MusselFinder: Automated dreissenid mussel density estimation via robotic benthic imaging

Abstract

Despite the widespread need to map the abundance of invasive quagga and zebr mussels ( Dreissena spp.) in lakes and reservoirs worldwide, current methods remain limited by data collection and analysis bottlenecks. Autonomous underwater vehicles (AUVs) combined with machine learning (ML) offer a scalable remote sensing approach for estimating mussel density (individuals m −2 ) across large spatial domains and a range of habitats that are difficult to sample with current techniques. We present MusselFinder: a deep neural network (DNN)-based pipeline that detects and enumerates live mussels in AUV-collected benthic images. MusselFinder estimates live mussel percent cover and density in nadir monocular images captured at known altitudes above the lakebed. In images of suitable clarity, MusselFinder achieved strong pixel-level localization (mussel-class Intersection over Union [IoU] > 0.76) and density prediction (log-density R 2 = 0.79–0.95) over a variety of bottom types. AUV-based image acquisition is more than three orders of magnitude faster than physical sampling by divers or ponar grabs, and DNN-based analysis can proceed at rates comparable to image acquisition. We used MusselFinder to analyze over 50,000 spatially referenced images at a spawning reef in Lake Michigan, representing a shift from small-scale, diver-collected datasets to large-scale monitoring in unconstrained environments. The method can be adapted to new imaging conditions using a modest training set and a user-friendly point-labeling technique, supporting broader-scale monitoring of one of the world's most invasive invertebrates.