Real-Time Detection of Strombolian Eruptions through Convolution Neural Network Identification in Infrared Images

Abstract:

A neural network model was built and trained to identify Strombolian eruptions in infrared image sequences that are typically too small to identify in seismic data from the Ray lava lake atop Mount Erebus. The neural network is smaller than the Inception v3 convolutional neural network used previously (Dye and Morra, 2020). The purpose-built sequential convolutional neural network was trained using the Adam optimizer with an input layer that consists of images downsampled to ¼ the original images. The first two layers are 2 2D convolutional layers with 8x8 single strides and a MaxPooling layer between each layer. The layer is flattened using MaxPooling and ReLU activation resulting in a 9600 parameter flattened layer that is fully connected to a 32 parameter dense layer. The top layer uses a sigmoid activation function on 3 nodes that represent the categories of “eruption”, “post-eruption”, and “no eruption”. The resulting network consists of 307,519 parameters. This smaller network achieved a binary cross-entropy validation loss of 0.0747 and a validation accuracy of 98.27% using categorical accuracy with a confidence interval of 95 and a 5% margin of error after 34 epochs of batches of 44 images running for 27 minutes using TensorFlow in CoLab.

Dye, B. C., & Morra, G. (2020). Machine learning as a detection method of Strombolian eruptions in infrared images from Mount Erebus, Antarctica. Physics of the Earth and Planetary Interiors, 305, 106508. doi:10.1016/j.pepi.2020.106508