Taking machine learning with a grain of sand: Sediment Analysis Neural‐network Data‐engine (SAND‐e) reveals sedimentological differences between turbid and clear‐water reefs

Abstract Sediment is an important facet of sand cay reefs as it is responsible for reef accretion and island formation, with shifts in the proportions of sediment producers being proxies for ecological shifts. However, manual sediment analyses require experts to identify thousands of sand grains by hand before beginning data analysis. To accelerate the process, we developed the Sediment Analysis Neural‐network Data‐engine (SAND‐e) to estimate the proportions of sediment producers, based on segmentation and classification of carbonate sand grains from microscope camera imagery. Sediment from Darvel Bay was used for training due to the variability of sand cay reefs available in that area. SAND‐e segmented 1686 images into 32 883 grains within 3.5 h. The grains were then fed through SAND‐e's classifier ensemble, containing four classifiers that voted to classify the grains into one of five classes (calcareous algae, coral, foraminifera, molluscs and ‘other’) in 1 hour. Both SAND‐e and 11 humans annotated grains from the same dataset to ensure that SAND‐e's accuracy was within the already accepted error rate deriving from multiple human annotators.