Dataset for: Colour-Agnostic Deep Learning for Automated Biofouling Assessment of Antifouling Coatings Using Soft Probability Labels

This record contains the image data and annotations accompanying the paper "Colour-Agnostic Deep Learning for Automated Biofouling Assessment of Antifouling Coatings Using Soft Probability Labels". The data supports a deep learning pipeline for automated biofouling coverage assessment on antifouling coated test panels. All panels were exposed and imaged at the CoaST Maritime Test Centre (CMTC), Hundested Harbour, Denmark, and captured underwater with a 64 MP imaging station. The pipeline composes two models, an algae model trained on continuous soft probability labels and a barnacle model trained on binary masks, producing the separate algae and animal coverage categories used for fouling resistance rating computation under ASTM D6990 and ECHA PT21. Two archives are included. algae_dataset.zip (20.27 GB) contains 1,524 annotated panel images of commercial antifouling coatings across five substrate colours, blue, green, grey, red, and reflective. Each image is paired with a continuous soft probability label map exported from ilastik, where every pixel encodes a fouling probability from 0 (confident clean) to 1 (confident fouled). The set is split into 1,067 training and 457 validation images, stratified to give approximately equal colour representation in both subsets. barnacles_dataset.zip (138.17 MB) contains 225 panel images annotated with binary barnacle segmentation masks. Annotations span all five substrate colours, with green panels withheld entirely from training and reserved as a 55 panel validation set on a substrate colour absent from the training distribution. The annotations are reference annotations produced by a single expert annotator in ilastik, not external ground truth. For the algae data the continuous probability maps preserve the diffuse, gradient nature of algal fouling boundaries. For the barnacle data the calcified structures carry well-defined edges, so binary masks are the appropriate representation. The code that consumes this data is published separately and linked under Related works. The accompanying manuscript is linked under Related works once a DOI is available.Financial support from the Hempel Foundation to CoaST (The Hempel Foundation Coatings Science and Technology Centre) is gratefully acknowledged. Licence: all files are released under CC BY 4.0. Attribution should cite this record and the accompanying paper.