When building Species Distribution Models (SDMs), researchers need “background” points to represent the available environment. In data-rich regions, “target-group backgrounds” are used to cancel out sampling bias.
However, in data-poor regions (like many parts of the Global South), target-group data does not exist. Randomly sampling background points causes the model to learn “where the roads are” rather than “where the species lives.”
biasProxyRbiasProxyR solves this by using open-source
infrastructure data (like OpenStreetMap road networks) to create a
“Sampling Effort Proxy” surface. This allows researchers to sample
background points that mathematically match the geographic bias of their
occurrence data.
First, we generate a sampling effort proxy for a country. For this tutorial, we will use Luxembourg because it is small and processes quickly.
``{r generate-effort, message=TRUE, cache=TRUE} #
Generate a 1000m resolution effort surface for Luxembourg effort <-
generate_effort_surface(“Luxembourg”, resolution = 1000)
terra::plot(effort, main = “Sampling Effort Proxy - Luxembourg”, col = terrain.colors(50))
*Brighter areas indicate higher probability of human sampling effort (e.g., near cities and major highways).*
## Step 2: Sample Bias-Aware Background Points
Instead of sampling background points randomly, we use the effort surface to weight our sampling. This ensures our background points have the same geographic bias as typical GBIF occurrence data.
```{r sample-bg, message=TRUE, cache=TRUE}`
# Sample 2,000 bias-aware background points
bg_points <- sample_bias_aware_background(effort, n_points = 2000)
# Plot the points over the effort surface
terra::plot(effort, main = "Bias-Aware Background Points", col = terrain.colors(50))
plot(bg_points, add = TRUE, col = "red", pch = 16, cex = 0.5)
If you already have occurrence data, you can use the diagnostic tool to see how biased it is before you even run an SDM.
``{r diagnose, message=TRUE, cache=TRUE} # Run the
diagnostic on our sampled points (pretending they are occurrence data)
diagnostic_results <- diagnose_sampling_bias(bg_points, effort)
print(diagnostic_results$bias_score_percent) ```
Now that you have bias-aware background points, you can feed them
directly into your favorite SDM package (like maxnet,
biomod2, or sdm) alongside your true presence
data to build ecologically realistic models!