Abstract
Earth remote sensing methods, used to detect and map oil spills, are considered. The research is aimed at solving problems in the field of aerospace monitoring of oil development sites and oil spills. The object of the research is images or a sequence of images of the natural environment. The subject of the research is presented by mathematical modeling and software for image processing and analysis, recognition and classification of images, to assess the environmental condition of offshore oilfields. The aim of the research is to detect oil pollution in the shelf area by means of semi-automated image analysis using super-accurate neural network algorithms (ResNetIO) with long- and short-term memory (LSTM) when processing materials from several information sources, which require spatial matching between the images. To recognize objects on the ground, visual data have been analyzed based on the formation of stable features using deep neural networks. Multidimensional heterogeneous (multiple images) remote sensing data has been proposed as the basis. The research tasks, which help to realize the aim are: to form a set of image sets, containing scenes in a changing environment (weather conditions, illumination, seasonality, viewing angle); to use own base, allowing to allocate stable image features; to develop methods and algorithms for processing and analysis of images for formation of stable features forming scenes from neural networks; to conduct computer experiments for comparative analysis and evaluation results of classification and categorization with the use of neural networks. For this purpose the author has made automatic registration of geometrical deformations (displacement, rotation, scale change), using bilinear interpolation and testing for possible variation of statistical model inside the inhomogeneous sliding window based on the semi-automatic approach for the shelf area of the Oil Rocks (Caspian Sea). Standard single-layer 2D LSTM networks [18] are used to solve the texture segmentation problem by classifying texture perpixels. The network accurately estimates texture regions and automatically adapts the different scale, orientation and shape of texture regions in the image. A simple way of using LSTM networks for texture segmentation is shown, and the efficiency (accuracy) is compared, using research-based measures of classification quality using a new similarity measure based on a statistical model (three versions of the nearest-neighbor rule and the maximum likelihood method) [6]. The results of the studies have generally confirmed the effectiveness of the proposed model. The second front of the research (object recognition analysis) has shown that context information must be taken into account when applying object recognition systems. It has become possible to find evidence that the natural formation of clusters indicates that contexts emerged, which appeared to be fundamental to performance results. However, it is important to emphasize that these experiments are empirical in nature and are conducted on a particular image base that is well known in the academic scientific community
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call