Hybrid Expert-Augmented Active Learning for Enhanced Electronic Records Management in Uganda’s Wildlife Sector

Hybrid Expert-Augmented Active Learning

Abstract

We propose a hybrid expert-augmented active learning framework to reformulate Uganda’s wildlife electronic records management system, addressing the critical challenges of data quality and decision-making efficiency in conservation efforts. The proposed method integrates a Bayesian neural network with human-in-the-loop annotation, dynamically prioritising uncertain records for expert validation while autonomously processing high-confidence data. The system consists of three core modules: an uncertainty-aware data ingestion layer that quantifies prediction reliability, a mobile-optimised expert interface for real-time annotation, and an adaptive training loop that incrementally refines the model using newly validated records. Moreover, the architecture substitutes conventional data pipelines by routing ambiguous inputs to human experts and archiving only machine-confident entries, thereby reducing noise in the central database. The implementation employs a Monte Carlo dropout transformer for robust uncertainty estimation and federated learning to aggregate distributed expert inputs without centralised data pooling. Unlike static systems, our approach establishes a closed-loop feedback mechanism between data quality and model performance, enabling continuous improvement in predictive accuracy and operational decision-making. The novelty lies in its context-aware annotation workflow and dynamic prioritisation of expert effort, which are tailored to the sparse and heterogeneous nature of wildlife data in resource-constrained environments. Field deployments demonstrate significant improvements in species identification accuracy and threat assessment reliability, highlighting the framework’s potential to transform electronic records management in conservation sectors globally.