Automatic Detection And Segmentation Of Cloud And Fog In Landsat Remote Sensing Images

Optical remote sensing images are vulnerable to the influence of cloud,fog and haze(CFH).CFH distribution decreases the availability of remote sensing images in ground target detection and classification,and has a lot of interference to image fusion and quantification of land-surface material.Currently,CFH detection and segmentation methods tend to be "dichotomy",which divide images into two types,"cloud" and "non-cloud".Dichotomy can not meet the demand in many practical RS application,because different CFH distribution on the images has different influence to the ground.This paper attempts to explore a method which detect CFH and automatically divides an image into three parts.Using 20 images of Landsat TM,ETM + and OLI&TIRS from the year of 1984 to 2013 as data source,based on the spectral characteristics of CFH,integrating the advantages of existing cloud detection algorithms,this paper builds the automatic CFH segmentation model,and process model validation.The main contents and conclusions are as follows:(1)Characteristics of CFH and type definitions in remote sensing imagesAccording to the scattering of CFH in different atmospheric conditions,the CFH respond of visible bands have significant differences with that of infrared and near-infrared bands.So CFH distribution area of remote sensing images was defined.Type1 is the region that no pixel values in the remote sensing image bands are influenced by CFH;Type2 is the region that only visible bands pixel values in the remote sensing image are influenced by CFH,and infrared and near-infrared bands are hardly influenced;Type3 is the region that all pixel values from visible to infrared bands in the remote sensing image are influenced by CFH,leading to information loss of all bands.This definitions have good consistency with the result of visual interpretation of(3,2,1)and(7,5,4)color composite image.(2)Analysis and validation of the typical CFH detection algorithmAccording to the results of previous research,we select Landsat images of Taihu Lake for the verification of two typical algorithm-object-based cloud detection(Fmask)and dark-channel-prior-based local dark-objective algorithm(HTM).The results show that,Fmask algorithm for Type3 has high detection accuracy,while Type2 detection accuracy is poor;HTM for Typel has high detection accuracy,but Type2,Type3 are not detected properly.(3)Construction and validation of automatic CFH segmentation modelWe build the automatic CFH segmentation model according to the spectral characteristics,integrating Fmask and HTM,improving the points such as decreasing the influence of highlighting features and using Otsu automatic thresholding algorithm to the segmentation.The validation of 20 images of Landsat of Taihu Lake from the year of 1984 to 2013 show that the segmentation model has good applicability.Visual interpretation of RGB(3,2,1)and(7,5,4)image were used to evaluate accuracy.For the part of water body such as Taihu lake,the correct rates of automatic segmentation model of this paper for Typel,Type2 and Type3 are respectively 94.5%,80%and 99.9%,but there are relati’vely high missing rates for Type2 and Type3 which is due to the use of experienced threshold.As for the non-water region,model’s segmentation correct accuracy for Typel,Type2,Type3 are 94.3%,89.3%and 97.5%.T-test to the segmentation results show that the differences between the three regions are significant with 0.001.The segmentation model and workflow of the paper is simple,high efficiency,and ease for remote sensing applications.