The process of image classification extracts information about semantic labels of pixels or objects (i.e. regions) from imagery. Apart of input imagery, the process requires an input set of target classes (classification scheme) for which their spectral (and other) properties have to be identified. A classification method has to be selected that transforms the image data and the classification scheme into semantic map information. In complement to the resulting sematic labelling products, a secondary outcome are instructions or rulesets with the used parameters that constitute the documentation of the classification process.
The input imagery consists of one or more images (optical and/or SAR data) of a specific geographic area, collected in multiple bands of the electromagnetic spectrum (that may have already undergone certain pre-processing steps; determined by the purpose). Additionally, the imagery may include derived spectral indices, principal components, filtered bands, or other features to support the classification process.
The classification purpose defines the information about the target classes. It includes classification schemes (taxonomies), spectral signatures for each class and, mental concepts and categories about the classes (that enable an analyst to distinguish classes by texture, spatial relationships etc.). Often, training areas are used to understand how an object of a particular class is discernible in the available imagery and separable from other classes. Both the input imagery and the chosen classification method determine which features of each class can be exploited for classification. For example, spectral signatures of the target classes (extracted from training areas with known class label) may be a suitable input for extracting information with a pixel-based classification. For shape features, objects are a pre-requirement, derived with segmentation. They are only available with object-based classification approaches.
Classification methods: Various methods exist that can be categorized according to the classification logic that they follow when transforming the input information into the output semantic labelling products. These can be parametric or nonparametric, supervised or unsupervised, per-pixel or object-oriented, semi-automated or fully automatic, and hybrid approaches. Classification methods are for example bayesian techniques like conditional probability or maximum likelihood, clustering (unsupervised), decision trees, deep learning and machine learning.