Automation of unmanned counting for spawners of chum and coho salmon with methods of artificial intelligence

Published methods of unmanned counting of pacific salmon allow to obtain high-quality photographic materials, but their processing is a rather hard labor. Practice of using the artificial intelligence methods, in particular deep learning models (otherwise — artificial neural networks), for this purpose is still insufficient, preliminary and concerns the counts of post-spawn chum salmon only. In this study, a deep learning model was created capable for effective automatic species-specific detection of chum and coho salmon spawners and post-spawn chum salmon in multi-species aggregations using enumerated data of unmanned surveys. This neural network is based on wide training materials and is able to process simultaneously images with several classes of objects and identify all of them. Detailed and reproducible technical description of the preparation, quality control, and operation with the deep learning model for pacific salmon is presented for a desktop GIS environment. Various factors affecting quality of the neural network are described, and recommendations for improving its work are given. Quality indicators of the model should be considered in the context of visual characteristics of target objects in processed photographic materials, which can vary significantly. Quality of the neural network output can be improved by using functionality of geoinformation software. Recommendations are given for optimizing the model development using the tools of photogrammetric and geoinformation software. Scheme for recalculating the neural network output based on processed photographic materials is proposed to obtain a corrected estimate of the total number of objects in the surveyed water area.