The five preceding articles in this series on satellite image segmentation all concerned what can be seen. This one concerns a modality where nothing is seen, in the ordinary sense of the word.
A radar image does not record reflected light but the echo of a signal emitted by the sensor itself. It therefore shows neither the colours nor the shades of a landscape, but its roughness, its geometry and its electrical properties. This article sets out what that changes. It extends the article on segmenting road networks.
Why radar is indispensable
Two properties justify it in satellite image segmentation. They have no equivalent in optical imagery. Independence from weather comes first. The signal passes through cloud, which guarantees acquisition whatever the cover, a decisive advantage in tropical regions and during weather events. Independence from illumination comes next. The sensor emits its own signal, which makes night acquisition as good as daytime acquisition. Those two properties combine to produce this modality’s most useful characteristic: an acquisition regularity optical imagery never reaches. A radar series does not carry the gaps the vegetation article described as the dominant difficulty of optical series. A third property deserves flagging. The signal is sensitive to moisture and structure, which gives access to information optical imagery does not supply at all, soil moisture and canopy structure notably.What a radar image measures
This understanding precedes any discussion of method. It explains the entirety of the difficulties that follow. The sensor emits a pulse and records what returns. The intensity measured depends on three properties of the surface. Roughness at the scale of the wavelength determines signal scattering. A surface smooth relative to that scale reflects away from the sensor and appears dark; a rough surface returns in all directions and appears bright. Geometry determines the direction of return. A surface facing the sensor appears bright, a surface facing away appears dark, which makes relief highly visible. And electrical properties determine absorption. Moisture considerably increases the returned signal, which makes the same soil look different according to its water content. One important consequence follows for annotation. An annotator used to optical imagery spontaneously interprets a radar image as a greyscale photograph, which produces systematic errors, and that reinterpretation requires specific training.Speckle in satellite image segmentation
This characteristic is the primary difficulty of the radar modality. It dominates the appearance of a radar image. A homogeneous surface does not produce a homogeneous value but a grainy aspect, each pixel varying strongly from its neighbours. That phenomenon results from interference between multiple echoes within a single pixel. Three properties clearly distinguish speckle from ordinary noise. It is multiplicative rather than additive, its amplitude growing with signal intensity. It is not random in the strict sense, the same configuration producing the same pattern. And it carries information, its statistics indicating something about the nature of the surface. Two technical responses are commonly used. Filtering, which reduces speckle at the price of resolution loss and smoothed limits. And temporal averaging, which exploits several acquisitions and assumes a surface stable between them. One practical consequence follows for a radar satellite image segmentation project. The degree of filtering must be identical between training corpus and production, a model learned on filtered images and applied to raw ones meeting a texture different from the one it knows.Geometric distortions
Three deformations specific to radar affect relief. They have no equivalent in optical imagery. Foreshortening compresses slopes facing the sensor, an entire hillside potentially being reduced to a few pixels. Layover occurs where a summit returns its echo before its base, which inverts the order of points in the image and makes the area uninterpretable. And radar shadow deprives slopes facing away from the sensor of any signal, producing zones with no information. Those three distortions have a common and particularly decisive consequence. They depend on the incidence angle, which means the same mountainous area appears differently on ascending and descending orbits, and a model trained on one configuration performs poorly on the other. One recommendation follows. A corpus intended for terrain with relief must contain both orbit configurations, an obvious requirement and regularly omitted.Polarimetry
This richness specific to radar is underused and constitutes a differentiating competence. The signal can be emitted and received under different polarisations, and the combination chosen indicates the scattering mechanism. Three scattering mechanisms are distinguishable and identifying them permits a physical rather than statistical interpretation. Surface scattering characterises bare soil and water surfaces. Volume scattering characterises vegetation canopies, where the signal bounces between leaves and branches. And double bounce characterises vertical structures on flat ground, buildings and trunks in flooded areas. That third signature deserves emphasis. It permits detecting flooding under forest cover, a situation the water article flagged as a limit of standard radar and which polarimetry partly removes. One consequence for annotation follows. A polarimetric corpus requires an annotator able to interpret those mechanisms, a competence markedly rarer than reading optical imagery and one justifying a distinct pricing position.Interferometry
This capability sits outside satellite image segmentation while informing it. Comparing the phase between two acquisitions of the same place permits measuring ground displacement on the order of a millimetre, which no other satellite technique approaches. Three distinct operational uses follow: subsidence monitoring, landslide tracking and detection of structural deformation. A related notion directly interests segmentation. Coherence measures signal stability between two acquisitions, and its loss indicates a change of surface. High coherence characterises stable surfaces, buildings and bare soil; low coherence characterises vegetation and water. That property supplies classification information at low cost, complementary to intensity, and it is underused in projects that often limit themselves to signal amplitude.What segments well in radar
Four classes lend themselves particularly to radar satellite image segmentation. It is worth identifying them before committing to a project. Open water presents the strongest contrast, a calm surface appearing very dark, which the water article detailed. Buildings present a strong signature through double bounce, which makes them clearly detectable, with the caveat that their geometry appears distorted. Agricultural surfaces are distinguished by their temporal evolution of roughness, ploughing, growth and harvest producing distinct successive signatures. And large canopy structures, forest against non-forest, are distinguished by volume scattering. One important observation runs across those four classes. Radar excels at structural distinctions and is weak at compositional ones, which places it exactly opposite optical imagery and grounds their complementarity.What segments badly in radar
An honest reading requires setting out the limits of this modality. Fine distinctions between vegetation covers are difficult, two crops of similar structure producing close signatures. Smooth non-aqueous surfaces are confused with water, a difficulty the water article flagged. Areas of rugged relief are partly uninterpretable, through the distortions set out above. And vegetation health largely escapes radar, that information belonging to the spectral response only optical imagery measures. Those four limits lead to a simple scoping recommendation. Establishing whether the information sought is structural or compositional determines the appropriate modality, and that question takes a minute to ask.Complementarity with optical imagery
This combined configuration is the most robust and it remains underused. The two modalities measure different properties of the same surface, which means their errors do not overlap. Three different arrangements are used in practice. Early fusion combines the bands of both sources at the input of a single model, which is simple and assumes strict geometric and temporal coincidence. Late fusion processes each source separately then combines the results, which tolerates date discrepancies better. And sequential use employs one modality to complete the other, radar filling the cloud gaps of an optical series. That third arrangement is the most accessible and the most immediately useful. It addresses the dominant difficulty of optical series the vegetation article set out, without requiring the strict coincidence the first two demand. One important consequence follows for annotation. A multimodal corpus must contain examples annotated in both modalities over the same areas, a resource few projects hold and whose construction is precisely the kind of work a specialised provider can supply.The conventions to fix
Six distinct decisions must be settled before production and they differ from those of an optical project. The processing level of the images supplied to annotators, raw or filtered, and the filtering parameters where applicable. The display scale, linear or logarithmic, which radically changes appearance and therefore the perception of limits. The treatment of shadow and layover zones, which must receive a class of their own rather than an arbitrary assignment. The orbit configuration adopted, and the treatment of areas covered by several configurations. The polarisations used and their composition where a colour representation is produced. And the treatment of sensor artefacts, spurious echoes and very bright point reflectors, frequent in urban and industrial areas.Available radar sources
They condition the economic feasibility of a project. An overview is needed before any scoping of radar satellite image segmentation. Three regimes coexist, as in optical imagery, with a noticeably different distribution. Public programmes supply free and open radar imagery with regular revisit and a multi-year archive, a considerable resource notably less exploited than its optical equivalent. Commercial operators supply higher resolutions and taskable acquisitions, in a rapidly growing market the previous series documented. And airborne campaigns supply configurations satellites do not offer, full polarimetry and different wavelength bands notably. One observation deserves adding on wavelength, a decisive parameter often ignored. Short bands interact with the top of the canopy and long bands penetrate further, reaching the ground under forest for the longest, which makes the choice of band as structuring as the choice of resolution. One practical consequence follows. A project concerning forest structure or flooding under canopy must verify the availability of a long band over its area, a constraint that can settle feasibility before any consideration of method.Radar annotation as a competence
This is the most commercially exploitable point of this article. Radar annotation is markedly less widespread than optical annotation, for three cumulative reasons. Reading it requires specific training, a competent optical annotator not being immediately operational. Public corpora are scarce, which research documented as restricted sensor modality coverage. And the tooling is less mature, part of the annotation toolset having been designed for natural imagery. Those three reasons produce a scarcity of supply while demand grows, driven by the crisis, maritime surveillance and tropical monitoring uses the previous series described. One direct strategic consequence follows. Training annotators in this modality is an investment whose value exceeds that of the same effort applied to optical imagery, precisely because few actors make it.The cost of a radar corpus
It differs markedly from an optical corpus. Five factors compose its structure in satellite image segmentation. Three raise it. The hourly rate is higher, the competence being scarce as the previous section set out. Throughput is lower, reading a radar image requiring more interpretation time than an optical one where objects are immediately recognisable. And verification is heavier, a checker often having to consult an optical image of the same area to settle a doubtful case. Two lower it. Acquisition availability is excellent, the absence of cloud constraint guaranteeing regular coverage of any area of interest. And the useful classes are fewer, radar lending itself to a small number of structural distinctions rather than to detailed nomenclatures. The balance of those five factors depends on the project. A radar corpus generally costs more per unit area than an equivalent optical corpus, and it carries a scarcity value that amply compensates that gap in a market where supply is limited.Evaluating a radar satellite image segmentation product
Three particularities distinguish evaluating a product on this modality. The reference poses a specific difficulty. Photo-interpretation on a radar image is less reliable than optical photo-interpretation, which often leads to establishing the reference on an optical image acquired at a nearby date, with the temporal gap that implies. The acquisition configuration must be documented and evaluation conducted separately by configuration, an aggregate score masking very uneven performance between orbits. And moisture variability affects reproducibility, the same area acquired after rain and in a dry period producing different signatures without the surface having changed. That third particularity is the least anticipated. It requires documenting acquisition conditions and, for an operational product, evaluating across a range of conditions rather than on one favourable acquisition.Approaching a radar project
Five questions scope a project on this modality and they differ from those of an optical project. Is the information sought structural or compositional. That question settles the choice of modality before any other consideration, and it avoids committing to a radar project on a distinction only optical imagery permits. Does the area have rugged relief. Its presence requires handling distortions and covering both orbit configurations, which doubles part of the work. Is temporal regularity a requirement. If so, radar brings an advantage optical imagery does not, and that argument carries particular weight in tropical regions. Is a long band available over the area. That constraint settles the feasibility of a project concerning forest structure or flooding under canopy. And does an optical source exist for establishing the reference. Its availability changes the validation method and its absence appreciably raises the corpus cost. Those five answers determine feasibility, source, reference method and volume of work. Asking them before drafting a proposal avoids the two most expensive errors in this domain, committing to radar on a compositional question and underestimating the effect of relief.What this use case teaches
This observation concerns the modality rather than the class. It closes the first set of this series on satellite image segmentation. The five preceding articles in this cluster examined different classes on one modality. This one examines a different modality, and it reveals an implicit assumption the others shared. That assumption is that the image shows the objects. It is true in optical imagery, where an annotator recognises a building because it looks like a building. It is false in radar, where a building signals itself through a physical mechanism one must know in order to identify it. Three consequences follow and they hold beyond radar. Annotation competence is not transferable between modalities, which is counterintuitive for a client used to treating annotation as a generic task. Training is an investment rather than a prerequisite, and its value grows with the scarcity of the modality. And a provider mastering several modalities occupies a position volume does not procure, an observation the previous series made about scarce corpora in general.Building a radar annotation capability
One practical satellite image segmentation question follows from the scarcity argument above, since the conclusion is only useful if the capability can actually be built. Four distinct elements make up that capability and they can be acquired in sequence rather than at once. Physical understanding comes first. An annotator needs to know why water is dark, why a building is bright and why a slope is compressed, which is a matter of days rather than months and which no amount of practice substitutes for. A worked case library comes second. Documented examples of each mechanism on real imagery, with the reasoning recorded, transfer understanding faster than any explanation. Paired optical reference comes third. Training with an optical image of the same area alongside builds the mapping between radar appearance and ground reality, and that support can be withdrawn progressively. And configuration awareness comes fourth. Knowing which orbit, which polarisation and which processing level is being viewed prevents the systematic errors that arise from treating all radar imagery as one thing. Those four together turn a competent optical annotator into a competent radar annotator, which is the transition the scarcity described above makes valuable. The investment is measured in weeks and the resulting position is one competitors reach slowly, because most do not attempt it at all.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Interpreting a radar image as a greyscale photograph.
- Using optical annotators without specific training.
- Filtering the corpus and the production data differently.
- Mixing ascending and descending orbits without distinguishing them.
- Neglecting shadow and layover zones in relief.
- Expecting compositional distinctions from radar.
- Ignoring the effect of moisture on reproducibility.
- Limiting to amplitude and neglecting polarimetry and coherence.
- Establishing an optical reference without measuring the temporal gap.
- Evaluating an aggregate score across several acquisition configurations.
- Ignoring wavelength, as structuring as resolution.
- Assuming an optical annotator is immediately operational on radar.
