Image processing and analysis comprises all relevant steps to reach from (raw) image data to [...] information via image interpretation and digital image classification. In traditional remote sensing workflows, this step follows the image acquisition process. There are two main components, i.e. (1) image processing, (2) analysis, which emphasizes the sequential nature of the process – while increasingly this dichotomy disappears.
The information production workflow aims at converting semantically rich, but unstructured image data into a set of classes, objects, arrangements, etc., to enable ultimately a complete image understanding and scene reconstruction. This scene reconstruction entails a mental component (“understanding”) and a technical one, by providing standardized classification results or even beyond, dedicated information products in form of digital maps and reports, tailored to the specific application domains and use cases, in order to make informed decisions. Such information products can be maps, reports, dashboards etc., overall it is the transformation from quantitative, semi-continuous digital numbers (“brightness”) to qualitative information using categories and figures, which can be stored and further used in a GIS environment.
The first part of the process entails image calibration, image correction (geometric, radiometric), data assimilation, and any type of enhancement (contrast manipulation, filtering, etc.) which aims to better condition the information extraction part. It ends where we achieve a significant milestone in the processing milestone, remarkably denoted as analysis-ready data (ARD). From there, we enter into the analysis realm, classically referred to as digital image classification, the process of assigning pixels to classes. In other words, the aggregation of pixel values according to their similarity into categorical (nominal) classes. The discrimination of these classes by and large depend on application domain, and ideally, these classes match with information classes. To address the issue of ambiguity and to overcome the so-called semantic gap in image interpretation by providing a stepping-stone in the information extraction process, the strategy of pre-classification (semi-concepts) has been introduced in the literature.
Today, boundaries between pre-processing and classification increasingly vanish, through an increasing level of automation in the pre-processing and image correction steps. In addition, new ways of analysis emerge, in particular in large time series, including image data cubes. Instead of a processing chain, which suggests a linear – and potentially irreversible – cascade of manipulations, the automation of large parts of this part allows us to see the process more reversible and approachable from either side.