Azov Sea Gobies (Gobiidae) Stock Assessment using Artificial Intelligence Approach Trained on Benthic Ecosystem Data

A novel approach to gobies stock assessment based on training an artificial intelligence model using long-term historical ecosystem data was proposed. This research uses data from the historical archive for annual parameters of the Azov Sea ecosystem, including the Don River flow volume, water salinity, zoobenthos relative biomass, sander catches, sturgeons catches and fishery biomass, gobies catches (predictor factors), gobies stock biomass (predictable variable) dynamics during the period 1970–2022. In the preliminary data analysis, no evidence of non-stationarity factors or predictable factors collinearity was found. Based on this datasets, three types of models were trained to predict the gobies stock biomass based on environmental factors: a random forest model, Bayesian regularized neural network model and generalized linear model. Lags between biomass of gobies and predictable factors was introduced from 1 to 4 years to make forecasting possible. The best results and diagnostics were observed with the RF model (R2 = 0.85 with no lags). The acceptable predictive models were found with a lag of up to 2 years based on the random forest model ( $$R_{{1 + }}^{2} = 0.81$$ , $$R_{{2 + }}^{2} = 0.73$$ ). Considering the forecast for the 2 years ahead, no prospects have been found for restoring the gobies population in the short term.