ANIMAR (2011-2013)
Automated annotation of images from the surface of Mars
Project details
Funding Institution
Fundação para a Ciência e a Tecnologia (FCT), contract PTDC/CTE-SPA/110909/2009
Period
January 2011 - June 2013
Principal Contractor
Centro de Recursos Naturais e Ambiente, Instituto Superior Técnico (CERENA/IST)
Principal Investigator
Pedro Pina (IST)
Participating Institutions
Instituto de Sistemas e Robótica (ISR/IST)
Team Members
CERENA/IST - José Saraiva, Lourenço Bandeira, Pedro Pina, 3 BIC to contract, 1 BM to contract
ISR/IST – Jorge Salvador Marques, 1 BPD to contract
Abstract
Other than the Earth, Mars is the most intensely studied planet of the Solar System. Automated probes have been sent to the planet since the dawn of the space age. There are many reasons for this interest, but chief among them are the many similarities with the Earth that point to the possible existence of life on the planet. Thus, the effort for the exploration of Mars will continue, and there are currently plans for several future missions. ESA, the European Space Agency, of which Portugal is a full member, has recently signed an agreement with NASA to strengthen cooperation in the exploration of the red planet. The amount and quality of data collected will thus continue to increase fast. This is especially relevant in the case of digital images, whose spatial resolution has increased about 10000 times since the 1960s. Furthermore, orbital probes remain now in operating conditions for a decade or more, acquiring thousands of hundreds of images. The teams that operate the cameras, which are normally responsible for the reception, pre-processing and storage of images, also make a first and generic annotation of the main feature of interest in each image. Though there are images that are later studied in detail and consequently further annotated by other researchers, the great majority of images remain labeled with the initial comment focused on a single feature, even if several interesting structures are present. Due to this, the search for features of particular interest is still painstakingly done manually, demanding too much time and/or too many people. The most recent regional and global studies on a diversified set of Martian structures continue to be based on manual surveys of catalogues containing tens, even hundreds of thousands of images.
The availability of an automated procedure capable of replacing the current manual surveys would certainly be extremely useful and would have a strong impact on the planetary science community, not only because it would be capable of generating thematic catalogues (based both on new images and on the already released image catalogues) with more accurate geomorphic labels, but also because it would provide the exact locations of those structures.
Thus, the main objective of this project is to develop an automatic procedure to delineate regions of occurrence of some of the most ubiquitous structures on the surface of Mars (namely impact craters, dunes, polygonal networks and dust devil tracks) and to provide adequate geomorphic labels to the corresponding images. A platform like this, or similar, does not exist today. It will perform analysis at several spatial scales and on a multi-temporal basis, by monitoring structures that may evolve due to some agent of erosion (wind, ice, heat and cold), and it will be able to indicate the existence of any changes on the surface.
The platform will be constructed using machine learning and pattern recognition techniques which are able to learn from experiments and to robustly adapt to every situation. It will follow two main steps: (i) extraction and selection of the features that are more appropriate to describe each structure; and (ii) application of a statistical learning mechanism which will apprehend the properties of the selected features for each type of terrain and geomorphic structure. We will use simple but effective features that have already proved their potential in a wide variety of image recognition situations (HOG-histogram of oriented gradients, Haar-like features, SIFT-scale invariant features). For the classification, we intend to use some of the most up-to-date efficient techniques available, namely SVM-support vector machines, boosting classifiers and classification trees.
The last two types have the significant advantage of being able to automatically select a small subset within the “best” features to perform the classification, which may significantly reduce the computational efforts and improve the overall performance of the procedure.
The main expected result consists of a platform that will be made available to the scientific community, and that will be able to indicate if a given geomorphic structure is present in an image and what are the spatial limits of its occurrence. This platform will also provide some basic geo-referenced measurements, pinpoint changes in distinct images of the same region and include a module for editing the contours of each object or structure. The partnership is constituted by two groups with complementary experience in the development of automated procedures to process remotely sensed images of planetary surfaces and in the use of the most up-to-date feature extraction and classification methods in other areas of research, namely pattern recognition and surveillance. A close and effective collaboration with two North-American institutions (LPI, USGS), world leaders on the development of automated approaches for planetary geologic cartography, will be also developed.
Funding Institution
Fundação para a Ciência e a Tecnologia (FCT), contract PTDC/CTE-SPA/110909/2009
Period
January 2011 - June 2013
Principal Contractor
Centro de Recursos Naturais e Ambiente, Instituto Superior Técnico (CERENA/IST)
Principal Investigator
Pedro Pina (IST)
Participating Institutions
Instituto de Sistemas e Robótica (ISR/IST)
Team Members
CERENA/IST - José Saraiva, Lourenço Bandeira, Pedro Pina, 3 BIC to contract, 1 BM to contract
ISR/IST – Jorge Salvador Marques, 1 BPD to contract
Abstract
Other than the Earth, Mars is the most intensely studied planet of the Solar System. Automated probes have been sent to the planet since the dawn of the space age. There are many reasons for this interest, but chief among them are the many similarities with the Earth that point to the possible existence of life on the planet. Thus, the effort for the exploration of Mars will continue, and there are currently plans for several future missions. ESA, the European Space Agency, of which Portugal is a full member, has recently signed an agreement with NASA to strengthen cooperation in the exploration of the red planet. The amount and quality of data collected will thus continue to increase fast. This is especially relevant in the case of digital images, whose spatial resolution has increased about 10000 times since the 1960s. Furthermore, orbital probes remain now in operating conditions for a decade or more, acquiring thousands of hundreds of images. The teams that operate the cameras, which are normally responsible for the reception, pre-processing and storage of images, also make a first and generic annotation of the main feature of interest in each image. Though there are images that are later studied in detail and consequently further annotated by other researchers, the great majority of images remain labeled with the initial comment focused on a single feature, even if several interesting structures are present. Due to this, the search for features of particular interest is still painstakingly done manually, demanding too much time and/or too many people. The most recent regional and global studies on a diversified set of Martian structures continue to be based on manual surveys of catalogues containing tens, even hundreds of thousands of images.
The availability of an automated procedure capable of replacing the current manual surveys would certainly be extremely useful and would have a strong impact on the planetary science community, not only because it would be capable of generating thematic catalogues (based both on new images and on the already released image catalogues) with more accurate geomorphic labels, but also because it would provide the exact locations of those structures.
Thus, the main objective of this project is to develop an automatic procedure to delineate regions of occurrence of some of the most ubiquitous structures on the surface of Mars (namely impact craters, dunes, polygonal networks and dust devil tracks) and to provide adequate geomorphic labels to the corresponding images. A platform like this, or similar, does not exist today. It will perform analysis at several spatial scales and on a multi-temporal basis, by monitoring structures that may evolve due to some agent of erosion (wind, ice, heat and cold), and it will be able to indicate the existence of any changes on the surface.
The platform will be constructed using machine learning and pattern recognition techniques which are able to learn from experiments and to robustly adapt to every situation. It will follow two main steps: (i) extraction and selection of the features that are more appropriate to describe each structure; and (ii) application of a statistical learning mechanism which will apprehend the properties of the selected features for each type of terrain and geomorphic structure. We will use simple but effective features that have already proved their potential in a wide variety of image recognition situations (HOG-histogram of oriented gradients, Haar-like features, SIFT-scale invariant features). For the classification, we intend to use some of the most up-to-date efficient techniques available, namely SVM-support vector machines, boosting classifiers and classification trees.
The last two types have the significant advantage of being able to automatically select a small subset within the “best” features to perform the classification, which may significantly reduce the computational efforts and improve the overall performance of the procedure.
The main expected result consists of a platform that will be made available to the scientific community, and that will be able to indicate if a given geomorphic structure is present in an image and what are the spatial limits of its occurrence. This platform will also provide some basic geo-referenced measurements, pinpoint changes in distinct images of the same region and include a module for editing the contours of each object or structure. The partnership is constituted by two groups with complementary experience in the development of automated procedures to process remotely sensed images of planetary surfaces and in the use of the most up-to-date feature extraction and classification methods in other areas of research, namely pattern recognition and surveillance. A close and effective collaboration with two North-American institutions (LPI, USGS), world leaders on the development of automated approaches for planetary geologic cartography, will be also developed.