使用雙向卷積長短期記憶網路的全持續時間閃焰檢測、分類和區域提取

Abstract

太陽閃焰對於地球有著深遠影響，因此一直是研究人員關注的焦點之一。目前已有許多關於太陽閃焰的預測以及偵測的相關研究，但是在檢測閃焰的整個持續時間仍然有待探索。在這項研究中，我們提出了一種用於在極紫外線範圍內的圖像上使用雙向 LSTM 在整個持續時間內偵測、分類和提取太陽閃焰區域的自動系統。與當前的許多研究不同，我們在短時間內使用圖像來訓練我們的網路。此外，我們也提出了一種自定義的資料集生成方法，該方法能夠在閃焰期間創建全太陽圖像序列，特別是閃焰區域。為了利用耀斑事件的時間和空間訊息，我們使用多個卷積 LSTM，從而得到一個相對輕量級的模型。我們的模型在僅使用圖像資料的持續時間的條件下，可以成功檢測閃焰的近似值，這是在該領域針對未探索問題的一種新穎方法在該領域。

Solar flares have been one of the focal interests among researchers, as they have a profound effect on Earth. Both prediction and detection of solar flares are well studied topics in the field, however detecting the full duration of a flare has yet to be explored. In this research, we propose an automatic system for detection, classification, and extraction of solar flare regions for their entire duration using bidirectional LSTMs on images in the Extreme Ultra Violet range. Unlike numerous current research, we use images within a short time window to train our network. Moreover, a custom dataset generation method has also been proposed, which is able to create sequences of images of the full sun during a flare, as well as flaring regions specifically. In order to exploit both the temporal and spatial information of the flare event, we use multiple convolutional LSTMs, resulting in a relatively lightweight model. Our model can successfully detect flares for their approximate duration using only image data, which is a novel approach at an unexplored problem in the field.