Direct application of residual neural network to riverine aerial photography for estimating fish distribution

Morphology and hydraulic condition of rivers explain habitat and actual species distribution of fish species. However, detailed ground surveys to obtain such habitat information are generally complicated and costly. Therefore, we aimed to explore the possibility of integrating aerial photographs and image recognition technique as a supplementing approach for field surveys. For this purpose, we focused on one benthic species (Acanthogobius flavimanus) and two migratory species (Nipponocypris temminckii and Plecoglossus altivelis) as representative species. Their distribution in the Kanto region of Japan was obtained from the national census on river environments-riparian zone, while aerial photographs of the corresponding river sections were collected from Geographical Survey Institute of Japan. Then, convolutional neural network (CNN) was applied to model the fish distribution based on the riverine photographs. As per the results from hypothesis tests, CNN was capable of learning relevant attributes of the river channel appearance for distribution of A. flavimanus. The model performance was significantly correlated with the number of the training data for this species. For this species, the relatively dark water surface and wide channel width seem to be possible key factors. At the same time, for the other two species, this modeling approach was not as successful as A. flavimanus, while the model performance did not show significant differences among the three species. Although practical application of this approach is still challenging in terms of available data and model validity, it is worthwhile further exploring its applicability to other riverine species and regions.