Michael Leuenberger, UNIL
Random Forest for Environmental Data Mining
PDF files: abstract – slides
Environmental phenomena don’t lie in a simple two or three dimensional scale. Most of them (e.g. forest fires, landslide, permafrost, etc.) lie in a high dimensional space and require powerful nonlinear methods to explain or visualize their real “shape”. Using anthropogenic and geographical feature data along with the random forest algorithm, this study aims at highlighting factors that most influence the occurrences of such phenomena and at identifying areas under risk.
Developed by L. Breiman and A. Cutler, the Random Forests (RF) algorithm provides an ensemble of classification and regression trees. By a pseudo-random variable selection for each split node, this method grows a variety of decision trees which return different results, and thus by a committee system, returns a value that has a better accuracy than other machine learning methods. This algorithm incorporates directly measurement of the variable’s importance which is used to display the factors affecting the phenomenon under study. Dealing with this parameter, several models can be fitted, and thus, a prediction can be made throughout validity domains.
In the study of forest fires in Canton Ticino, comprehensive RF analysis was carried out in order to 1) understand the importance of anthropogenic and topographic features, 2) to assess the predictability of forest fires using environmental variables, and 3) to compare RF with other machine learning algorithms for this particular case study.
Risk maps were plotted by estimating the burned area taking into account the environmental variables and according to the most relevant selected variables.
Carmen Vega Orozco, UNIL
Spatial Statistics & R: our recent developments
PDF files: abstract – slides
Many environmental, socio-economic and other data can be studied as stochastic point processes where events are represented by points (geographical coordinates) and marks (attributes) within a confined geographical region. Frequently, such events exhibit scaling behavior indicating clustering of their spatial/temporal patterns. In this regard, spatial statistical methods are used to discover patterns in their spatial association (spatial clustering). Thus, clustering analysis can reveal information about the patterns of the underlying process and their relationship with the phenomenon under study.
The present paper presents some clustering measures of point pattern analysis for estimating the degree of clustering of spatial point patterns such as the box-counting fractal dimension, the generalized Rényi dimensions and the Morisita index. This study aims at analyzing the spatial variability of forest fires which is a very complex process as a result of an intermixture of human, topographic, meteorological and vegetation factors. The considered database consisted on 2,402 georeferenced forest fires ignition-points occurring from 1969 to 2008 in Canton Ticino. This canton is located in the Southern Swiss Alps and it is the most fire-prone region in Switzerland.
To compute these measures of clustering in a complex-shape region as is the case of the Swiss Alps, the concept of validity domain is applied to constrain the spatial dimensionality of the phenomenon on the mapping space. The validity domains are regions of interest related to the phenomenon under study and are crucial for predictions. Within the validity domain is possible to generate spatial randomly distributed events which structure properties are well known. These properties can be compared to the real phenomena and the deviation between these measures can quantify the clustering of the original data. These simulations within validity domains allowed as well avoiding performing statistical tests to claim for statistical significance.
Computations were carried out using R free software for statistical computing. It is a free software environment integrating facilities for data manipulation, calculation and graphical display. The spatial point pattern analyses considered in the present study were completely developed by the author.
Devis Tuia, EPFL
Helping remote sensing analysts with active learning
PDF files: abstract – slides
Very high resolution imagery is nowadays accessible to the general public, and can be used to monitor the processes occurring at the surface in an efficient way. However, when resorting to automatic classification or detection models, the bottleneck remains the availability of labeled examples to train the model.
In this talk, I present some ideas to overcome the lack of labeled information by interaction with the machine. Active learning approaches, that rank the unlabeled pixels according to their interest for inclusion in the current model, is discussed in detail. I also discuss the question of learning the skills of the analyst in labeling and the integration of such skills in the learning algorithm.
Anthony Lehmann, UNIGE
enviroGRIDS – retour sur 4 ans de renforcement des capacités d’observation de l’environnement dans le bassin de la Mer Noire
PDF files: abstract – slides
Magda Naparus, EPFL
Digital Elevation Models: regular grid versus Triangulated Irregular Network. A comparative assessment on their suitability for wavelet-based analysis over complex terrains
PDF files: abstract – slides
Although numerous studies are dealing with complex terrain representations, the ability to model natural phenomena linked with an accurate geomorphological description in order to assess landscape’s wind energy potential is still limited. Here we evaluate the response of the multi-resolution digital elevation models and their applicability within wavelet analysis over complex terrains, regarding the setting of the wind farms. The reciprocal interference between wind flows and terrain configuration is scale dependent, and therefore better understood in the context of multi-resolution terrain representation, both as TIN and DEM. Wavelets techniques are available for regular and irregular data sets. For instance, the wavelet coefficients, derived by filtering high resolution DEMs, may provide useful complementary information on the relationship between geomorphological processes and topographical aspects. The Wavelet Triangulated Irregular Network is a multi-resolution representation of TINs, which is more powerful when using a TIN model derived directly from LiDAR data. By comparing both tracks (DEM and TIN), we aim to explore and test their possible advantages and limitations in 3D GIS analyses.
Pierre Lacroix, UNIGE
Cartographic Methods for Visualizing the Explosive Remnants of War
PDF files: abstract – slides
This study aimed to answer the question how cartography can help decision makers visualize the problem of contamination by explosive remnants of war (ERW). We thus explored a set of six cartographic visualization methods and systematically evaluated their usefulness with respect to four categories of stakeholders in the humanitarian demining process (i.e., database administrators, operations officers, directors of national mine action authorities, and donors) at four geographical scales, ranging from municipal to global. The main application of our work is for stakeholders involved in humanitarian demining. We provide them with a comprehensive framework for visualizing ERW hazards at the geographical scale at which they have to make decisions, as well as customized cartographic visualization tools and recommendations to help them make informed decisions. For example, we provide potential donors with a method for obtaining a global overview of ERW contamination while remaining aware of regional variation and hot spots. We also enhance cartographic visualization capabilities using traditional kernel density estimation by customizing key parameters. Specifically, we propose a method for adjusting kernel bandwidth for datasets with highly heterogeneous spatial distributions and a method for generating kernel surfaces from polygon data that consists of infilling the polygons with points before using them as inputs in the kernel density estimation.
Stephanie Rogers, UNIFR
GlaciArch: applying glaciological methods for gauging archaeological potential using GIS
PDF files: abstract – slides
Recent climate changes have led to an increase in the exposure of archaeological remains in frozen environments due to the melting of glaciers, ice patches, and permafrost on a global scale. In the majority of cases, the discovery of glacial archaeological remains has occurred due to chance. In order to avoid the risk of losing exceptional, often organic, cultural remains due to decomposition, systematic and predictive methodologies should be employed to locate areas of high archaeological potential. In this paper, we merge existing glaciological knowledge and methodologies with archaeological and historical information in GIS to gain a better understanding about how people interacted with frozen environments in the recent past, as well as to create a model to determine areas of high archaeological potential for the future based on glacier melting rates. First, glacier outlines from the years 1850, 1973, and 2010, as well as topographic properties such as slope and aspect, will be compared to archaeological and historical databases to validate the relationship between artifact discoveries and glacial extents over time. Next, a simple but robust, broad-scale hypsometric glacier recession model (Paul et al. 2007) will be employed along with several climate change scenarios for the Pennine Alps region. This approach is based on the relationship between current glacier extents and changes in the balanced-budget equilibrium line altitude (ELA0) and will be used to predict where glaciers will recede fastest. The results will specify locations in which archaeological investigations should be conducted first. Ground Penetrating Radar (GPR) data and the spatial distribution of modeled mass balance will then be used to calculate a high resolution glacier evolution model (Huss et al. 2008; Jouvet et al. 2009) at two control sites, the Theodule glacier and the Haut glacier d’Arolla, in order to check the accuracy of the regional model (Linsbauer et al. 2013). The results of archaeological and historical database investigations will increase understanding about the location and retrieval of artifacts in relation to glacier extents over time. The variations in glacier dynamics between sites makes it difficult to predict where archaeological remains might be located; however, we hope database investigation along with broad- and local-scale glacier modeling will provide insight into patterns of artifact location and retrieval to aid in archaeological prospection and investigation in the future.
Olivier Ertz, HEIG-VD
OGO :: cas de géocontamination à COMEM+
PDF files: abstract – slides
La filière d’ingénierie des médias du département COMEM+ à la HEIG-VD incorpore un enseignement orienté géoinformation connu sous le nom de cours OGO. C’est l’interdisciplinarité qui caractérise cette filière, intégrant aux mêmes niveaux d’importances les dimensions technologiques, humaines, organisationnelles et économiques pour la gestion des médias au sens large de contenu informationnel dans un processus de communication.
OGO est donc un nom de code pour un cours bloc donné depuis 10 ans en 3ème année de Bachelor. D’abord optionel, il est devenu progressivement indispensable à mesure que la démocratisation des techniques et usages des SIG se confirme. C’est LE cours à COMEM+ qui met en relief l’information géographique (IG).
Les compétences d’ingénieur des médias peuvent être mises à contribution à toutes les étapes de construction d’un Système d’Information Géographique (SIG), mais plu particulièrement sur les points d’interface avec l’utilisateur final et partout où il est important de valoriser du contenu géoinformationnel.
Ainsi, l’ingénieur des médias « géocontaminé » n’est pas un spécialiste des SIG, mais un intégrateur de composants logiciels pour valoriser l’IG. Par exemple sous la forme de services géographiques, cartographiques, basés sur la localisation et assemblés autour d’une logique de présentation bien pensée au sein d’une application web. Tout cela en considérant autant les aspects technologiques (comme la diversité des supports de publication, statique sur papier, interactive sur mobile etc) que les aspects humains, organisationnels et économiques.
Pour en savoir plus sur le plan de cours et découvrir quelques réalisations : http://ogo.heig-vd.ch
Alain Dubois, HEPIA
De la géodonnée à la représentation 4D: panorama des activités du laboratoire MIP
PDF files: abstract – slides
Le groupe de recherche modélisation informatique du paysage (MIP) est actif aussi bien dans le monde des systèmes d’informations géographiques (SIG) que dans la représentation graphique en trois dimensions (3D).
En créant des passerelles entre le monde du SIG et le monde de la 3D, il est possible de générer rapidement et facilement des portions entières de territoire sous forme de maquette virtuelle. Avec la puissance des logiciels de 3D, on obtient des images de grande qualité graphique, tout en conservant la précision et la rigueur des données géographiques.
Après insertion d’un projet d’aménagement, ces maquettes sont utiles non seulement comme outil d’aide à la décision aussi bien pour les concepteurs d’aménagements que pour les maîtres d’ouvrage, mais ce sont également de précieux outils de communication.
Marc Soutter, EPFL
E-Atlas de l’Oronte
PDF files: abstract – slides
Atlas électronique développé dans le cadre d’un projet de gestion des ressources en eau de la vallée de l’Oronte, rivière qui prend sa source dans la vallée de la Bekaa au Liban et s’écoule vers le Nord, traverse l’est de la Syrie puis la région de Hatay en Turquie pour se jeter dans la méditerranée vers Antioche.
Les ressources en eaux souterraines, bien que relativement abondantes en raison des précipitations sur les reliefs du Mt-Liban et de l’Anti-Liban, sont déjà aujourd’hui surexploitées avec pour conséquence un abaissement marqué des nappes phréatiques dans plusieurs secteurs. Avec une diminution prévue de la pluviométrie et un fort accroissement démographique, les perspectives d’avenir sont peu réjouissantes, à moins que la gestion de la ressource s’oriente vers une meilleure efficacité. Ceci implique une concertation entre les divers acteurs des trois pays concernés.
L’atlas électronique doit fournir un support de discussion et d’aide à la décision dans cette démarche de concertation. La première étape de mise en place de l’Atlas a essentiellement consisté à réunir les informations et données disponibles. Une version préliminaire encore très expérimentale, avec des contenus provisoires essentiellement présents à titre d’illustration, est consultable.
Daniel Rappo
Applications distribuées et géoservices
PDF files: abstract – slides
Le projet IN-TRACK a été financé par le RCSO-TIC (Réseau de Compétences de Suisse Occidentale en Technologies de l’Information et de la communication de la HES-SO).
Réalisé entre 2010 et 2012, et faisant appel à divers spécialistes en informatique, télécommunications et ingénierie des médias, il avait pour objectif de proposer une plateforme unifiée pour le développement d’applications de type LBS (Location Based Services).
Le projet s’est concentré sur la réalisation d’une plateforme de géolocalisation intégrée, sur la conception d’un middleware spécialisé et sur le développement de modules de tracking, notamment indoor.
Divers démonstrateurs et applications métiers ont été réalisés pour participer d’une part à la spécification de l’architecture et de ses composants, et d’autre part pour en valider l’intégration.
C’est dans ce cadre que l’application myFest a vu le jour. Développée à l’occasion du festival Paléo 2010, elle a permis d’acquérir une bonne expertise en matière de conception d’APIs (Application Programming Interface) de type RESTful. Les web services réalisés exposaient une soixantaine de méthodes consommées par trois applications différentes (iOS, Android et HTML).
Quand bien même elles n’étaient pas au cœur des préoccupations du projet, les spécificités liées au développement d’applications mobiles ont été riches en enseignement (disponibilité des librairies, processeurs, bande passante, etc.).
Olivier Ertz et Daniel Rappo, HEIG-VD
Text Geotagger
PDF files: abstract – slides