Discussion of statistical models and analysis applied to the data.
Method: Statistical and visual identification of the plumes will be done manually to generate the training data on actual plumes. The training data would also include clouds (a significant source of false alarms), and non-activity (true negatives) through statistical methods, human vetting, and simulations. A region based convolutional neural network(R-CNN) would be used as the starting point which would have the Alexnet configuration with an SVM as the final layer for predictions. We would augment the training data (for instance using adversarial networks) in order to avoid overfitting. For detection of temporal activity i.e. tracking of the plumes, we plan to use the Region Convolutional 3D Network (R-C3D) to extract spatio-temporal features capturing activities, accurately localizing the start and end times of each plume.