Research On Cloud Detection Technology Of Remote Sensing Image

Modern satellite remote sensing images have been extensively applied in people's production and life.However,as long as there is a cloud barrier between the satellite and the ground,the available information of ground objects contained in satellite remote sensing images will be greatly reduced,or even cannot be used effectively.At present,elimination processing of satellite downlink images is mainly performed by the ground receiving station,which consumes time and effort,and wastes the precious storage space of the satellite.Hence,if cloud detection of images may be performed in the satellite to automatically delete the images with high cloud coverage and only transmit effective images,it will considerably save the satellite storage space for images,improve the efficiency of image processing on the ground,which is of great significance to the on-orbit applications of remote sensing satellites.The main research work and achievements covered in this paper include:(1)The development and research status of cloud detection technology for remote sensing images is investigated and researched.The applicable scenarios,advantages and disadvantages of commonly used cloud detection methods for remote sensing images both domestic and abroad are compared and analyzed.The classification criteria for cloud and underlying surface are determined through extraction and analysis of cloud features.In addition,a relevant satellite remote sensing image database is built by selecting 600 Landsat satellite remote sensing images containing all kinds of clouds and underlying surface information from Geospatial Data Cloud website and preprocessing the remote sensing images through the manual marking of the cloud and underlying surface area.(2)The cloud detection method based on spectral characteristics is researched and simple threshold method,multi-threshold method,and adaptive threshold method based on maximum interclass variance(OTSU)are designed and implemented.In addition,a cloud detection method based on statistical characteristics is innovatively proposed and in-depth analysis is performed to the effect and advantages and disadvantages of each method.The research result shows that the cloud detection method based on statistical characteristics offers the best recognition effect.The accuracy of identification reached90.87%.(3)The cloud detection method based on texture characteristics is studied.By calculating the gray-level co-occurrence matrix of images,the five classification indicators of images: energy,entropy,inverse moment,autocorrelation and average gray value are extracted,and the Support Vector Machine(SVM)is used for classification.In view of the problem that this method has few discrimination indicators and poor classification effect,a new discrimination method based on the texture and structural characters of images and “3D gray level co-occurrence matrix” is proposed,which greatly increases the recognition accuracy of remote sensing images.With respect to the discrimination of image cloud sectors,this paper presents a cloud sector discrimination method based on overlapped blocks which can effectively improve the precision and accuracy of the discrimination.(4)To address the problem that high-lighted special ground features such as snowfield,snow mountains,tundra,and desert and so on are easily misjudged as clouds,this paper comes up with a cloud detection technology based on Preference Support Vector Machine(PSVM),which generates preference support vector machine model using preference training,and carries out cloud detection together with conventional support vector machine model.This method increases the accuracy of the algorithm in the recognition of cloud and common landforms to a certain extent,it also improves the accuracy of it in the recognition of clouds and special landforms by more than three times.The experimental results show that the accuracies of the PSVM method in the recognition of both average recognition and special landforms are 97.66% and 99.31%respectively.(5)A human-computer interface is set up.You can select the type of algorithm and the image to be processed in the interface.The algorithm automatically identifies the cloud sectors in remote sensing images and marks them with distinct colors.Furthermore,it automatically calculates the cloud coverage of images,and those with high cloud coverage will be automatically eliminated by means of manually setting the threshold value from the interface.