From object detection to modeling spatial intensity of marine biodiversity: joining machine learn- ing and bayesian statistics

In ecology, photogrammetry is a crucial method for efficiently acquiring non-destructive samples of natural environments. When the goal is to estimate the spatial distribution of animals, detecting objects in large-scale images becomes essential. Object detection models enable large-scale analysis but introduce uncertainty, as the probability of detection depends on various factors. A key aspect of this process is the selection of the confidence threshold used during detection. A conservative threshold ensures high precision but reduces sensitivity, which can lead to an underestimation of community size and bias in species distribution models. In our proposal, we utilize YOLOv11; however, the main advantage of our approach is its flexibility, allowing us to use any detector. To address detection bias, we model the distribution of holothurians (sea cucumbers) in an area near the coast of Giglio Island using a Thinned Non-Homogeneous Poisson Process (NHPP). We assume that a ”true” intensity function accurately describes the distribution, while the observed process, resulting from independent thinning, is represented by a ”degraded” intensity. The detection function regulates the thinning mechanism, influenced by the object’s location and other detection-related features.