Aggregation of multi-source information for classification of remotely sensed images

Abstract

Classification of remotely sensed images is typically performed by applying a decision rule to each pixel and classify it based on spectral information. To further improve classification accuracy, researchers had proposed the use of multi-source information. In addition to the spectral information, multi-source information also includes other image features such as texture, structural, and etc, as well as any ancillary data that pro-vides useful information for the classification. When using multi-source information for classification purposes, different sources exhibit different degree of source reliability. Fur-ther more, these information sources also, and frequently do, exhibit drastically differing class-dependent reliability for the assessment of different classes. Most conventional clas-sification techniques overlooked such deferring reliability in their classification process. Existing techniques incorporating mechanisms in quantifying the source's reliability for classification purposes only deal with the reliability of the information source as a whole.