Even if the human vision system is the most advanced existing, it has limitations toward speed and the analysis of very complex data sets. The Section for Earth Observation focuses on developing methodology for analysis of huge volumes of remote sensing data and when the data are too complex to be analysed by the human being in an effective way (e.g, a large number of spectral channels).

NRs methodological research and development within the area of remote sensing includes: generic methodology for multi-sensor, multi-scale, multi-temporal and hyperspectral data; retrieval of geophysical variables in the cryosphere; retrieval of biophysical variables in vegetation and forests; detection and classification of amorphous objects; detection of man-made objects in very-high resolution imagery; and high-accuracy pre-processing algorithms. Each of these methodological disciplines are briefly described in the following:

• Generic methodology for multi-sensor, multi-scale, multi-temporal and hyperspectral data
• Retrieval of geophysical variables in the cryosphere
• Detection and classification of amorphous objects
• Detection of man-made objects in very-high resolution imagery
• High-accuracy pre-processing

Generic methodology for multi-sensor, multi-scale, multi-temporal and hyperspectral data: With a growing number of satellite sensors the coverage of the earth in space, time and electromagnetic spectrum is increasing fast. To be able to utilize all this information, there is a need for methods that can handle multi-sensor, multi-temporal and multi- to hyperspectral images. Applications involving environmental monitoring, and mapping and monitoring of resources may benefit from methods that enable a higher exploitation of these data types. New approaches in this field are currently in high demand as such methods can increase the accuracy, efficiency and regularity of mapping and of monitoring of natural resources, the environment and the climate.

Retrieval of geophysical variables in the cryosphere: The most important application areas of snow variables are in hydrology, meteorology and climate monitoring. NR has a long record in retrieval of geophysical variables in the cryosphere and is at the research frontier for snow variable retrieval algorithms. Currently, NR is developing new methods for optical well as multi-sensor time-series algorithms for retrieval of snow variables, such as snow cover area, snow surface temperature, snow wetness, snowmelt onset, snow grain size and snow surface albedo. The increased coverage in space and time made possible by steadily more satellites and more sensors in orbit has significantly increased the capability of remote sensing of the cryosphere. However, there are significant multi-sensor challenges related to the fact that snow optical and microwave sensors observe completely different physical phenomena.

Detection and classification of amorphous objects: Examples of amorphous image objects are clouds, oil spills and remains of cultural heritage sites in agricultural areas. Amorphous objects are characterised by being without a specific shape. However, amorphous objects can usually be categorised into classes based on specific features and in certain cases by their relation to other amorphous objects. The classes might not be very distinct, making classification a challenging problem. NR has developed a three-stage methodology for detection and classification of amorphous objects. The methodology was developed for detection of oil spills, and there are few, if any, other approaches that can compete with the classification accuracy obtained. Recently, we have started to use this approach for detection of cultural heritage sites. Further methodological research is going along two lines: 1) Refinement of the generic approach in order to obtain overall better performance for most applications; and 2) Tailoring to specific applications in order to obtain further improvements only possible by using domain-specific information.

Detection of man-made objects in very-high resolution imagery: Changes in societal infrastructure such as buildings, roads, train tracks, airports, harbours etc. happen at an ever-increasing rate. In order to keep up to date with these changes, frequent map revisions are necessary. Such revisions are however very costly. An alternative to the largely manual map revision procedures common today is based on semi- or fully automatic feature extraction from aerial and satellite imagery. Most of our research in this field concentrates on roads and buildings. The primary research challenges are related to the models applied for the object interpretation and hypotheses generation for obtaining correspondence between data and model. Another application of detection of man-made objects is detection of vehicle locations and movements. Such information represents important indicators of the capabilities of the infrastructure to serve transport needs. The main challenges are related to the low resolution of the images (0.5-1.0 m for very-high resolution satellite images) compared to vehicle sizes, and advanced utilization of context information and scene knowledge is therefore needed.

High-accuracy pre-processing: Time series of satellite observations are crucial for environmental and climate change applications. However, there are still problems with precise co-registration of such data because the geometry varies between the observations. In order to determine temporal changes in variables locally (down to one pixel), better co-registration than available today is needed. Furthermore, geo-referencing is necessary to determine the position in a geographic map system. NR has worked both in the area of image geo-referencing and in image co-registration. However, more research is still needed to make it sufficiently robust when co-registering long time series of data. The research includes improvements of area-based registration techniques by taking into account specific and local properties in the image data and approaches for deformable image registration. With respect to the radiometry of the data, time series of images should be represented by the spectral reflectance (or brightness temperatures, when relevant) of the ground surface. These physical properties should be corrected for effects caused by the atmosphere and variations in the illumination using state-of-the-art or improved methods.