JellyNet: The convolutional neural network jellyfish bloom detector
Ben Mcilwaine, Mónica Rivas Casado
Abstract
Coastal industries face disruption on a global scale due to the threat of large blooms of jellyfish. They can decimate coastal fisheries and clog the water intake systems of desalination and nuclear power plants. This can lead to losses of revenue and power output. This paper presents JellyNet: a convolutional neural network (CNN) jellyfish bloom detection model trained on high resolution remote sensing imagery collected by unmanned aerial vehicles (UAVs). JellyNet provides the detection capability for an early (6–8 h) bloom warning system. 1539 images were collected from flights at 2 locations: Croabh Haven, UK and Pruth Bay, Canada. The training/test dataset was manually labelled, and split into two classes: ‘Bloom present’ and ‘No bloom present’. 500 × 500 pixel images were used to increase fine-grained pattern detection of the jellyfish blooms. Model testing was completed using a 75/25% training/test split with hyperparameters selected prior to model training using a held-out validation dataset. Transfer learning using VGG-16 architecture, and a jellyfish bloom specific binary classifier surpassed an accuracy of 90%. Test model performance peaked at 97.5% accuracy. This paper exhibits the first example of a high resolution, multi-sensor jellyfish bloom detection capability, with integrated robustness from two oceans to tackle real world detection challenges.