Road extracted from satellite imagery have been used for many different purposes, e.g. military, map publishing, transportation, and car navigations, etc. Many method such as, neural network, Knowledge-based, Optimal search, Snake model, Semantic model, Road operator model, etc. was researched to identify road from satellite image, but because of complicated characteristics of road and image itself, and automated road network extraction still remains a challenge problem, and no existing software is able to perform the task reliably. This paper presents a hybrid method which combines Fuzzy-C-Means with back-propagation neural network and knowledge processing technique to detect roads in SPOT image.
The basic idea of the paper is "easiest first" principal, and firstly focus to extract local salient road segments most easily and reliably, then use contextual knowledge and supervised back-propagation neural network model to extract fuzzy road segments among salient road segment, and then grouping these extracted pixel as seed point, candidate point, and not-road point, and then according to appropriate knowledge rule to traversal and join, guide the further road link in the whole image. At last, some post-processing steps are taken to refine the result. The resultant image shows this hybrid identification method performs better than only using knowledge-based method or neural network techniques.
In this paper, a new classified method for satellite sensing data is proposed by using both a neural network and knowledge reasoning technique. Neural network model can better distinguish land of level I, but if more subdivision needed, it seldom get satisfied result. Knowledge reasoning system can use human geographical knowledge to improve the classification results, but it needs a large amount of assistant knowledge to classify the data correctly. The new method makes use of the advantages of both the neural network and knowledge reasoning technique, and fulfils layered intelligent extraction of linear object and plane-like object for satellite sensing image. It firstly extracts water and road information by neural network and pixel-based knowledge post-processing method, then remove them from original image, and then segments other plane-like object by neural network model, and respectively computes their features, including texture, elevation, slope, shape etc., then extracts them by polygon-based uncertain reasoning method. At last experimental results indicates that the new method outperforms the single neural network method and moreover avoids the complexity of single knowledge reasoning technique.