Knowledge-based feature extraction and spectral image enhancement from remotely sensed images.

摘要

Scene development plays the first step for synthetic image generation using DIRSIG (The Digital Imaging and Remote Sensing Image Generation Model). Traditionally the scenes are built manually; the procedure is very time consuming especially for complex urban scenes. The research focuses on contributing to the DIRSIG scene model development based on information retrieval from high-resolution multispectral images, such as WorldView-2 sensor imagery. The proposed approach takes advantage of a sequence of image processing routines to enhance the spectral images and extracts key geographical features for the man-made road network and naturally occurring water bodies. These routines take into account the spatial as well as spectral signatures in the multispectral images. They constitute a chained process, which includes several steps: pan-sharpening, image filtering, classification, segmentation, morphological processing, vectorization and final refinement. In the first step of the process, a novel and highly parallel nearest-neighbor diffusion based pan-sharpening procedure (NNDiffuse) is designed to fuse a high spatial resolution panchromatic image with the spectral image. Image filtering using trilateral filters for multispectral images is devised to process the image, removing small variances in the image as well as preserving significant edges. Spectral features such as Spectral Angle Mapper (SAM) are used to locate natural resource coverage such as water bodies. Multispectral flood fill technique, a graph based connected component technique and a knowledge-based system is used to extract the road networks. Both the road network and the water bodies can be refined and exported as vectorized ArcGIS shapefile. The outcome of the research is a workflow to facilitate scene development from spectral images; it also contributes to the development in the field of cartographic feature extraction, photogrammetry and target detection.