Our MaskSDM paper is out in Methods in Ecology and Evolution🥳
https://t.co/pUPx67RySy
MaskSDM is a multimodal masked-modeling method predicting species distributions from any subset of variables or modalities
Nina van Tiel @gencersumbul@ChiaraVanalli@b_kellenb@devistuia
🧵
@gencersumbul@ChiaraVanalli@b_kellenb@devistuia Huge thanks to my coauthors and collaborators for making this possible🙌
MaskSDM is just the beginning: there’s a lot more to explore in combining AI and ecology to tackle the biodiversity crisis💚🌍
Our MaskSDM paper is out in Methods in Ecology and Evolution🥳
https://t.co/pUPx67RySy
MaskSDM is a multimodal masked-modeling method predicting species distributions from any subset of variables or modalities
Nina van Tiel @gencersumbul@ChiaraVanalli@b_kellenb@devistuia
🧵
@gencersumbul@ChiaraVanalli@b_kellenb@devistuia Many more results are available in the paper (and in its Appendix):
📄 https://t.co/pUPx67RySy
We hope MaskSDM becomes a practical tool for ecologists and conservation scientists:
💻 https://t.co/koKaWimDhN
@gencersumbul@ChiaraVanalli@b_kellenb@devistuia A major contribution is our new Shapley value computation method, which avoids common linear assumptions by leveraging MaskSDM’s flexible input design. This provides more precise insights into how different environmental factors shape species distributions, locally and globally📊
@gencersumbul@ChiaraVanalli@b_kellenb@devistuia A single MaskSDM model performs nearly as well on each tested subset of inputs as an oracle model trained specifically on that subset. This makes it possible to obtain predictions, performance metrics, and maps for any variable subset using only simple inference passes 📈🗺️
@gencersumbul@ChiaraVanalli@b_kellenb@devistuia Leveraging MaskSDM, we modeled the distributions of 12,738 plant species worldwide using the sPlotOpen dataset. MaskSDM can be applied anywhere and adapts to the data available, making it ideal for global biodiversity assessments 🌍🌏🌎
@gencersumbul@ChiaraVanalli@b_kellenb@devistuia MaskSDM uses transformer-based masked modeling (e.g., BERT, MAE, 4M) adapted for ecology. This lets the model learn from incomplete inputs and still predicts reliably. It’s also multimodal: tabular data, satellite images, time series, and more. More coming soon… stay tuned😉
MaskSDM brings three key advantages for species distribution modeling (SDM):
🧩 Flexibility — choose which variables and modalities to use
⚙️ Robustness — works even with missing data
🔍 Explainability — a new Shapley values method shows which factors matter most for each species
@VictorKristof Je consulte désormais toujours Predikon en premier, car les estimations sont beaucoup plus fiables que celles de GFS Bern qui a annoncé à 12h01 que BienPropre serait accepté !
I would like to think about this as an invitation that physics extends towards AI/ML to think like physicists: -"Let's figure out how the world works". This is unlike computer scientists, who often see worst cases and adversaries. Nor mathematicians who need theorems everywhere.
Interested in underwater SfM/SLAM?🪸📸🌊🤿
I will present our approach to remove detrimental caustics and backscatter from single images at #ECCV2024 in Milan this week. Work done at ECEO @EPFL_en with @devistuia!
Come say hi on Thursday Oct 3 at 16:30! https://t.co/WQBPwQ1zTn
🚨 job alert!
We just opened a PhD position in my lab for a crazy interesting and important project with @ETH_en and the @ICRC
We will look into new ways to use nightlight 🛰 data for characterising humanitarian crisis situations!
Apply: https://t.co/vbTTV1Qd3o
@EssTechEPFL
Today, I have started my new position as Lecturer/researcher in the People & Nature Lab at University College London (@ucl, @UCLCBER )!
I will be working with fabulous @ProfKateJones on technology and machine learning for environmental good. Stay tuned for news and opportunities!
What is the best time to go to lunch to avoid queues at EPFL? 🕧🥗
Are more energy drinks and coffees bought during exam sessions? 🥤☕
These are the types of questions we explored in this study, to which I contributed during my master's!
Check out our new work, now out in Frontiers in Nutrition:
"Measuring and Shaping the Nutritional Environment via Food Sales Logs: Case Studies of Campus-Wide Food Choice and a Call to Action"! https://t.co/T6WNFLgckS
Our paper on pseudo-absences in Species Distribution Modeling with Deep Learning is published!
https://t.co/ejbr2CqFOW
With @nina_vTiel @b_kellenb@LloydHughesZA@devistuia, we address geographic biases, class imbalance, and presence-only samples in citizen science datasets
🧵
7/7
Additionally, take a look at our other paper (https://t.co/QZ2elTG9zi), presented at the CCAI workshop at ICLR 2024; we tested our loss on other citizen science datasets, achieving great performance on GeoLifeCLEF 2023.
Code: https://t.co/iAJ8Sz3i92
Feel free to reach out!