Automatic Recognition and Extraction of Oil Tanks from High-resolution Remotely Sensed Images

Nowadays, recognition of oil storage tanks that are crucial targets is vital. Making use of remote sensed images becomes more and more important in the field of targets recognition. In recent twenty years, numerous achievements in the field of object recognition, such as roads, buildings, planes and oil storage tanks. There are plentiful kinds of algorithms presented in target recognition and extraction, with the development of computer-assistant recognition and extraction, such as neural networks, wavelet transform, image segmentation, mathematical morphology and genetic algorithm, most of which are based on low-resolution remote sensed images. In recent years, high-resolution remote sensed images have become one of the most important products, because they become not so expensive as before and they have some characteristics such as tremendous data, complex feature details and dependence on scale. Thus algorithms based on high-resolution remote sensed images are necessary and helpful in extraction. However, there lacks general algorithm aim at recognizing oil storage tanks, especially making use of high-resolution remote sensed images. Furthermore, high-resolution images still have some disadvantages, such as too many information to find a particular object. So we bring forth a new technique for information extraction and target recognition from high resolution remote sensing image. This paper brings forward an automatic target recognition algorithm which is based on the knowledge-driven strategy combined with the data-driven strategy. Firstly, we introduce the up-down strategy that collecting large amount data of oil tanks to customize a knowledge database aim at our target, based on the analyze of their size, texture, color, distribution and other features information. Secondly, we utilize data mining method, taking spatial relationship among them and GIS into account as well, to get special priori knowledge at last. Thirdly, we present a new automatic down-up strategy inherited from template matching method and watershed algorithm with adaptive parameters, in order to get a primary segmentation result. Last but not the least, based on the supervisory learning method combining with the primary result, we get precision result of oil tanks extraction and recognition from high resolution remote sensing images. The primary test result shows that this new automatic algorithm has excellent performance in pertinence and speed.