Generating pseudo-absences for presence-only data
When dealing with presence-only species distribution data, two main approaches are commonly used to handle the absence of true absence points. The first approach is to model the presence points directly as a point process, which treats presences as an observed spatial pattern. This approach doesn’t require absence data and instead estimates the intensity of occurrences across the study area. The second approach is to generate pseudo-absence points, creating false absences across the study region, allowing the model to learn from contrasting presence and absence locations.
In GLOSSA, we take the second approach as we fit a Bayesian Additive Regression Trees (BART) model using a binary response variable (presence = 1, absence = 0) with a probit link function. Generating pseudo-absences allows the model to estimate the probability of species presence by learning from both the observed presences and the generated pseudo-absences.
The challenge of generating pseudo-absences
Generating pseudo-absences for presence-only data is a critical decision in species distribution modeling. There exist multiple methods for generating pseudo-absences, each with its own strengths and limitations, and there is currently no consensus on a standard approach. Some common methods include random sampling, geographic exclusion (e.g., excluding areas close to presence points), and environmental filtering (e.g., selecting pseudo-absences in contrasting environmental conditions).
Each method can influence model outcomes differently, depending on the study area, species, and modeling objectives. Given the absence of a definitive standard, GLOSSA uses a balanced random sampling approach that provides broad applicability and ease of use across diverse datasets.
Pseudo-absence generation with GLOSSA
When you upload an occurrence file with presence-only data (the pa column just includes ones or is missing), GLOSSA uses a balanced random sampling approach for generating pseudo-absences, that is, GLOSSA generates pseudo-absence points randomly across the defined study area, ensuring that the number of pseudo-absence points matches the number of presence points.
The spatial extent within which pseudo-absences are generated is an essential factor.
When the presence data includes several timestamps, GLOSSA generates the same number of pseudo-absences as presences for each time period. This approach ensures that the model captures temporal variation in species occurrence, which can be important for species with temporal patterns.
If you prefer an alternative method for generating pseudo-absences, you can create them outside of GLOSSA using your preferred method and include them in your occurrence putting a 0 in the pa column of these records.