Study Of Auroral Oval Segmentation And Substorm Detection In UVI Images

Aurora is a display of natural lights in the sky, particularly in the high latitude regions, caused by the collision of energetic charged particles with atoms in the thermosphere. Images captured by ultraviolet imager (UVI) from space reflect the global distribution of aurora phenomenon, which can benefit the research on the interaction between solar wind and the earth's magnetosphere because of its all-day and full-area observation. Auroral particles centered on the magnetic pole settle to the Arctic and Antarctic regions, which forms the auroral oval with an annular ring shape. When the earth's magnetosphere exhibits strong disturbance, the area of auroral oval would first increase with an expansion in luminosity, and then revert back to normal, forming the auroral substorm. The study of auroral oval segmentation and substorm detection in UVI images contributes to the in-depth analysis of energy storage and release process in solar-terrestrial space and the modeling of auroral particle precipitation, which exhibits great research significance and application value.Existing auroral oval segmentation methods are low in accuracy and sensitive to noise, while existing substorm detection methods are susceptible to subjective errors and time-delay effect. To this end, this dissertation enhances the level set method to achieve accurate, efficient and robust auroral oval segmentation, and improves the sparse and low-rank decomposition method to achieve automatic and precise substorm detection. The main contributions are summarized below:1. A shape-initialized and intensity-adaptive level set method is proposed to improve the accuracy of auroral oval segmentation. To solve the problem of the sensitivity of initial curves to segmentation results, we present a shape-knowledge-based initialization method, i.e., construct a saliency morphological map by introducing the elliptical shape knowledge of auroral oval, and regard its contour as the initial evolving curve. To solve the problem of weak boundary leakage, we propose an intensity-adaptive level set equation, i.e., embed the neighborhood information into the additional speed function and edge stop function, which makes the level set evolution adaptive to local intensity. The proposed method successfully obtains the rugged details of internal boundary, avoids the weak boundary leakage and applies for both full oval and gap oval images, and therefore achieves high segmentation accuracy.2. A generalized lattice Boltzmann level set method is proposed to accelerate the speed of auroral oval segmentation. To solve the problem of restricted numerical scheme, we propose a generalized lattice Boltzmann-based numerical scheme, i.e., construct a lattice Boltzmann equation with an external force term by simulating the process of particle collision and streaming, and leverage it to solve the level set equation. To solve the problem of high computational cost in global strategy computation, we employ the sparse field method which only updates pixels near the evolving curve in each iteration. The proposed method reduces the calculated amount in each iteration, allows large time step to decrease the number of iterations, and therefore achieves fast auroral oval segmentation.3. A Markov embedded level set method is proposed to enhance the robustness of auroral oval segmentation. To solve the problem of the sensitivity of level set evolution to noise interference, we adopt the Markov random filed to build the correlation of a pixel with its neighbors and embed a Markov energy term into the level set energy functional, which encourages the adjacent pixels to fall into the same region, i.e., object or background. The proposed method prevents the evolving curve from converging to local minima by considering the neighbor information, and therefore enhance the robustness against noise. In addition, it achieves high segmentation accuracy in noisy synthetic images, SAR images, medical images and natural images, which demonstrates that its application for more types of images is feasible and promising.4. A shape-constrained sparse and low-rank decomposition method is proposed to achieve automatic substorm detection. To reduce noise interference inherent in current sparse and low-rank decomposition methods, we introduce a shape constraint set with the magnetic information of regions with high substorm appearance probability, which forces the noise to be assigned to the low-rank part (stationary background). The proposed method ensures the accuracy of the sparse part (moving object) and improves the performance of sequence motion analysis. On this basis, we devise an automatic auroral substorm detection system. Firstly, rough detection is conducted to select time points with explosive increase in auroral luminosity as the candidate substorm onsets. Secondly, motion analysis based on the shape-constrained sparse and low-rank decomposition method is implemented for candidate auroral sequences to extract their motion diagrams. Finally, fine detection is performed by checking whether the motion diagrams possess the characteristic that the luminosity and area of auroral oval change from expansion to recovery, and eligible ones are regarded as the real substorm onsets. The proposed system successfully avoids the time-delay effect and the results are highly consistent with real physical situations. Therefore, it improves the accuracy of substorm detection without the decrease in speed.In conclusion, this dissertation is a crossed research between polar space physics and computer vision, which covers the contents of solar-terrestrial space interaction, image segmentation and event detection. In specific, it solves the problems of accurate, efficient and robust auroral oval segmentation and automatic substorm detection, and therefore establish a solid theoretical basis for the in-depth analysis of UVI images. In addition, it creates a new way of applying engineering algorithms to solve natural scientific problems.