Automatic Cloud Detection From High-resolution Remote Sensing Satellite Imagery

Recently,with the flourishing in the techniques and systems of earth observation,the development of the high-resolution remote sensing satellite,such as ZY 3,GF 1,GF 2,Pleiades,etc.,has reached advanced levels in the world.The number of images is increasing exponentially,and the level of economy marketization is continuously improving.The high-resolution remote sensing satellite imagery(we call it as high-resolution imagery in following)and their derivatives have been widely used in many fields,such as change detection,oceanic development,disaster prevention and mitigation,environment protection,the research of agriculture and urbanization,etc.,which has provides an important decision support and information security for some policies,such as geographical sate of the census,“a map”project construction of land resources,the belt and road,etc.However,not all the high-resolution imagery can meet the later production requirement,and the biggest problem is cloud cover.In practice,the process of satellite imagery usually assumes that there is no or few clouds in the image,which however,can't be ignored.Clouds not only cover the lands,but also change the spectrum and the texture information of the image,which leads to lots of problems in the products of mapping,such as matching,radiation correction,uniform color,image interpretation,etc.Cloud detection from satellite imagery is always a popular researching field in remote sensing.Lots of scholars have engaged in this problem for many years,but most of them use hyperspectral images that have abundant bands information.We can obtain an accurate cloud detection result by those useful bands.However,since there are less bands in the high-resolution imagery(usually only red,green,blue and near infrared bands),reliable cloud detection result can't be obtained by those traditional methods.Based on the above issues,we proposed to detect cloud from the high-resolution imagery by focusing on the thorny issues.This paper mainly includes the following:(1)Review the research background of cloud detection at home and aboard,and classify the clouds in the high-resolution imagery based on existing research,which provides advantage for the following work.(2)Research on the features of clouds,such as spectrum,texture,frequency,geometric,etc.and find effective features that used to detect each kind of cloud.(3)Use the knowledge of digital image processing to detect clouds from high-resolution imagery,including color change model,features extraction,morphological processing,etc.In addition,to solve the problems of losing contour information and missing small cloud area,a suitable optimum proposal is proposed in the paper.(4)To solve the problems that exist in the algorithm of cloud extracting based on multi-feature.Some theories of pattern recognition and computer vision to detect clouds from high-resolution imagery,including superpixel segmentation,support vector machine,probabilistic latent semantic analysis model,graph cut method,etc.The experimental results show that the method that based on features can detect cloud automatically in real time,compared to the detection results of other automatic and interactive methods,the overall accuracy of our proposed method achieves nearly 5%improvement,and it is capable of improving the efficiency of cloud detection significantly.When compared to different cloud detection methods in the literature,the overall accuracy of the proposed cloud detection method that based on latent semantic analysis and support vector machine was up to 90 percent for ZY-3 and GF-1 images,which is about a 6.8 percent improvement over the traditional spectral-based methods.The experimental results show that the proposed method can automatically and accurately detect clouds using the multispectral information of the available four bands.