Comparing random forests and convoluted neural networks for mapping ghost crab burrows using imagery from an unmanned aerial vehicle

Sandy beaches are important ecosystems that line a third of the world's ice-free coastlines. Unfortunately, these environments and the life they support are often threatened by various anthropogenic and natural factors. Monitoring sandy beach health is important to aid in appropriate management decisions. One such method of quantifying environmental health is using bioindicators. Ghost crabs are a commonly used sandy beach bioindicator. Current techniques for assessing ghost crab abundance and distribution data involve manually counting each individual burrow opening, which can intrusive and timely for a large area. The aim of this study was to assesses the use of imagery from an unmanned aerial vehicle and machine-learning algorithms as an alternative approach to monitoring ghost crab burrows. The accuracy and transferability of random forest (RF) and convolutional neural networks (CNN) classifiers within an Object-Based Image Analysis (OBIA) framework were tested using hyper-resolution (0.04 m) orthomosaics from four different dates. CNN was a more robust classifier with higher accuracies (max F-score 0.84). Transferability of rule sets and models was limited for both classifiers, particularly when applied to sub-optimal imagery. Overall, we present a feasible workflow that provides ecologist and environmental managers with a cost-effective and less invasive alternative to mapping ghost crab burrows.