There are many approaches to sampling “in the wild,” where the subjects of interest are mobile and free to move. As I understand it, marine surveys are particularly challenging, because the distribution of subjects across an area is typically very “patchy” – the subjects of interest are clustered, and the clusters are distributed non-randomly across large areas. Start with this reading, which is on Research Gate: https://www.researchgate.net/publication/303683508_Sampling_techniques_for_molluscan_fauna

The approach you choose will be shaped by what you are measuring and what you are trying to test or explain. For instance, if you wish to estimate the occurrence of the nudibranchs in a well-defined space – you are doing a census of some area – you might want to establish a reference grid and then count everything in randomly sampled cells within the grid. This would yield a very rough estimate, because the nudibranchs’ presence and activity probably depend on a host of unknown factors (e.g., weather, tides, sun angle, water temperature), so any given sample will be problematic. In that case, multiple samples spread over hours and days would be wise, so that your final dataset reflects the various unknown conditions that shape the subjects’ presence and activity.

On the other hand, if your research questions pertain to specific nudibranch behavior you may want to do a controlled experiment and worry more about internal validity than external validity or ecological validity in the short run. For instance, if nudibranch foraging behavior is not random but driven by odors/chemicals in the water, you would expect to see different numbers of subjects in the area when a scent was introduced experimentally. For example, if you defined a fixed area, say a 3-meter square cell, you can introduce different conditions and measure the nudibranchs behavior under different conditions. Note: the random ordering of the treatment conditions will be important.

Measuring behavior “in the wild” is challenging, because there is so much going on at once. So visual records (e.g., underwater photographs) would be useful. Set up a camera over a sampled site and record activity at a specified rate (time sampling). You can manually code the behavior or count the subjects, if the dataset is small. But if you have a large number of observations, consider using machine learning to automate the process. There are lots of open-source machine learning tools available for object recognition (e.g., OpenCV , Python, R). Also, AWS offers an affordable commercial object recognition service, and Google (Cloud AutoML) also automates custom training from your tagged images.

If you provide more details about your precise research goals, I am sure the community can hone in on specific sampling recommendations or research design ideas. Good luck!

Gary Bridge thank you so much for your answer.