A Predictive Mobility Model Based on Crocodile Mobility Data for Human-Crocodile Conflict Management in Namibia

Abstract

Uuman-Wildlife Conflict (HWC) is a dynamic and complex issue of global magnitude which has negatively impacted the livelihoods of both humans and wildlife. In Namibia, the escalating and increasingly devastating human-crocodile conflicts in the Kavango-East Region have culminated into numerous incidents over the years, leading to the intolerable loss of both human and wildlife lives. It has been an important and long-time interest of the conservation community to quantify and characterize wildlife movements as a fundamental element in the study of wildlife for sustainable conservation management, but the conventional approaches to studying wildlife movement offered sparse and limited knowledge on wildlife behaviour. With technological advancements, new approaches to studying wildlife movement emerged utilising geospatial technologies like Geographical Positioning Systems (GPS) to collect mobility data. However, GPS is predisposed to technical limitations that can hinder it from accurately determining the location of a moving object causing gaps in the mobility data. This study presents a predictive location model based on historical GPS mobility data of crocodiles' movement in the Kavango region in Namibia to bridge the gaps in the mobility data when GPS technology fails by leveraging the versatility of Machine Learning (ML) models. Several individual supervised ML models were trained and tested on the crocodile GPS mobility dataset, and the best-performing regression algorithms were combined into several ensemble ML models to optimise prediction accuracy. The best-performing ensemble ML model is a combination of Extra Trees Regressor, Decision Tree Regressor and XGBoost Regressor referred to as edXGBoost Regressor. The results indicated that the edXGBoost Regressor model outperformed the individual and alternative ensemble ML models evaluated in this study by achieving an R2 and RMSE score of 93% and 0.003 respectively making it the optimal model.