Comparing the performance of generalized linear model (GLM) and random forest (RF) models in predicting catch distribution of Caspian Kutum (Rutilus frisii)

The present study aimed to assess the performance of generalized linear model (GLM) and random forest (RF) model in predicting Caspian Kutum (Rutilus frisii) catch distribution. Caspian Kutum catch per unit of effort (CPUE) data was used as the response variable. Remotely-sensed data of five environmental parameters were used as model predictors as well, including daily sea surface temperature (SST), chlorophyll-a concentration (CHL), aerosol optical thickness (ASL), particulate organic carbon (POC) and particulate inorganic carbon (PIC) concentrations. The coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) scores were used as measures of model performance and accuracy. The best fitted GLM had only Log(PIC) and POC as significant parameters, while the RF model contained all predictors. RF showed higher explaining potential compared to GLM (RF: R2=0.47; GLM: R2=0.053). Also, higher accuracy was observed using RF (MAE=972.4; RMSE=1326.1 (kg/hour.seine)) than GLM (MAE=1328.7; RMSE=1465.6 (kg/hour.seine)). ASL (33.31%) and CHL (28.87%) were parameters with the highest and lowest relative influence in the RF model. Based on the results, random forest modelling is suggested as a practical technique for predicting fish catch distribution.