10:45 AM - 11:00 AM
[PEM09-07] Time-Series Prediction of SDO Ultraviolet Full Disk Images using a Video Prediction Method with Deep Learning
Keywords:Solar, UV, Deep Learning
Space weather disturbances around the Earth have a significant impact on aircraft and satellite operations, and the importance of space weather forecasting is increasing year by year due to the widespread use of space in recent years. Ultraviolet images of global images of the Sun are often used to predict solar activity, which has an impact on space weather. Video Prediction is a deep learning model that takes a portion of a video as input to a model and produces frames that are predicted to follow the input and output. Video prediction has made remarkable progress in recent years with the advent of ConvLSTM (Shi, Xingjian, et al 2015), Pred-RNN (Wang, Yunbo, et al 2017), which models spatial features and temporal changes with a unified memory flow, and the introduction of 3D convolutional E3D-LSTM (Wang, Yunbo, et al. 2018), which has been proposed. In this study, the Motion-Aware Unit (MAU) based on Pred-RNN proposed in Chang, Zheng, et al. (2021) was used to try to predict solar global UV images.
The dataset used was SDO/AIA211Å global images, which clearly show large-scale structures such as coronal holes and active regions. 2010-2022 data were sampled at 4-hour intervals. 48 hours of data, i.e. 12 global images, were used as input. and a model was created to estimate the subsequent 48 hours at 4-hourly intervals.
The model created generally reproduced the large-scale structures that had been identified on the sphere at the time of the previous input, along the differential rotation. The correlation coefficients were calculated by comparing the intensity of the same pixel in the predicted and correct images for 10 such active regions, and were 0.86 at 4 hours, 0.76 at 24 hours and 0.63 at 48 hours. The distribution of brightness intensity was also approximately reproduced by the active regions that existed on the outer edge of the east at the time of the previous input and appeared on the sphere after a time lapse. These results demonstrate the usefulness of deep learning video prediction techniques in space weather forecasting.
The dataset used was SDO/AIA211Å global images, which clearly show large-scale structures such as coronal holes and active regions. 2010-2022 data were sampled at 4-hour intervals. 48 hours of data, i.e. 12 global images, were used as input. and a model was created to estimate the subsequent 48 hours at 4-hourly intervals.
The model created generally reproduced the large-scale structures that had been identified on the sphere at the time of the previous input, along the differential rotation. The correlation coefficients were calculated by comparing the intensity of the same pixel in the predicted and correct images for 10 such active regions, and were 0.86 at 4 hours, 0.76 at 24 hours and 0.63 at 48 hours. The distribution of brightness intensity was also approximately reproduced by the active regions that existed on the outer edge of the east at the time of the previous input and appeared on the sphere after a time lapse. These results demonstrate the usefulness of deep learning video prediction techniques in space weather forecasting.