Data Fusion

Data fusion

A framework of tools and methods for combining information from different sources so the result carries more useful information than any single source alone.

Pan-sharpening

Using a single high-spatial-resolution band to boost the spatial detail of a lower-resolution multi-band image.

Panchromatic (PAN) band

A single broad wavelength band captured at fine spatial detail but with no colour/spectral discrimination.

Multispectral (MS) image

An image with several narrower wavelength bands, rich in spectral detail but coarser in spatial detail than a single broadband image.

Pixel-level fusion

Combining raw sensor measurements directly, such as merging a fine-detail band with a coarser multi-band image.

Feature-level fusion

Combining derived attributes — like a vegetation index, a texture measure, or a shape descriptor — rather than the raw measurements themselves.

Decision-level fusion

Combining the outputs of several separate classifiers or models, e.g. through a voting scheme, into one final result.

Data-aggregation fusion

Combining outputs after each source has already been separately classified.

Spatial resolution

The ground area represented by a single measurement — finer means smaller ground area per value.

Spectral resolution

How many, and how narrow, the wavelength bands recorded by a sensor are.

Temporal resolution

How often a sensor revisits and re-images the same location.

Resolution Merge

A generic label (used in one image-processing package) for boosting spatial detail using a finer single-band image.

Modified IHS merge

A method that separates a colour composite into brightness, colour identity and colour purity, swaps the brightness component for a sharper single band, then recombines.

PCA-based merge

A method that decorrelates a multi-band stack into uncorrelated components, substitutes the component holding the most shared structure with a sharper single band, then reverses the transform.

Brovey / Colour Normalization

A ratio-based method (used in one image-processing package) for boosting spatial detail while preserving overall colour balance.

Local-regression sharpening (Zhang, 2002)

A method that fits a small local statistical relationship between each band and a sharper single band within a moving window, rather than one relationship for the whole scene.

Hue-Saturation-Intensity split

A colour-space conversion that separates an RGB image into "what colour," "how pure," and "how bright" components.

Q4 / SAM / ERGAS

Metrics that check how well a spatially-enhanced product still matches the spectral character of its multi-band source.

Overall Accuracy / Kappa / Producer's & User's Accuracy

Statistics computed from a table comparing predicted classes to known reference classes, used to judge how good a classification is.

Feature bloat

The problem of stacking so many overlapping or irrelevant derived layers into a model that it starts fitting noise instead of the real pattern.

GLCM texture

A derived measure of how grey values co-occur spatially, used as an extra input layer describing surface roughness or pattern.

Majority / weighted voting

Rules for reconciling several separate classification outputs into one map by letting the most common (or most trusted) label win.