Research On Water Body Extraction From Remote Sensing Image Based On Machine Learning

Surface water body extraction has great signification for many aspects about flood analysis, environmental protection, water resources exploitation and utilization, etc. With the development of remote sensing technology, the main position is increasingly occupied by remote sensing technology for obtaining water body information. At present, traditional water body extraction methods such as spectral relationship method and normalized difference water index could not satisfy current large-scale application, more and more researchers adopt machine learning methods to do it. Shallow machine learning models have achieved certain effect, but they need complex artificial feature analysis and selection. Deep learning is a new field of machine learning research and becomes current hot research topic of artificial intelligence. Its deep network model has strong expression ability and can self-learn more useful features from samples, making many breakthroughs on speech analysis, image recognition and natural language processing. In this thesis, machine learning methods are adopted to build related models for water body extraction. The main work includes the following aspects:(1) We summarize traditional water body extraction methods and analyze advantages and disadvantages of these methods. Then preprocessing methods of remote sensing images and model of support vector machine(SVM) are introduced in detail, preparing for experiments in later paragraphs.(2) A method is proposed for extracting river water based on BP neural network. It integrates several effective extraction methods of water especially small water and takes full advantage of features combination, extracting features of spectral relationship, modified normalized difference water index, the third component of K-T transformation(TC3) and IHS colorful space from ETM+ images to train the network.(3) A method for water body extraction based on stacked autoencoders is proposed. This method combines advantages of unsupervised and supervised learning. A deep network model is built by stacking sparse autoencoders and softmax classifier. The greedy layerwise approach is adopted to train each layer in turn. Features are learnt without supervision from the pixel level, avoiding the process of explicit feature extraction. The features having learnt and corresponding labels are used to train softmax classifier with supervision. Back propagation algorithm is adopted to optimize the whole model.(4) A method is proposed for water body extraction based on convolutional neural network. This method makes full use of both spectral and spatial information of remote sensing images. It adopts the hierarchical structure similar to the brain, extracting low-level to high-level and concrete to abstract features from raw data directly. Its local receptive field and weight-sharing greatly reduce the number of network parameters which need to be trained, and the complexity of network model. Its sub-sampling structure has high invariance for translation, scaling and other forms of deformation, which has unique advantages in image processing.Experiments conducted on related data show that, contrasted to some existed methods, water body extraction models built in this thesis can effectively extract required water body information from remote sensing images and improve the accuracy and automation level, being more applicable.