Predicting avian roadkill risks in Shanghai: insights from ensemble modelling of small, opportunistic data

Avian roadkill represents a pervasive yet understudied ecological concern in China. Data scarcity alongside sample biases hinder mechanistic understanding of its spatial dynamics. To address these challenges and gain insights into spatial patterns and drivers of avian roadkill, we developed an ensemble modelling framework to predict avian roadkill risks in Shanghai using 46 opportunistically recorded roadkill occurrences alongside 18 explanatory variables. We contrast the conventional ensembles of small models (ESMs) with a stacking framework to combine sub-models built upon grouped explanatory variables. Three distinct algorithms (Maximum Entropy, Boosted Regression Tree, and Random Forest) were employed for sub-model predictions, while Elastic Net Regression was used to create the meta-model. Performance evaluation (cross-validation AUC = 0.964 ± 0.027; independent testing AUC = 0.772 ± 0.025) suggests that our ensemble model could provide a useful indication of the spatial patterns and key drivers of avian roadkill occurrences even with small, opportunistic data. Predicted risk hotspots were primarily concentrated in the central urban area, the south-eastern coastal zones, and Chongming district in Shanghai, with distance to industrial sites, watercourse density, secondary road density, tertiary road density, Enhanced Vegetation Index annual range, and tree cover identified as key drivers. These results confirm the utility of our ensemble model for both risk mapping and mechanistic inference in data-limited contexts and can serve as useful tools to better understand avian roadkill dynamics and to inform targeted mitigation in conservation planning.