Automatic Forecasting of Volcanoes Eruption Time

Abstract

Detecting volcanic eruptions before they occur is a significant issue that has traditionally proved to be challenging. However, it is an interesting concern because of its great potential in saving millions of lives by accurately forecasting the eruption time of volcanoes in the proximity of the inhabited areas, which would provide additional time and facilitate the evacuation processes of people early enough before the catastrophe. In order to forecast the eruption of volcanoes, 10 sensors were mounted around 4431 volcanoes to monitor the eruption time. The goal is to estimate the time of the eruption of another 4520 volcanoes in the coming years. Now that we have established the problem, our approach is formulated using various regression models of machine learning to see the best solution for prediction. These are: Random Forest (RF), Artificial Neural Network (ANN), Convolution Neural Network (CNN), and Long ShortTerm Memory (LSTM). We used the Mean Absolute Error (MAE) as our loss function as well as performance metric, since the time to erupt values are large, the loss will be as well. We have labeled data for training and validation, and unlabeled data for testing. The best unlabeled loss is 6,042,665 belongs to CNN model. Finally, we analyzed the results and compared between the four models and found that RF doesn’t work well with noisy data. On the other hand LSTM, ANN, and CNN have nearly the same behaviour but the latter gave better results.