Automatic segmentation and classification of ESF patterns using a U-NET-based model

Equatorial Spread F (ESF) is a phenomenon that occurs in the magnetic equatorial ionosphere and can disturb radio signal propagation at night. This phenomenon is caused by depleted areas of plasma density (also known as bubbles) that begin at the bottom of the ionospheric F region. The Jicamarca Radio Observatory (JRO) in Peru has made it possible to study ESF using the 50 MHz Jicamarca ionospheric radar in the low-power mode called JULIA (Jicamarca Unattended Long-Term Studies of the Ionosphere and Atmosphere). The radar detects backscattered signals caused by the ESF structures, generating Range-Time-Intensity (RTI) power maps that show the temporal and spatial (altitude) occurrence of ESF. The Madrigal database contains over 20 years of RTI maps measured at Jicamarca, allowing us to identify different ESF morphological patterns or structures, such as Bottom-type, Bottomside, and Radar plumes. These patterns show the evolution of the ESF and can be used to forecast it. However, manually identifying these structures is a time-consuming process. To overcome this issue, a deep learning model using the U-Net convolutional neural network architecture was implemented to segment and classify four morphological patterns automatically. The model was trained using various features such as backscatter power, the F10.7 solar flux index, the disturbance storm time index, the Moon phase, the vertical drift, the zonal drift, and the statistical texture information of backscatter power and vertical drift. The proposed model achieved an accuracy of 90.08% in segmenting and classifying the structures. This model was applied to the RTI database, allowing us to obtain climatology statistics for each morphological pattern.