Visualization-informed noise elimination and its application in processing high-spatial-resolution remote sensing imagery

Noise removal is perhaps one of the most fundamental and challenging tasks for extracting useful information from a spatial data set. One of the challenges is that there is no general agreement on the definition of noise that can be universally applied to all different domains. This paper proposes a novel technique called Visualization-Informed Noise Elimination (VINE) to support a customized noise removal through incorporation of domain knowledge. The VINE technique consists of three steps of consecutive operations. First, a k-mutual neighbor graph is derived from a spatial data set to model the spatial proximity among data points. Next, a fast partitioning method is employed to reassemble graph nodes into groups. Last, a 3-dimensional (3D) visualization model is created to provide a layered view of the partitioned data, which allows an informed identification and elimination of noise by tailoring to the requirements of a specific domain. The flexibility and customizability provided by this novel technique ensures an effective differentiation of noise from valid data and demonstrates various advantages over traditional methods with improved results. When adapted in post-classification smoothing of high-spatial-resolution remotely sensed images, this approach was able to discover and reassign noise (such as shadows often seen in high-spatial-resolution images) to its proper target class. By incorporating domain knowledge and making use of spatial contextual information, the VINE technique could produce results significantly superior to existing approaches such as majority filter and size-based noise removal. r 2007 Elsevier Ltd. All rights reserved.