AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-23
Robustness Verification of an Autonomous Underwater Vehicle-based Plankton Classifier
Plankton classifiers on autonomous underwater vehicles often misclassify images when there is environmental noise or non-biological debris, so marine biologists must check the results by hand. The authors propose a robustness verification framework based on reachability analysis and apply it to a continuous-time neural ODE classifier that uses images from the SilCam particle imager. Formal verification against environmental perturbations is reported to work as an automated filter that gives formal guarantees of model stability on ambiguous data, reducing post-processing workload.