Back to search

ROMFORSK-Program for romforskning

Change detection in heterogeneous remote sensing images

Alternative title: Endringsdeteksjon i heterogene satellittbilder

Awarded: NOK 3.2 mill.

Comparing apples and oranges - and making sense of it Change detection is an important application in earth observation with satellite images. It is traditionally done by comparing images recorded under conditions that should be as similar as possible. The quest posed at the University of Tromsø - The Arctic University of Norway (UiT) has been one less obvious: How to compare different images from different sensors in order to find changes between them? The project «Change detection in heterogeneous remote sensing images» has been undertaken by a PhD student at UiT The Arctic University of Norway in collaboration with research partners at the University of Genoa, Italy. The project has been lead by Associate professor Stian Normann Anfinsen from the Machine Learning Group at UiT. - It is a bit like comparing apples and oranges - in a sensible fashion. The concept of heterogeneous or multimodal change detection is still new, and some of the earlier attempts have been made by our project partners in Italy. We have addressing the challenge with new methods from statistics and machine learning, says Anfinsen. Change detection can for instance be used to detect deforestation or flooded areas. The normal approach is to compare images from the same instrument and sensor mode, such that the measurements at two separate times can be directly compared and where a difference close to zero represents «no change». This requires, however, exact corrections and co-calibration of the images, which is difficult due to common variation of sensor geometry and environmental parameters that change with weather and season. The motivation is both to perform faster and more extensive analyses. With heterogeneous change detection there is no need to wait for an image taken by the same sensor after an event of interest. Since the method uses all kinds of images, changes can be detected as soon as any given image is available. The ability to compare different sensors also provides longer and denser time series of data, that can be used to extend the analysis in time span and temporal resolution. This makes for better exploitation of the vast amounts of satellite images that are captured. The methods that have been developed use the last innovations within deep learning with artificial neural networks, and the work has also contributed new innovative methodological solutions. Demonstrations of the ability to detect changes in multimodal satellite images are included in several scientific paper submitted for publication.

1) Prosjektet har utviklet metoder for endringsdeteksjon i heterogene/multimodale satellittbilder med større nøyaktighet enn noen kjente algoritmer for forskningslitteraturen. 2) Algoritmene er gjort åpent tilgjengelige gjennom kodedatabasen Github og kan tas i bruk av lokal og nasjonal industri 3) Maskinlæringsgruppa har finansiering fra Tromsø Forskningsstiftelse for å fortsette utviklinga av metodene ved å overføre dem fra satellittbilder til medisinske bilder i samarbeid med PET-sentrene i Tromsø, Trondheim og Bergen. 4) Forskningsprosjektet har åpnet en ny forskningsretning innenfor dyp læring med kunstige nevrale nettverk gjennom de metodiske nyvinningene som er gjort i prosjektet.

This project will employ a PhD student who shall develop a suite of algorithms capable of detecting changes in both pairs and longer time series of co-located remote sensing images captured by different sensors or with different acquisition modes. Traditional algorithms for change detection in images typically assume that the images are perfectly co-registered and co-calibrated, such that no change corresponds to zero difference or unit ratio between the compared pixels. This assumption is in most practical applications unrealistic due to variations in viewing geometry and environmental conditions, phenological evolution of vegetation, and other factors that cannot be perfectly compensated for. We proposed an alternative strategy: A pattern of typical temporal evolution is established by use of distance measures that quantify the statistical similarity of sensor measurements. These can be used to contrast single pixels or groups of pixels along both the spatial and the temporal dimension. Clusters of pixels that behave similarly in terms of temporal evolution define a thematic ground cover class with no change. Changes will be detected as divergence and bifurcation within such a cluster, as the spatial configurations or temporal profiles will deviate from the main pattern when changes occur. The distance-based approach will be implemented with kernel methods from recent award winning work on machine learning. The design will rely on a judicious combination of: (i) parametric models that capture prior knowledge based on a physical understanding of the sensors, with (ii) nonparametric models that allow the required flexibility to combine heterogeneous sources of information and to model the sampling distributions of the test statistics. The PhD student will be trained in collaboration with the University of Genoa, Italy, allowing research visits to and co-tutoring by leading experts and authors of seminal work on the topic.

Publications from Cristin

No publications found

No publications found

No publications found

Funding scheme:

ROMFORSK-Program for romforskning