Machine Learning Models in Wildlife Population Estimation

Abstract

Accurate wildlife population estimation is fundamental to evidence-based conservation management, yet traditional methods such as mark-recapture, distance sampling, and aerial transect surveys are logistically demanding, expensive, and often limited in spatial and temporal coverage. Machine learning (ML) approaches -- including random forests, gradient boosting machines, convolutional neural networks (CNNs), and recurrent neural networks (RNNs) -- offer scalable alternatives by integrating heterogeneous data streams such as camera-trap images, acoustic recordings, satellite remote sensing, and citizen-science occurrence records into unified predictive frameworks. This study benchmarked seven ML architectures against conventional distance-sampling baselines across six target species representing diverse taxa, survey contexts, and data availability scenarios: African elephant (Loxodonta africana), snow leopard (Panthera uncia), Amur tiger (Panthera tigris altaica), European wolf (Canis lupus), humpback whale (Megaptera novaeangliae), and Iberian lynx (Lynx pardinus). CNN-based image classification applied to camera-trap arrays achieved the highest overall accuracy for terrestrial megafauna (mean MAE = 4.8% +- 1.2% across species), outperforming conventional distance sampling (MAE = 9.4% +- 2.8%) and random forest models (MAE = 6.7% +- 1.9%). Gradient boosting machines integrating multi-source environmental covariates provided the best performance for cryptic and low-density species where camera-trap encounter rates were insufficient for deep learning. Ensemble approaches combining CNN detections with spatially explicit occupancy models reduced population estimate uncertainty by 38.4% relative to single-method baselines. These findings demonstrate that ML-integrated survey pipelines can substantially improve the precision, cost-efficiency, and geographic scalability of wildlife population monitoring for conservation decision-making.