Automatic Detection and Feature Extraction of Equatorial Plasma Bubbles From All‐Sky Airglow Image Based on Machine Learning

Abstract We developed a deep convolutional neural network‐based program code that automatically detects equatorial plasma bubbles (EPBs), segments EPB morphologies, and then extracts their features (including EPBs' northernmost foot latitudes and zonal drift velocities) from OI 630 nm airglow images. From 2012 to 2022, all‐sky airglow images from Qujing Station, China (geographic: 25°N, 104°E; geomagnetic: 15.1°N, 176.7°E), were manually labeled as EPBs or non‐EPBs. Some images showing typical EPB morphologies also had their contours annotated. This created two unique data sets which are both suitable for supervised learning models. Based on the above data sets, numerous experiments were conducted to train the EPB recognition model (EPB‐RM) and the EPB morphology segmentation model (EPB‐MSM). The results on the test set indicate that ResNet18+CBAM model can perform the best in recognizing EPB (precision: 99%, recall: 91%), and all models perform better during geomagnetically quiet periods and high solar activity periods (F10.7 value >140) than during geomagnetically disturbed periods (Dst index ≤ −30 nT or Kp index ≥3) and low solar activity periods. The EPB‐MSM based on Deeplabv3plus had the highest accuracy (88.2%) for extracting the EPB morphology. Statistical features extracted using the EPB feature extraction program (EPB‐FEP) were highly consistent with those extracted manually. The test results verified that machine learning is an excellent method for automatically detecting and extracting EPB characteristics. Our study provides a convenient tool for analyzing massive EPB images.