Project

Back to overview

“KernelCD phase2”: Change Detection in Remote Sensing Images Using Kernel Based Machine Learning Algorithms

English title “KernelCD phase2”: Change Detection in Remote Sensing Images Using Kernel Based Machine Learning Algorithms
Applicant Kanevski Mikhail
Number 144135
Funding scheme Project funding (Div. I-III)
Research institution Inst. de Géomatique & d'Analyse du Risque Fac. des Géosciences & de l'Environnement Université de Lausanne
Institution of higher education University of Lausanne - LA
Main discipline Other disciplines of Environmental Sciences
Start/End 01.10.2012 - 30.09.2013
Approved amount 57'484.00
Show all

All Disciplines (2)

Discipline
Other disciplines of Environmental Sciences
Other disciplines of Earth Sciences

Keywords (8)

Remote sensing; Change detection; Multi-source; Kernel methods; Very high resolution; Ground cover classification; Environmental monitoring; Hyperspectral, Synthetic aperture radar, LiDAR

Lay Summary (English)

Lead
Lay summary

“KernelCD phase 2: Change Detection in Remote Sensing Images Using Kernel Based Machine Learning Algorithms” (SNSF No. 200020-144135)

Multi-modal, multi-source and multi-temporal remote sensing image analysis (M³KCD)

The KernelCD phase II project (M³KCD) aims at dealing with fundamental research in the field of multi-modal, multi-source and multi-temporal remote sensing image analysis. Particularly, robust and flexible change detection and multi-temporal classification systems for airborne and spaceborne remotely sensed imagery are built by exploiting kernel-based learning algorithms.

One of the most challenging problems in Earth observation is the detection of changes occurred between two acquisitions of the same geographical area at different times. Such detections are crucial for environmental and urban monitoring, post-catastrophe assessment and natural hazard mapping. The methods designed during the M³KCD project will be able to deal with noisy environments and will benefit from heterogeneous input sources. They will maintain strict accuracy standards and information extraction capabilities. As for the “KernelCD phase I” project (SNSF No. 20021-126505), these problems will be tackled by considering the most recent developments in the machine learning and pattern recognition communities.

Kernel-methods have been poorly documented in the remote sensing community for multi-temporal and data-fusion applications: we aim at filling this gap by building on the knowledge acquired from the “KernelCD phase I” project, and by proposing high-level research targeted at introducing novel techniques for change detection and image classification under these considerations. Particular attention will be paid to nonlinear feature extraction / manifold learning and dimensionality reduction methods, which will be applied to data-fusion tasks. On these bases, once the data structure is described and learned, a comprehensive change detector (generating change maps or multi-temporal classifications) will be developed on the transformed and fused information. Considering heterogeneous available sources is a compulsory step, as the changes react differently depending on the type of sensor. For the new systems, the lacks of information of a specific sensor shall no longer constitute a limitation and the multi-source nature of the Earth observation science . The issue of detecting changes among acquisitions from different sensors will  be approached via statistical learning methods already studied in the KernelCD phase I project. This continuation allows a natural ending of the aforementioned project: the development of novel advanced techniques dedicated to multi-temporal multi-source data processing will contribute positively to both theoretical and applicative scientific communities.

The “KernelCD” phase 2 starts from the achievements of its precursor, which have been appreciated at international level by publications, invited talks and fruitful collaborations. It is therefore a natural evolution of the project that will increase its impact also in the applicative communities. 

Direct link to Lay Summary Last update: 21.02.2013

Responsible applicant and co-applicants

Employees

Publications

Publication
Classification of urban multi-angular image sequences by aligning their manifolds
Trolliet Maxime, Tuia Devis, Volpi Michele (2013), Classification of urban multi-angular image sequences by aligning their manifolds, in Joint Urban Remote Sensing Event JURSE, Sao Paolo, BRA.
Flooding Extent Cartography with Landsat TM Imagery and Regularized Kernel Fisher' s Discriminant Analysis
Volpi Michele, Petropoulos Gerge P., Kanevski Mikhail (2013), Flooding Extent Cartography with Landsat TM Imagery and Regularized Kernel Fisher' s Discriminant Analysis, in Computers and Geosciences, 57, 24-31.
Multi-view feature extraction for hyperspectral image classification
Volpi Michele, Matasci Giona, Kanevski Mikhail, Tuia Devis (2013), Multi-view feature extraction for hyperspectral image classification, in European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Bruges, B.
Create the relevant spatial filterbank in the hyperspectral jungle,
Tuia Devis, Volpi Michele, Dalla Mura Mauro, Rakotomamonjy Alain, Flamary Rémi, Create the relevant spatial filterbank in the hyperspectral jungle,, in IEEE International Geosciences and Remote Sensing Symposium IGARSS,, Melbourne, AUS.
Multisensor alignment of image manifolds
Tuia Devis, Trolliet Maxime, Volpi Michele, Multisensor alignment of image manifolds, in IEEE International Geosciences and Remote Sensing Symposium IGARSS, Melbourne, AUS.
Multi-sensor change detection based on nonlinear canonical correlations
Volpi Michele, de Morsier Frank, Camps-Valls Gustavo, Kanevski Mikhail, Tuia Devis, Multi-sensor change detection based on nonlinear canonical correlations, in EEE International Geosciences and Remote Sensing Symposium IGARSS, Melbourne, AUS.

Collaboration

Group / person Country
Types of collaboration
Prof. JP Thiran, EPFL Lausanne Switzerland (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Prof. G. Camp-Valls, Image Processing Laboratory, University of Valencia Spain (Europe)
- in-depth/constructive exchanges on approaches, methods or results
- Publication
Dr. F. Holecz, L. Copa, SARMAP company Switzerland (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Dr. A. Pozdnoukhov, National Centre for Geocomputation, Maynooth University Ireland (Europe)
- in-depth/constructive exchanges on approaches, methods or results
- Publication
Dr. Yorgos George Petropoulos, Institute of Geography and Earth Sciences, Aberystwyth University Great Britain and Northern Ireland (Europe)
- in-depth/constructive exchanges on approaches, methods or results
- Publication
Dr. L. Telesca, Institute of Methodologies for Environmental Analysis (IMAA) Italy (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Prof. A. Plaza, University Extremadura Spain (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Prof. F. Golay, Dr. S. Yoost, EPFL Lausanne Switzerland (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Dr. F. Pacifici, Digital globe United States of America (North America)
- in-depth/constructive exchanges on approaches, methods or results
Prof. S. Canu, INSA; Rouen France (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Prof. A. Rakotomamonjy, University of Rouen France (Europe)
- in-depth/constructive exchanges on approaches, methods or results
- Publication
Dr. R. Lasaponara, IMAA Institute /CNR) Italy (Europe)
- in-depth/constructive exchanges on approaches, methods or results

Awards

Title Year
Prix de Faculté, outstanding PhD award 2013

Associated projects

Number Title Start Funding scheme
126505 “KernelCD”: Change Detection in Remote Sensing Images Using Kernel Based Machine Learning Algorithms 01.10.2009 Project funding (Div. I-III)
150593 Multimodal machine learning for remote sensing information fusion 01.12.2014 SNSF Professorships

Abstract

One of the most challenging problems in remote sensing and Earth observation is the detection of changes occurred between two acquisitions of the same geographical area at different times. Such detections are crucial for environmental and urban monitoring, post-catastrophe assessment and natural hazard mapping. Many techniques that aim at addressing the problem starting from different assumptions exist, and among which difference image analysis and post classification comparisons are the most utilized.In the current “KernelCD” project (SNSF Grant No. 200021-126505), the problem of detecting changes in optical images was successfully addressed by applying methods coming from the machine learning community. In particular, supervised and unsupervised kernel-based techniques were developed specifically for change detection. For the supervised methods, the approaches proposed were designed with accent on accuracy, while light computational load was preferred for the unsupervised ones. Finally, robustness was the main criterion for the success of feature extraction techniques, last topic proposed in the current “KernelCD” project. Despite the success of this first part of the project, the potential of such developments for the applicative domain remains underexploited. One of the main limiting factors remains the recourse to single-source sensors. In order to fully accomplish the aim of “KernelCD”, new methods specific for multi-source data need to be developed. To this end, we ask for a second part of “KernelCD” (“KernelCD” phase 2) in this proposal.In this second part of the project, a one year prolongation, we aim at developing approaches specific to multi-source, multi-modal and multi-temporal remote sensing image processing, with focus to accurate and robust change detection. On the one hand, the methods designed will be able to deal with noisy environments and will be robust with respect to heterogeneous input sources. But on the other hand, they will maintain strict accuracy standards and information extraction capabilities. As for the current “KernelCD” project, these problems will be tackled by considering the most recent developments in the machine learning and pattern recognition communities. Kernel-methods have been poorly documented in the remote sensing community for multi-temporal data-fusion applications: we aim at filling this gap by building on the knowledge acquired from the current “KernelCD” project, and by proposing high-level research by introducing novel techniques for change detection under these conditions. Particular attention will be paid to kernel-based nonlinear feature extraction / manifold learning and dimensionality reduction methods, which will be applied to data-fusion tasks (thus adopting a feature-level fusion approach). On these bases, once the data structure is described, a comprehensive change detector (generating change maps) will be developed. Considering of heterogeneous sources available is a compulsory step, as the changes react differently depending on the type of sensor. For the new detector, the lacks of information of a specific sensor will no more constitute a limitation. This continuation allows a natural ending of the “KernelCD” project: the development of novel advanced techniques dedicated to multi-temporal multi-source data processing. The “KernelCD” phase 2 starts from the achievements of its precursor, which have been appreciated at international level by publications, invited talks and fruitful collaborations. It is therefore a natural evolution of the project that will increase its positive impact on to both theoretical and applicative scientific communities.
-