Research Of Cloud Detection Algorithm For Landsat Multi-spectral Image Based On Support Vector Machine

Landsat satellite images have been widely applied in the fields of resource investigation, agricultural production, environmental monitoring, ecological protection and so on. Due to the influence of the weather, the remote sensing images tend to have some cloud covering areas which seriously affected the interpretation of the images. So detecting the cloud pixels of Landsat image data accurately is important to the further procession and application of the image, such as image classification, recognition and object detection.In this paper, clouds and land features of Landsat satellite multispectral images are extracted. For different cloud detection applications, cloud detection problems of Landsat satellite images are researched by using different SVM classification algorithm. The main research contents and results are divided into the following aspects:1. The purpose and significance of cloud detection algorithm for Landsat satellite image and the development of domestic and foreign were introduced. The classification theory of SVM and TSVM are analyzed briefly.2. A cloud detection method for Landsat satellite multi-spectral image based on least squares twin support vector machines (LS-TSVM) is researched. Firstly, the spectral feature of the cloud pixel was acquired on the basis of the atmospheric radiation characteristics of cloud in different bands and the spectral characteristics of Landsat7ETM+image data. Then the texture feature of the corresponding pixel was obtained by extracting the gray level co-occurrence matrix of the each image block. Using the spectral properties and texture feature of pixels to construct the feature vectors and train LS-TSVM classification algorithm to detect the cloud pixels of Landsat7ETM+image. Experimental results showed the validity of the method.3. A cloud detection method combining ACCA (automatic cloud cover assessment) with WSVM (weighted support vector machine) is proposed. Firstly, the ACCA method is used to divide image pixels into cloud pixels, non-cloud pixels and undetermined pixels. Then using the spectral properties of cloud to construct feature vectors, and determining the weight coefficient by building the weight function on the geometry information of the training sample and using WSVM method to detect the image pixels. Experimental results show that this method has the good detection effect for the thin cloud hardly identified by traditional methods of cloud detection.4. The chapter studies MLTK method and semi-supervised learning theory to apply to remote sensing image cloud detection. A classification model containing unlabeled samples is constructed by using the semi-supervised learning theory. Propose a cloud detection method combining MLTK (modified Luo-Trishchenko-Khlopenkov) with STSVM (semi-supervised twin support vector machine) to detect the cloud pixels of Landsat satellite image. Firstly, the MLTK method is used to divide image pixels into cloud pixels and undetermined pixels. Then using the spectral and texture features of cloud to construct the feature vectors and training STSVM to acquire the hyperplane parameter, which has well detection effect for the thin cloud identified hardly by MLTK. Compared with the TSVM method, experimental results show that this method can effectively improve visual quality and quantitative evaluation.