Distributed Water Body Recognition Research From Landsat Imagery

Accuracy water body information extraction is the crux of water resource survey, flood and drought detection. As the remote sensing technology matures, analyzing water body information based on remote sensing images which are acquired by artificial satellite has been a popular study point. At present, there are three typical methods for water body extraction. The first method is called single-band threshold method. It is easy to implement and sets threshold value on a single band. The second method is multiband spectral relationship method. This method can consider multi-band so that it can dig more features than single-band threshold method. Therefore, multiband spectral relationship method is able to extract linear water bodies. The last method is water index method. By performing logic calculation on multi-band, water index method is capable of abstracting water spectral character accurately. But through a lot researches, all methods which are mentioned above have low degree of automation and regional limitations. For the areas which have different terrain, most of these methods rely on large amounts of prior knowledge and carefully hand-engineered threshold setting. Besides, it also need a feature filtering and artificial selection. So the current methods are hard to meet the demand of many applications. Hence, it is particularly important to put forward a high precision method for water body extraction based on the former research results.In the first place, in order to resolve the problem of data-intensive computing and computing-intensive, the idea of parallel computing is introduced into image processing and information extraction for high real-time demand from application. Multiple hosts are used for building distributed Hadoop cluster so that remote sensing images can be stored and managed by distributed approach. Several kinds of methods are chosen to establish model for water body extraction. And Weigan River was selected as study area for experiment dataset. The experimental results demonstrate that the proposed model can extract water precisely. Meanwhile, the model possesses good scalability and scalability.Aim at the main reasons for hindering the degree of automation, in the second place, we put forward a new model that with no need of manual threshold setting and feature selection. Based on BP(Back Propagation) algorithm, an automatic identification model of water body is designed. Several typical characteristics of the water body are dug from remote sensing image as input unit for training the BP neural network. Select some region as study area for validation experiment. Through several experiments, we can draw a conclusion that the proposed model has better degree of automation and higher identification accuracy.By introducing the machine learning model and distributed computing algorithm, the efficiency, accuracy and the degree of automation about water body information extraction has increases on some degree. But because the difficulty of acquiring labeled data, the data which is used for training model is relatively small. It becomes enormous challenges for the extraction of massive remote sensing data information. The same as brain, Deep Learning has “learning ability” that can abstract highly deeper features from the characteristics of the original data. On account of good characteristics of learning ability, deep learning model is widely used in the field of remote sensing image information extraction. As a consequence, we proposed a water body extraction model based on deep learning. By taking into account the influence of neighboring pixels, a Feature Expansion Algorithm(FEA) is designed for finding more features. Linking the original image features and expanded features, the proposed model is trained, for extracting water body information correctly. The experimental results showed that our model outperformed Support Vector Machine(SVM) and traditional artificial neural network(ANN).