Evaluating the Impact of Diverse Marine Environments on Image Classification Performance

Monitoring marine environments leverages various imaging technologies from synthetic aperture sonar to consumer-grade cameras. A typical task involves classifying images of various aquatic species and structures. Data quality and model performance is affected by several factors including turbidity, depth, light, backscatter, resolution, camera calibration, blur, and distance. Training data captured from a single environment was found to be insufficient for classifying coral images from diverse environments. A methodology for evaluating models determined the generalizability of models across a range of species and underwater conditions. This provided insights towards further field data collection, augmentation, and model tuning. The contributions facilitate the development of models that can be generalized for long-term, species-level monitoring of coral reefs.