p(zjjY = 1) is estimated from a KDE on the labeled data only: f^(zj; zl) and P (Y = 1) is a unknown constant c. We are observing:
P (Y = 1jzj) c = p(zjjY = 1) p(zj)
A threshold can be set on the cell probability from equation 5 under which it is assumed to be drawn from the generic distribution p(zj), while above it is assumed drawn from the distribution of the labeled data p(zj; Y = 1). In practice, we set the threshold to one.
To chose the best set of parameters for the method, a
performance measure need to be de¯ned. To this end, our
labelled set is divided in three subsets, the ¯rst one to train
the KDE and the second to select the parameters. We want
the estimated densities of the training and testing sets to be
(6)
(7)
similar, nevertheless both should di®er from the random set
density. Indeed, if the density of the random set and the one
issued from the labeled set are similar, the ratio for equation
5 will tend to one. This means that the geographic features
where badly chosen and are not able to distinguish properly
the random data from the labelled ones. In [
Where r is the recall (the proportion of correctly predicted data in the testing set) and rpos is the ratio of positive predicted locations in the random set.
In this study, we propose another criterion, more adapted for the density estimation process: we want to ensure that the bandwidth h ¯ts well for both the training and testing sets and thereby their density should be similar. We want to maximize C in equation 7:
C = 2 sR
2 = s2T ¡ st 2 sR (sT ¡ st)(sT + st)
Where sR is 1 ¡ rpos; sT = (1 ¡ rT ) with rT being the recall for the training set and st = (1 ¡ rt), with rt being the recall for the testing set. Thus, C is very similar to Fpu but the component (sT ¡ st) ensures that the PDF estimation of the training and testing set are similar. 2.3
In order to take into account the possible correlation
between the geographic features and the non-linearity in the
data distribution, we propose to estimate the density of the
data either in the original space or in a feature space spanned
by the mapping Á(:) induced by the KPCA. Ho®mann [
The kernel functions project the data on a hypersphere
which has the same dimension number than the number of
training data. OCSVM searches for a hyperplan with two
constraints. First, the intersection between the plan and the
sphere enclose a ratio of the points equal to 1 ¡ ¹, where ¹
is the ratio of the outliers. Second, the plan has to be as
far as possible from the origin. SVDD [
As in [
First, we will present the data and geographic features considered and how the image locations di®er from the distribution of the random locations. Then, we conducted some experiment using di®erent combination and di®erent amount of geographic features. Finally, we will present the resulting probability maps. 3.1
The experiments consider one of the political regions of Switzerland located in the Swiss Alps: \Valais - Wallis". The area is bounded by the Geneva Lake to the West and by the highest summits of Switzerland. It encloses some of the most touristic spots in Switzerland: Zermatt, the Matterhorn and the Aletsch glacier. The altitude gradient is important between the lowest area on the Geneva lake shore (450m) and the highest summit the \Dufourspitze" (4634m). The area is a valley, whose bottom hosts small to mid-sized villages. Climbing the °anks, the low altitudes are generally covered with vineyards (500-700m); then forest, mountain villages and resorts are found in the range from 700 to 1400m; above 1400m pastures and slopes dedicated to ski give access to the highest peaks, playground for the alpinists.
The Swiss Topographic Agency (Swisstopo) provides a
DEM; among the available resolutions of 25m and 200m,
we retained the ¯rst. Swisstopo also provides a TLM
containing vector layers of several territorial objects. For this
study, we selected the roads, other transportation facilities,
forest and lake layers. The image locations are extracted
from the Flickr database and were provided by [
In order to measure the signi¯cance of these features, the unlabeled data and the picture location distributions are compared. For each of the geographic feature chosen, their distribution diverge (tested by a Kolmogorov-Smirnov test with ® = 1%). Despite their statistical di®erences, some features are more discriminative than others. The following list presents the geographic features selected and explains how their value are di®erent at true image locations. ² Altitude (Z): Since we are focusing on landscape photography, few images are taken between 450 and 1500m. In contrary, the range between 1800 to 2200m, where the ski slopes are found is very attractive. There is a small mode above 4000m representing pictures taken by alpinists at high altitude, ¯gure 2 (a). ² Curvature (Curv): Positive curvatures (convex area) are more represented; indeed the ridges are more attractive than the valley. ² Slope: The °atter areas from 5% to 30%) are preferred. ² Visible Sky (Sky): This feature con¯rms the result observed with the curvature: cells with a high ratio of visible sky (>90%) are more often chosen. ² Distance to nearest lake (DLake): People tend to take more pictures in a radius of 200m around the lakes. ² Distance to nearest road (DRoad) and presence of roads (BRoad): The cells close to roads and paths are more active. Approximately 70% of the pictures are taken in a cell containing a road or a path, ¯gure 2 (b). ² Distance to the nearest forest (DFor): The random and image distributions are very similar. ² Distance to the nearest cable car (DLift): Half of the pictures are taken within a range of 1500m surrounding a lift. ! (a) (b)
In table 2, we report 10 experiments, obtained by considering di®erent combination of geographic feature for GOCD. In this table only the best experiments are presented (C >= 20). Thus, the less-signi¯cant geographic features are less represented (DLift, DFor). The best combination of geographic features is altitude, slope, visible sky, curvature, distance to the nearest road and the binary roads (Exp. 9 in Table 2). The superiority of this experiment is observed for the two methods proposed but the KDE(KPCA) has the best result on the independent validation set, corresponding to a recall of 0.85. The KDE in the original space obtains slightly lower performance with a recall at 0.78. Thereby, it appears that the KDE and KDE(KPCA) have a similar behavior for the combination of few signi¯cant geographic feature (Exp. 1-4, 6, 7). However, if more features (Exp. 8 10) or binary features (Exp. 5) are added, the KDE(KPCA) is more suitable to describe the data relations between geographic features, resulting to better results. Both KDE and KDE(KPCA) are more suited than KPCA and OCSVM for this problem. By inserting the unlabeled data in the classi¯cation process, we ensure that less unlabeled data are classi¯ed positive and thus less data in the independent validation set are misclassi¯ed. 3.4
On the map in ¯gure 3 (a), probabilities for the Zermatt area are presented. The ¯lled dots are the training locations, the triangles are the testing locations while the empty circles are the validation locations. Image locations are superposed to the results of KDE(KPCA) and correspond to the estimate of p(Y=1jz) at each location. A misclassi¯cation corresponds to a circle that would be located on an area of low probability (blue). The validation set is very speci¯c in this area: the pictures on the south are above 4000m and accessible via a cable car (arrow A on the map in ¯gure 3 (a)). This area is also a skiing region during winter and summer. First, some of the misclassi¯ed data are found along the cable car and are easily explicable. Indeed, these pictures aren't taken from the ground (but from the lift) and thus these locations aren't related to the geographic features. This also explains why the \DLift" geographic feature is not in the set of the best geographic features. Another set of locations, on the south of the \Breithorn" are badly classi¯ed (arrow B on the map in ¯gure 3 (a)). They are shot on the way to this peak which is one of the most easily accessible 4000 summit in the Alps. However, the path to the summit is not in the roads / paths database. From this map, we can understand the geographic features related to the image locations. First, the mountain paths are easily recognizable in the ¯gure. Indeed, they are always more attractive than the other locations. However, the paths on steep slopes are less probable than the other ones. Then, the ridges (Gornergrat), passes and summits (Matterhorn) are more attractive than other areas.
At the scale of the whole area, as seen in ¯gure 4, it is interesting to note that the method is able to extract di®erent behaviours. For instance, the paths are expected to have a large probability. By combining the geographic features, it appears that it is true at medium altitude. However, in the valley, where the roads have more tra±c, locations between the roads are preferred. Moreover, at higher altitude, where ski slopes are found, people are more disposed to move away from the path. It is intuitive that the mountain peaks are good places to shot pictures. The strength of our method is to rank the peaks according to their altitude and shape. Finally, the less attractive locations are the steep slopes, bottom of deep valleys and °at areas at high altitude (glacier). Indeed, these regions are hardly accessible.
The GPS measured positions of shared images are more accurate than locations provided by the user. In this paper, we propose to use the GPS coordinates of a set of landscape pictures to train a classi¯er of likely and unelikely image locations. Every map location is described with geographic features extracted from a DEM and a TLM. The method proposed projects the geographic features in a space 1 2 3 4 5 6 7 8 9 10
considered
DRoad, Curv
DRoad, DLake
DRoad, BRoad
DRoad, DFor
DRoad, DLift
C 33.3 19.2 20.43 27.23 29.6 18.8 20.7 21.85 R(V) 0.83 0.73 0.83 0.78 0.78 0.69 0.75 0.79
Npc 4 4 7 5 7 4 5 7 7 8 C 34.5 24.86 35.6 26.95 46.27 25.7 23.54 34.44 R(V) 0.83 0.72 0.86 0.78 0.84 0.55 0.7 0.86 Npc 3 5 6 6 4 4 2 6 5 2 Fpu 1.94 2.43 1.59 2.10 3.12 1.52 2.11 1.52
0.57 0.01 0.45 0.55 0.42 0.53
Fpu 1.66 1.19 1.23 1.3 1.89 1.18 1.21 1.16 º 0.4 0.45 0.35 0.35 0.45 0.25 0.2 0.25
R(V) 0.41 0.48 0.52 0.42 0.37 0.64 0.59 0.63 of higher dimension using a KPCA. Then, the spatial probability density function of the picture locations and random locations are estimated with a density function (KDE). The relation between them is used to compute the probability of each map cell to be a likely location given the geographic features. The recall on the independent validation set surpasses the KPCA and OCVM classi¯cation in their regular implementation.
This preliminary study could be re¯ned in several ways. First, the geographic features computed were selected from a priori expected correlations. However, other geographic features could also be correlated to the image locations (orientation, DEM based features at ¯ner or coarser scales, rivers, cable car departure and arrival stations only etc.). Second, our studies focus on a small area in the Alps, a similar approach could be applied to the entire Alps. Indeed, a bigger amount of picture locations would re¯ne our results and increase their robustness and generalization power at a larger scale. Third, in this work we avoided the KDE bandwidth setting using the Scott's rule. A more appropriate bandwidth could improve the results. Finally, the improvement between the proposed method (KDE and KDE(KPCA)) and the standard ones (OCSVM, KPCA) is mainly due to the insertion of the unlabeled data in the classi¯cation process. Our method is easy to implement and shows good results without ¯ne tuning.
We proved that the locations of a map are not equiprobable relative to landscape image locations. By describing each map location with geographic features, one can extract the most probable regions. The generated map di®ers from the density maps based on the Northing and Easting of the picture locations in two ways: ¯rstly, using the geographic features, probabilities are also computed for region without image locations. Second, by taking into account the locations accessibility, the computed probabilities are more realistic. The force of our method is to learn attractive combinations of geographic features with the density estimation. Indeed, the relation between a geographic feature and an the picture location are intuitive (eg. paths are more probable, convex area are preferred...), but their combinations is more powerful to classify correctly picture locations.
Currently, there is a huge interest in computer vision for the pose estimation of shared images at a local or worldwide scale. Our results show that it is possible to use geographic features as a prior knowledge to either discredit some unlikely poses or to promote the more probable ones.
This work has been partly supported by the Swiss National Science Foundation (grant PZ00P2-136827).