Graph-attention Bayesian modelling of zero-inflated waterbird counts: an informatics framework for habitat-quality indicators in irrigation pondscapes

Turning imperfect ecological survey counts into reliable habitat-quality indicators is a key ecological informatics challenge because field observations are overdispersed, often zero-inflated, spatially dependent, and limited in sample size. This repository accompanies the study of wintering waterbirds in Taiwan's Taoyuan Tableland (45 irrigation ponds, 9,485 individuals, 22 species) and provides the complete reproducible workflow, including data processing, statistical analyses, graph construction, GATv2–Bayesian–ZINB model implementation, variational inference, and figure generation. The proposed framework combines a graph-attention (GATv2) encoder with an adaptively generated Gaussian Bayesian prior, whose parameters are learned from graph embeddings rather than specified subjectively, under a zero-inflated negative binomial likelihood. Results show that modelling overdispersion is essential, whereas explicit zero inflation provides little additional benefit. Shoreline complexity (perimeter–area fractal dimension, MPFD) emerges as a contrasting habitat-quality indicator, being negatively associated with dominant open-water waterbird abundance while positively associated with community evenness. All code, processed data, and documentation are released to support transparency, reproducibility, and future applications of graph-based Bayesian models for ecological indicator analyses.