Classification of multispectral satellite imagery

Earth observation is a topic of increasing importance. Sensors, for example mounted on satellites, provide measurement data of e.g. the Earth's vegetation, agriculture, oceans, urban areas and cryosphere. The great challenge is the extraction of key environmental parameters from data, for example with respect to biomass, forest degradation, snow cover etc, which provide us with updated knowledge about the Earth's current state.

The data analysis methods that are used are crucial components, as their capabilities determine the type and quality of the knowledge discovery. In the research field known as machine learning, new types of data analysis tools are developed, currently used e.g. in bioinformatics and web information retrieval, by extracting knowledge in a learning process from representative data examples. These methods are expected to have a potential also within remote sensing.

This project will study methods for automatic classification of land cover classes in multispectral satellite images. Current state-of-the-art methods in Earth observation will be reviewed and analyzed. In addition, recent machine learning classifiers will be examined, and their performance will be contrasted to the established techniques.

Contact person: Assoc. Prof. Camilla Brekke and Assoc. Prof. Robert Jenssen

Geo-referencing and data fusion

Earth observation data is often acquired in custom coordinates systems for each sensor, such as the range-azimuth coordinates of a satellite SAR sensor or the pixel indices of a photograph, and is often different for each scene because of geometric considerations. If data from different scenes or sources needs to be combined (data fusion), including using ground truth data for validation, then it will be essential to have a common reference system. Geo-referencing means to define a location in terms of an Earth centred geographical coordinate system, i.e. latitude and longitude, or a well defined map projection, such as the Universal Transverse Mercator (UTM). Working with such location based data is the scope of geographic information system (GIS) applications.

The student will implement geo-referencing of satellite SAR images, and thereby learn how to calculate the image location from the satellite orbit geometry and projecting to different mapping co-ordinates. The geo-referencing accuracy may be tested with real images by comparing known GPS ground truth points visible in the images, and by measuring the co-location of image structures geo-referenced from different scenes or sensors.

Recommended courses: Fys-3001, Fys-3023

Contact person: Prof. Torbjørn Eltoft and Dr. Anthony Doulgeris

Polarimetric decomposition analysis

Polarimetric decompositions are methods to break down the measurements of a polarimetric synthetic aperture radar (PolSAR) in terms of basic scattering processes that make up the backscattered signal and to quantify their relative contribution. They provide an intuitive interpretation of PolSAR images as simple mechanisms, such as single-bounce surface scattering, double-bounce corner scattering and multiple random scattering from a volume. Some have become universally used for visualisation of PolSAR, such as the Pauli RGB decomposition and the Freeman-Durden decomposition. However, there is an easy tendency to over-interpret the information and care must be taken to fully understand the assumptions and limitations of the decomposition schemes.

The student will review the commonly used decomposition schemes and understand their model assumptions, derivations, numerical solutions and limitations. Their value can then be compared by application to real PolSAR images to determine the discrimination ability of each method. Once the methods of solving these kinds of decompositions are understood, the study can be extended to investigate designing new decomposition methods that may be specialised to individual application problems, such as sea ice or glaciers, as decompositions of known target signatures.

Required courses: Fys-2007, Fys-3011/Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3023

Contact person: Prof. Torbjørn Eltoft and Dr. Anthony Doulgeris

Data fusion of multifrequency/polarimetric SAR data

There are now several polarimetric synthetic aperture radar (PolSAR) instruments used in Earth observation that give a choice of operating frequencies: X-band at 8-12 GHz (approx. 3 cm wavelength); C-band at 4-8 GHz (approx. 5 cm); L-band at 1-2 GHz (15-30 cm); and P-band at 0.3-1 GHz (30-100 cm). The interaction of the microwaves is strongly dependent upon the operating frequency and results in quite different features being observed by the different frequency sensors.

This study will compare the observable differences between several co-located images at different frequencies and try to give a physical interpretation in terms of wavelength dependent scales, penetration depth and scattering properties. The student will then proceed to explore different methods to combine the multi-frequency data and investigate any benefits of using data fusion over the individual frequency results.

Required courses: Fys-2007, Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3011, Fys-3023

Contact person: Prof. Torbjørn Eltoft and Dr. Anthony Doulgeris

Feature extraction from polarimetric SAR data

The basic level of any synthetic aperture radar (SAR) image is a complex valued image, representing a magnitude and a phase. A polarimetric SAR (PolSAR) image has four such complex channels, for each of the combinations of transmit and receive for horizontal and vertical polarisations. Furthermore, radar images are expected to have a mean value of zero and so must be analysed at the second order statistical level to extract useful measured information. For PolSAR images, that means that the basic measurable data consists of 4 by 4 covariance matrices. Obviously, working in this matrix-variate space is difficult to analyse, often computationally slow, and impossible to visualise directly.

This study will investigate feature extraction from PolSAR data to allow simpler, faster analysis methods and to aid visualisation. Features will be chosen from known decomposition schemes, inspired by the mathematical and physical characteristics of the radar measurement system, and other textural and image analysis methods. The objective is to potentially find new features and to explore each feature's characteristics with respect to robustness and statistical properties for the purpose of simple image classification. A large variety of SAR and PolSAR images are available to explore.

Required courses: Fys-2007, Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3011, Fys-3023

Contact person: Prof. Torbjørn Eltoft and Dr. Anthony Doulgeris

Simulation and modelling of speckle statistics

The speckle phenomenon of all coherent imaging systems, including synthetic aperture radar (SAR) systems, originates from the coherent interference of many reflected signals generated by individual scatterers within the image resolution cell. The actual number and position of the scatterers is unknown and varies from cell to cell, and means that statistical methods must be employed to analyse such images. Under certain conditions, the speckle has a Gaussian statistical distribution, but in many situations the signals exhibit non-Gaussian behaviour. Both empirical and theoretically based models for non-Gaussian analysis have progressed greatly in recent years, but the physical interpretation of the non-Gaussian model characteristics is still not very well known.

This study will employ numerical simulation and modelling to investigate the physical basis of speckle's statistical characteristics. The student will design and implement high resolution surface models and perform forward simulation of radar scattering to explore the observed interference (or integrated) signal properties. Surface concepts, such as scale, angle and roughness, will be investigated, as well as the medium's properties, such as the dielectric strength and anisotropy, and the effect of mixtures of different media. The latest methods of matrix log-cumulants shall be used to explore the statistical behaviour of the signals and may be compared to the equivalent results from real SAR images to improve our understanding and interpretation of the real world.

Required courses: Fys-2007, Fys-3011/Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3023

Contact person: Prof. Torbjørn Eltoft, Dr. Anthony Doulgeris and Dr. Stian Normann Anfinsen

Entropy component analysis of hyperspectral images

Kernel entropy component analysis (k-ECA) is a new method for data transformation and dimensionality reduction. Similar to the well-known principal component analysis (PCA) and kernel-PCA, the k-ECA technique is based on an eigenvalue decomposition of the data set, which finds orthogonal (uncorrelated) components in the data. While PCA/k-PCA selects the principal components as those with the largest variance (corresponding to the highest eigenvalues), k-ECA selects the components that maximise an entropy measure. The resulting projection may be very different from that of k-PCA, and has been shown to produce better results for certain applications.

Hyperspectral imaging is an important technique in many fields of science, where quantitative and qualitative information about an object is extracted by imaging it in very many spectral channels (frequency bands). The high number of channels makes dimensionality reduction and feature extraction important tasks, where the problem is to project the data into a low-dimensional subspace, while preserving the maximum amount of information. For this project, images from the Earth observation instrument AVIRIS will be made available. The aim is to apply k-ECA to hyperspectral AVIRIS images, to compare and assess it with respect to alternative methods for relevant image processing tasks.

Required courses: Fys-2007, Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3023

Contact person: Assoc. Prof. Robert Jenssen and Dr. Stian Normann Anfinsen

Change detection in heterogeneous remote sensing images

Detection of changes in multitemporal remote sensing imagery is an important image analysis task with applications in various areas, such as monitoring of deforestation, glacier dynamics, land use, and urban development. Traditional change detection algorithms assume that the images to be compared are spatially co-registered and the measurements are cross-calibrated, such that a direct pixel-by-pixel comparison can be made between the images. More recently, attempts have been made to design algorithms that detect changes based on heterogeneous image sources. The multitemporal data set may contain images from different sensors, or from the same sensor, but produced with in different acquisition modes, such that change is no longer represented merely by a difference in the pixel values.

The goal of this project is to develop methods that can be used to detect changes between radar images acquired at different frequencies, polarisations, and radar geometries (incidence angles), or even between radar images and optical images. The work will be based on advanced methods from multivariate statistics, such as multivariate probability density functions, copulas, and transformations of probability distributions.

Required courses: Fys-2007, Fys-3011/Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3023, Sta-3002

Contact person: Prof. Torbjørn Eltoft and Dr. Stian Normann Anfinsen

Radiometric calibration of SAR data

A synthetic aperture radar measures the electromagnetic wave backscattered from an area illuminated by the radar. The backscatter intensity is used for different image analysis tasks, such as classification, change detection, and estimation of biophysical parameters. However, the measured intensity will depend on the slope of the illuminated area relative to the radar look angle. Variations in local incidence angle due to terrain topography will lead to differences in the effective illuminated area, and thus to intensity variations that could be mistakenly interpreted as variations in the surface medium and its properties, unless they are corrected for.

This project will study different methods that use digital elevation models for radiometric calibration of SAR data. Forest monitoring will be used as an application to demonstrate the impact of the radiometric calibration, since correction for local incidence angle is of great importance to both forest classification, detection of forest disturbance, and biomass estimation in hilly terrain. The aim is to implement an effective radiometric calibration and, as far as time allows, to demonstrate that effective radiometric calibration prevents topographic variations from influencing classification maps, enables meaningful comparisons of images with different radar geometry, and prevents bias in biomass estimation due to the influence of topography.

Required courses: Fys-2007, Fys-3001/Fys-3023

Recommended courses: Fys-2010, Fys-3012

Contact person: Prof. Torbjørn Eltoft and Dr. Stian Normann Anfinsen

Characterisation of SAR measurements of tropical forest

Synthetic aperture radar (SAR) is a suitable instrument for monitoring of tropical forest, since the microwave signal penetrates the cloud cover, which is abundant in these areas. The capabilities of SAR for change detection, forest classification, and biomass estimation in topical forest has been demonstrated, but limitations are also documented. On of the limiting factors is variations in SAR measurements due to changes in environmental conditions, and in particular related to the moisture level of both soil and forest canopy.

The objective of this project is to characterise the SAR measurements of tropical forest within and between geographically separated forest biomes and land cover classes, to investigate and interpret their variability at different polarisations and radar frequencies. The goal is to find invariant features or empirical models that can improve the robustness of change detection and classification algorithms.

Required courses: Fys-2007, Fys-3001/Fys-3023

Recommended courses: Fys-2010, Fys-3012

Contact person: Prof. Torbjørn Eltoft and Dr. Stian Normann Anfinsen

Evaluation of remote sensing for diagnostics of industrial impact

The aim of the project is to review the potential of using remote sensing as a tool to assess the environmental impact of industrial activity. A target area which is potentially disturbed by industrial activities will be chosen, as well as an appropriate undisturbed reference area. Physical and biological parameters that carry information about the environmental impact must then be identified. Thereafter, a survey of available remote sensing methods and data sets with the potential of detecting abnormal changes in the selected parameters will be conducted. The outcome of this process may be to carry out an evaluation of industrial impact on the target area based on available data, or to make recommendations about data collection for a future project.

Required courses: Fys-2007, Fys-3001/Fys-3023

Recommended courses: Fys-2010, Fys-3012

Contact person: Prof. Torbjørn Eltoft and Dr. Stian Normann Anfinsen

Glacier monitoring with SAR data

Remote sensing of glaciers is very relevant for climate monitoring, and satellite SAR is particularly useful with its weather and sun independence. Glaciologist need to clearly identify the main three glacier zones (glacier facies): glacier ice, superimposed ice and firn.

The project would be to identify measurable features for each zone that have a high discrimination power for image classification. The student will gain experience in SAR imaging, speckle and statistical modeling, as well as feature selection, discrimination measures and classification methods from the field of pattern recognition. The study sites are: Holtedahlsfonna and/or Kongsvegen at Svalbard.

Required courses: Fys-2007, Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3023

Contact person: Prof. Torbjørn Eltoft and Dr. Anthony Doulgeris

Sea ice tracking in SAR images

With respect to marine safety and environmental studies, monitoring the large and remote regions of the Arctic is important. During the spring of 2010, large sea ice floes close to Svalbard were tracked with GPS. In addition, a large number of Synthetic Aperture Radar (SAR) images were acquired with the RADARSAT-2, ALOS and the ENVISAT satellites.

The main task in this project is to explore the possibility of tracking sea ice floes in the multi temporal, multi sensor satellite SAR image data sets. This project will be done in close collaboration with Kongsberg Satellite Services (KSAT).

Required courses: Fys-2007, Fys-3011/Fys-3012

Recommended courses: Fys-2010, Fys-3001, Fys-3023

Contact person: Assoc. Prof. Camilla Brekke and Prof. Torbjørn Eltoft

News

Mar 17, 2017
PhD defence
Congratulations to Ane Fors for successfully defending her PhD dissertation titled: "Investigations of summer sea ice with X and C-band multi-polarimetric synthetic aperture radar (SAR)" on 17 March.