by Dietrich Hagen

The history of Earth exploration has entered a new phase: After centuries of personal endeavour by sometimes daring, sometimes success-obsessed explorers and almost 200 years of Earth observation from aircraft, the surface of the planet can now be studied from space. Observation techniques have changed with the increasing distance. Initially, the naked eye was used, followed by optical aids such as binoculars and camera systems, which have recently been supplemented by optoelectronic scanners. In addition to visible light, such devices also register wavebands that lie beyond the perception of the human eye.

Vision is a multi-stage process that can be described roughly as follows: a broad band of electromagnetic radiation coming from the sun reaches the earth's surface, is filtered by the properties of the various surfaces, absorbed and partly reflected back, part of which in turn falls on the eye and there, focussed by the lens, hits the light-sensitive cells in the back of the eye. The optic nerve transmits the stimulus to the central nervous system, where the actual seeing takes place, i.e. the image of the world is created in the brain. Analogue to this process, which has not yet been fully explained in detail, the contactless exploration of the earth's surface with satellite-supported recording systems also takes place.

The wide range of electromagnetic radiation, at the short-wave end of which are the very hard gamma rays and at the long-wave end the radio waves, occurs in the range of 0.4 to 0.7 µ (1µ = 1/1000 mm) as visible light with the colours blue, green and red. Below this is the ultraviolet, above this the infrared range. In the infrared in particular, a further distinction is made between near, mid and far infrared for more precise labelling. This distinction is useful because many surfaces of interest to earth exploration reflect waves of this type. These include, in particular, mesophyll-bearing leaf tissue or, more generally speaking, vegetation. However, the other wavebands also react typically in their interaction with earth materials. For example, blue is able to penetrate (clear) water to a certain degree and is then reflected by algae, plankton or the ground. Green allows green vegetation to be recognised, whereas the aforementioned near infrared is used specifically to assess the vitality of plants. The far infrared is, among other things, an expression of the earth's own thermal radiation, which originates from the radioactive decay of the earth's core. The back radiation from deserts, house roofs and paved roads, on the other hand, is similar in a radiometric sense. However, it only resembles the "vegetation-free soil" class if it is dry. Even a small amount of water changes the composition of the reflection and allows moist soils to be reliably distinguished from other surfaces. The following therefore applies: the same wavelengths are reflected differently by different surfaces; on the other hand, the same surface reacts differently to different wavelengths.

The data used here is provided by the TM = Thematic Mapper sensor, which is mounted on board a Landsat series satellite. This satellite orbits the earth in a sun-synchronised orbit near the poles at an average altitude of 705.3 km in 98.9 minutes. This means that it passes over a specific point at the same time of day. Due to a slight shift in each revolution, the starting point is reached again after 16 days. During the overflight, the sensor scans the earth's surface line by line in a 185 x 170 km field with a sensitivity of 256 levels in seven channels. The channels are the sections of the electromagnetic spectrum. The image lines are divided into units (pixels) of 30 x 30 m edge length. Summarised, this results in the following assessment: the spectral resolution of this data is very good, the radiometric resolution is also very good and the area resolution is quite good. Other systems show objects with an edge length of up to 10 m, even up to 1 m, but the radiometric resolution is reduced by half and spectrally there is often only one panchromatic channel available.

This raw product must be corrected for errors caused by earth orbit fluctuations, satellite wobble, sensor differences and atmospheric disturbances, e.g. water vapour and carbon dioxide (CO²) in the atmosphere, using various calculation and filtering operations.

The typical area of a satellite scene (185 x 170 km) comprises 35 to 50 megabytes of data (millions of pieces of information). Such amounts of data can only be processed in very powerful PCs or workstations. However, it is not the individual information that is important, but the totality and distribution of identical or similar data. Analysing techniques are therefore statistical procedures and display operations that reveal orders in the data structure.

The different spectral channels have the following tasks, among others: It was mentioned that visible blue (here: 0.45-0.52 µ, channel 1) penetrates clear water and is only reflected by solid components. Near infrared (0.76-0.90 µ, channel 4), on the other hand, is not reflected by water. This can be utilised in such a way that the channel combination 1 + 4 can be used to identify the boundary between land and water on the one hand and the content of suspended matter on the other. In the visible green (0.52-0.60 µ, channel 2) there is a reflection peak of the chlorophyll-containing vegetation, whereas in the visible red (0.63-0.69 µ, channel 3) there is the maximum chlorophyll absorption. The relationship between red and near infrared is therefore a key to qualitative vegetation mapping. On the other hand, the absence of near infrared - usually together with other features - can be used to infer the spread of settlements. Settlements such as the city of Oldenburg, which are known for their high proportion of greenery, present difficulties: Street trees, public parks and numerous house gardens, together with the buildings and the road network, result in a very small-scale mosaic of reflectance values in most spectral classes. The processing of satellite images can therefore not do without good ground knowledge of at least parts of the terrain to be mapped.

If automatic classification is not an option, known areas are therefore digitised in the subsequent step using predefined keys, i.e. transferred to the computer, e.g. 'sparse residential development', 'city centre', 'bodies of water', 'coniferous forest'. The system uses the training areas within these boundaries to calculate the characteristic composition of the reflectance spectra. In this way, the system "learns" to search for the same or similar spectra in the other areas of the scene and to display them.

The mapping of land use in the narrower sense begins with a task to be described as precisely as possible or the expected result. The research process then proceeds by clarifying the necessary prerequisites until the "simple" facts are available or products can be taken from other sources, for example a vegetation map:

The flow chart can be created according to this logical tree, although the branches in the intermediate steps are not included in the following:

• Determination of the degree of differentiation of the result map according to the number and type of units to be mapped
• Preparation of the data (error correction, contrast stretching, delimitation of the working area)
• Selection and digitisation of the training areas
• Selection of the classification method e.g. automatic, unsupervised methods, semi-automatic, supervised methods etc.
• Multispectral classification
• Initial quality check: e.g. principal component analysis (what percentage of the image content is explained by the most informative channel, what percentage by the second most informative channel, etc.)
• If necessary, correction and restart at step 3.
• Summarisation of highly differentiated spectral patterns from various channel combinations into meaningful mapping units
• Fitting the mapping units into the geodetic framework of the Gauss-Krüger network of the official topographic map (geocoding)
• Second quality check: comparison with real utilisation mapping, aerial photographs, maps of different origin
• If necessary, correction by restarting at step 8. or 3.
• Map production: map frame, legend, separation of colours in preparation for printing
• Products:
  a) Printed land use map b) Information layer for a digital landscape model c) Interpretations, tables, diagrams.

Maps differ from digital satellite scenes for a reason that is as simple as it is momentous: in maps, an image of the earth's spatial substance is produced through the cartographic process of selection, simplification, emphasis, suppression, addition (of signatures, boundaries, names, etc.); a satellite image, on the other hand, shows what this substance looks like from above. A direct view, for example, provides a differentiated picture of the canopy in a forest stand according to composition, age classes, stand density and vitality, where a map at best shows the signature "mixed forest". However, this circumstance cannot be played off in favour of a simple 'better-bad' representation. The content and formal structuring of the maps is offset by the high level of detail and topicality of the satellite scenes.

The ratio of detail to area is important in satellite image analysis. Strictly speaking, a ground element would have to be twice the size of a pixel for it to clearly measure the radiation intensity of only this part of the earth, or even four times the size if an overlap in the direction of flight is taken into account, i.e. 60 x 60 metres. In reality, however, mixed pixels cannot be avoided (example: half barn roof, half pasture close to the farm), which thus change the spectral signature of a ground element. This problem is all the more serious the smaller the utilisation pattern to be recorded (city centres).

The recording of the same spectral signatures at one point in time does not necessarily capture one type of utilisation, because arable land is cultivated with different crops, it can also have progressed differently in its vegetation phase up to and including harvest (bare field), or an undersown crop has emerged and corresponds to the reflectance characteristics of a meadow that is just sprouting. On the other hand, according to the real mapping, similar utilisations take on different values on different substrates, depending on whether the field is located on a moor, in the marsh or on the Geest.

Training areas for the automated spectral pattern calculation must therefore be selected from uses that are far apart but of the same type in order to increase the reliability of the assignment. With this procedure, the hit rate is between 90 and 98 per cent.

Finally, two examples of why a complete assignment of all pixels cannot be achieved: A group of trees on a sports field is recognised by the system as 'deciduous forest', as is an open copse in the field; however, a mapper on the ground would enter 'sports field' in one case and 'shrubbery' in another, possibly even 'fallow land', depending on the predominant land use.

Unused railway facilities such as those in Oldenburg-Krusenbusch can develop into biotopes that spectrally resemble a grassland-scrub formation. However, they should be mapped under their original use, such as 'abandoned industrial/transport areas', possibly with the addition: 'biologically valuable'. This also makes it clear that the pair of terms 'right-wrong' is inappropriate and that the problem must be clarified in terms of content. This requires co-operation between different disciplines.

The analysis of satellite imagery provides an instrument that utilises the wavebands beyond the optical slit to explore the earth. Although some standard algorithms of image statistics are widely recognised, their further development and application to specific space poses problems that can only be solved in conjunction with good subject and regional knowledge. The creation of an automated land use classification, which is currently being developed, is at the same time the basis for a multi-temporal analysis of spatial changes, which are observed and queried from many sides: Expansion of settlements (also qualitative), growth of cities, changes in agricultural land, harvest estimates, monitoring of nature and landscape conservation areas, monitoring of current directions and sand transport on the Wadden coast. Suitable data is collected on an ongoing basis and is in principle available. However, before long-term studies can be carried out, the initial information must be made available.

Dr Dietrich Hagen studied geography, German studies and geology in Berlin, Hamburg and Cologne. He has been teaching physical geography and cartography in Oldenburg since 1971. He has been a visiting scientist at the universities of Torun, Gdansk and Towson, MD (USA). His work currently focuses on computer-aided mapping methods, remote sensing and geographic information systems.