11:00 AM - 11:15 AM
[PEM09-08] Development of an automatic solar filament detection method using deep learning
Keywords:Space Weather, Deep Learning, Object Detection
The influence of solar activity on the Earth's atmosphere is called space weather, which is important in the current utilization of outer space. In particular, the eruption of gas masses called solar filaments can affect the satellites and GPS. On the other hand, filament detection from solar images has been difficult with conventional rule-based methods. Ahmadzadeh et al. (2019) used Mask R-CNN, an object detection model using deep learning, for filament detection and achieved high detection accuracy. However, it has been reported that a single filament is detected as multiple filaments.
Against this background, this study aims to improve the accuracy of filament detection through appropriate learning settings based on the divided problem. Ahmadzadeh et al. (2019) used Mask R-CNN as a detection model, while this study uses Faster R-CNN. Faster R-CNN is a detection model that outputs object coordinates and classes from images. Compared to Mask R-CNN, Faster R-CNN has a simpler architecture because image segmentation is not performed during detection. Considering that Faster R-CNN is less expensive in computation and that the divided problem may be caused by the region proposal process, which is common to both Faster R-CNN and Mask R-CNN, the change of detection model in this study is reasonable.
H-α images taken by the Big Bear Solar Observatory (BBSO) are used as image data, and filament location information reported in the Heliophysics Events Knowledgebase (HEK) is used as annotation data. The 1,068 image data for the period 2012-2016 was used.
First, we performed hyperparameter tuning. The objective is to perform detection tailored to the individual shape of the solar filament by considering its size and shape and tuning the hyperparameters. We used Average Precision (AP) as our evaluation metric, which is a typical metric in object detection tasks. As a result of the tuning, the AP was 0.322. This result indicates that only hyperparameter tuning is not sufficient for proper filament detection. Next, we performed training data selection. In order to emphasize large filaments, which are considered important in space weather, we restricted small filaments from the teaching data. As a result, the AP was 0.441, confirming the improvement in accuracy. This result indicates that the limitation of training data is useful in filament detection. Finally, we performed a dataset modification. The dataset used in this study contains some improperly annotated data, and we expect to improve the accuracy by modifying these errors. As a result, an AP of 0.480 was achieved, confirming a further improvement in accuracy. This experiment partially resolved the divided problem, and suggests that the divided problem is caused by improperly annotation of the dataset. Resolving this problem is expected to further improve accuracy.
Against this background, this study aims to improve the accuracy of filament detection through appropriate learning settings based on the divided problem. Ahmadzadeh et al. (2019) used Mask R-CNN as a detection model, while this study uses Faster R-CNN. Faster R-CNN is a detection model that outputs object coordinates and classes from images. Compared to Mask R-CNN, Faster R-CNN has a simpler architecture because image segmentation is not performed during detection. Considering that Faster R-CNN is less expensive in computation and that the divided problem may be caused by the region proposal process, which is common to both Faster R-CNN and Mask R-CNN, the change of detection model in this study is reasonable.
H-α images taken by the Big Bear Solar Observatory (BBSO) are used as image data, and filament location information reported in the Heliophysics Events Knowledgebase (HEK) is used as annotation data. The 1,068 image data for the period 2012-2016 was used.
First, we performed hyperparameter tuning. The objective is to perform detection tailored to the individual shape of the solar filament by considering its size and shape and tuning the hyperparameters. We used Average Precision (AP) as our evaluation metric, which is a typical metric in object detection tasks. As a result of the tuning, the AP was 0.322. This result indicates that only hyperparameter tuning is not sufficient for proper filament detection. Next, we performed training data selection. In order to emphasize large filaments, which are considered important in space weather, we restricted small filaments from the teaching data. As a result, the AP was 0.441, confirming the improvement in accuracy. This result indicates that the limitation of training data is useful in filament detection. Finally, we performed a dataset modification. The dataset used in this study contains some improperly annotated data, and we expect to improve the accuracy by modifying these errors. As a result, an AP of 0.480 was achieved, confirming a further improvement in accuracy. This experiment partially resolved the divided problem, and suggests that the divided problem is caused by improperly annotation of the dataset. Resolving this problem is expected to further improve accuracy.