BIRDGen: Multimodal Conditional Inference of Latent Unbiased Species Distributions

Richard Yuxuan Zhu, Cathy Hou, Katherine McPhie

Keywords: Amortized Inference, Multimodal Learning, Bias Correction, Species Distribution Modeling, Citizen Science

TL;DR: A multimodal deep model that predicts standardized-survey bird detection rates from environmental and citizen-science inputs, using sparse rigorous surveys as ground truth and dense biased observations as an auxiliary reconstruction signal.

Abstract: Citizen-science platforms like eBird aggregate billions of bird observations that form heavily biased distributions confounded by observer behavior. We treat these biased observations as noisy proxies for a latent unbiased bird species distribution, which can be partially observed through sparse but statistically rigorous randomized surveys conducted by domain experts. To bridge this gap, we present BIRDGen, which performs multimodal conditional inference of latent unbiased species distributions from (i) biased citizen-science observations and (ii) environmental covariates over structured geospatial data. By conditionally aligning dense biased inputs (from eBird) with sparse unbiased targets (from rigorous surveys), BIRDGen predicts debiased species distributions, demonstrating that deep latent-variable reasoning can effectively isolate ecological signal from observer bias.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 264