Machine Learning Prediction of Human–Snake Encounters in Expanding Urban Areas

Abstract

The rapid urban growth is turning scrubland, wetlands and agricultural fringes and forest edges into impervious surfaces, fragmenting snake habitat and putting snakes and people in closer contact with one another. Most cities still respond reactively to these encounters, sending out rescuers only when they are notified that a snake was found, and traditional risk maps are static and don't account for the changes in risk seasonally or monthly caused by rain, or by building activity. This paper introduces an interpretable machine-learning model to forecast the location and time of human–snake interactions in an urban area that is growing in size. Ten-year snakebite emergency reports (2016–2025) were overlain at 500 m × 500 m resolution with monthly verified rescue and encounter data (n = 52,630) as well as land-use, vegetation, and LST indices derived from satellite imagery, infrastructure and population and waste indicators. Four classifiers were compared: Logistic Regression, Random Forest, Support Vector Machine and XGBoost, and a multilayer perceptron, with spatial-block and temporal validation. XGBoost achieved the best performance on the held-out test period (2024–2025) and the most important features were: historical encounter frequency, antecedent rainfall, built-up expansion rate, proximity to water bodies, or drains. The predicted probabilities were mapped as low, moderate and high-risk class and displayed as GIS hotspot maps with factor level explanations. The framework can help with pre-monsoon awareness campaigns, deployment of rescue teams, hospital antivenom preparedness and planning for habitat sensitivity, moving the focus of urban wildlife management from reactive response to prevention, while at the same time not encouraging the unnecessary removal of snakes