A Methodology For Built-up Areas Detecting In High Resolution Remote Sensing Image

Built-up area is the major place of human activity, it is a kind of very important target in disaster monitoring, urban planning and military reconnaissance.High resolution remote sensing image can provides a wealth of information, but also introduces more details, which brings more challenges into detecting and identifing. To achieve the automated and precise extraction of built-up area, we conduct a series of significant research and exploration, on the basis of previous work,which includes the following aspects.First, review the methods of characteristics analysis commonly used in built-up areas detecting. Combined with previous work, the characteristics of built-up area which is often used to detect are summarized into three categories: the spectral characteristics, the local key points and texture features. For each feature,we introduced their mathematical description and extraction methods, and thorough elaborate the typical case in detail.Then, we propose a method of built-up areas detection using multi-kernel learning, multi-scale Integrating and multi-hypothesis voting. Multi-kernel learning includes integration of multiple features and learning weights of the corresponding kernel function of each feature;Multi-scale Integrating includes dividing the image into many non-overlapping blocks at different scales, respectively using discriminative learning model,and integrating the results of the scales, thus enhance the stability of the target detection further; multi-hypothesis is namely,spliting the image into superpixels and voting with different parameter hypothesis, so that converting the block-level image interpretation result into pixel-leveling results with accurate edges and shapes.Finally, we propose a method detecting built-up area based on pre-classification of mountains and plains. Due to the wide huge differences between different types of built-up area and background,firstly the original data is divided into the class of mountains or plains based on DEM data, and then different detection algorithms are proposed with their different characteristics.Experiments showed the rate and effect are greatly improved.