Evaluating spatial resolution and imperfect detection effects on the predictive performance of inhomogeneous spatial point process models trained with simulated presence-only data

Species distribution models (SDMs) are crucial in ecology, conservation, and ecosystem management. Numerous SDMs have been developed over time, and studies have shown that these tools can be affected by a range of factors, such as data type, spatial resolution, number of explanatory variables, sample characteristics, and collinearity between environmental variables. New SDMs have often been developed to address some of these issues. Understanding the performance of new statistical tools is crucial for researchers in various fields. Thus, we assessed the predictive ability of the Poisson point process and log Gaussian Cox process models, considered as new SDMs, by simulating two factors, spatial resolution and imperfect detection, which are likely to have significant effects on SDMs, and considering Gabon as the study area. The observed model performance metrics, such as the Area Under the Receiver Operating Characteristic Curve (AUC), Mean Absolute Error (MAE), and Pearson correlation (CORR) between the true and predicted intensities, were used to evaluate the predictive performance of these models. The results showed that, although most of these models failed to estimate the intercept \({\alpha }_{0}\) ${\alpha }_0$ and covariate coefficients ( \({\varvec{\beta }}_{x}{\textbf {x}}(z)\) ${\mathit{\beta }}_x\mathbf{x}(z)$ ) correctly, they at least had the merit of demonstrated good performance (AUC more than 70%, CORR more than 67%, and MAE less than 0.61%). However, the spatial resolution of the environmental variables and imperfect detection of simulated species occurrences significantly affected the predictive performance of the two models (P < .0001). This study offers important insights for ecologist modellers, environmentalists, and conservators.