<3 <3 <3

🚀 What’s next?
We’re applying this method to 600+ alien amphibians and reptiles 🦎🐸 to map their global invasion potential & understand global biotic homogenization 🌍
(8/8)

🔧 Our fix:
We sub-sample both presences & backgrounds across the full environmental gradient available to a species.
This helps SDMs better:
📌 Explain niches
📍 Predict current distributions
🕰️ Project future/range-shifting scenarios
(7/8)

💥 What went wrong?
Environmental sampling bias!
The model didn’t capture the true species-environment relationship—it just overfit to the oversampled Australian conditions.
➡️ The data were unbalanced in environmental space, and it showed.
(6/8)

🎯 The scenario:
Goal: Predict the global invaded range of the cane toad 🐸
Data:
🇧🇷 1000 records from native Brazil
🇦🇺 2000 from invaded Australia
🇵🇭 100 from invaded Philippines

👉 All thinned using standard methods (SOA)

Result: SDMs accurately predicted AU and BR(overfit), but mehhh in PH

(5/8)

🤺 I’ve had beef with this issue since my MSc days. It all started here:
📄 doi.org/10.1038/s415...
(4/8)

✅ The Solution:
Apply 'Habitat Stratified Sampling Design' when thinning both data.

Our new methods:
1️⃣ Environmental clustering
2️⃣ Environmental distance thinning

Both outperform conventional approaches for:
🌍 Explaining
📍 Predicting
⏳ Projecting species distributions

(3/8)

🧩 The Problem:
Environmental sampling bias = when some environmental conditions are oversampled just because they’re common or widespread across the landscape.
➡️ This skews the models and messes with predictive accuracy.

(2/8)

Me too please!

Not too late for the afterparty 💜💜💜