Binary Classification Edge Cases in Cyclones Using AlexNet and ResNet Neural Networks

Two convolutional neural network binary classifiers were developed to classify images with and without hurricanes. These binary classifiers were modelled after ResNet, and AlexNet architectures. The cyclone data was gathered with the assistance of the National Aeronautics and Space Administration’s Interagency Implementation and Advanced Concepts Team (NASA IMPACT). The IMPACT Team archived Tag Image File Format (TIFF) images from the NASA Worldview application. This data store containing over 250 TIFF files of cyclone data, with examples of true and false positives of this cyclone phenomena, was then augmented through shear transforms, random rotations, and additional techniques to generate the model data. With this augmented data set, batch normalization was implemented for model training to develop a model that could detect nuanced differences. Ultimately, the ResNet model overfit toward the classification of hurricanes and was unable to correctly classify the false hurricane labelled images.The AlexNet model correctly classified these same images with 75 percent accuracy. This final result proved more so an experiment on a classifier network’s ability to generate improved feature detection when given prior false positive, and true positive data. Additionally, the results proved that when remote sensing rare phenomena use a representative, and stratified sample.