Using Remote Sensing Imagery and Machine Learning to Predict Poaching in Wildlife Conservation Parks

Rachel GuoAAAI 2021 · DOI

Introduction

Illegal poaching remains a critical threat to global biodiversity. Wildlife rangers, tasked with protecting vast and often remote territories, are frequently under-resourced and overwhelmed. While machine learning tools like the Protection Assistant for Wildlife Security (PAWS) have been developed to predict high-risk poaching areas, they often rely on high-quality geospatial data that many parks lack. This creates a "data divide" where only well-funded parks can leverage advanced AI.

Global distribution of wildlife parks that could benefit from automated data pipelines.

This research introduces an automated pipeline that extracts free, publicly available satellite data from Google Earth Engine (GEE) to bridge this gap. By providing a scalable way to generate predictive features, the authors aim to democratize AI-driven conservation tools for under-resourced parks worldwide.

Methodology

The proposed system automates the extraction of 16 key environmental features from GEE’s global satellite catalog. These features are categorized into:

  • Static Features: Including land cover, surface water, and elevation.
  • Dynamic Features: Capturing seasonal variations through monthly temperature, precipitation, and net primary productivity.

To process this data, park boundaries are discretized into a 1x1 km grid. High-resolution satellite imagery is rasterized to match this resolution, and dynamic climate data is aggregated into three-month intervals to capture seasonal trends.

The core predictive model, iWare-E (imperfect-observation aware Ensemble model), addresses the inherent bias in poaching data. Since rangers only record poaching activity in areas they patrol, the dataset contains many "false negatives" (areas where poaching occurred but was not observed). iWare-E uses an ensemble of decision trees to account for this uncertainty, providing more robust risk assessments.

Experimental Evaluation

The pipeline was evaluated across three diverse national parks: Murchison Falls (MFNP) and Queen Elizabeth (QENP) in Uganda, and Cross River (CRNP) in Nigeria. The study compared the performance of models trained on manually curated "Baseline" features against those using only "GEE-extracted" features.

Poaching risk maps generated for MFNP and QENP using the automated pipeline.

Key Results

The researchers used the Area Under the ROC Curve (AUC) to measure predictive accuracy:

  • Performance Parity: In MFNP and QENP, GEE-only models achieved AUCs of 0.665 and 0.706, respectively—nearly matching the performance of models using expensive, manually curated baseline data (0.687 and 0.712).
  • Superior Performance in Data-Poor Environments: In CRNP, the GEE-only model actually outperformed the baseline (0.710 vs 0.698), suggesting that satellite data can capture more relevant environmental context than limited local records.
  • Synergistic Effects: While combining both GEE and local features yielded the highest overall performance, the GEE-only results demonstrate that high-quality predictions are possible even without local geospatial infrastructure.

Discussion and Future Work

The primary contribution of this work is the democratization of AI for conservation. By replacing bespoke geospatial data with automated remote sensing, the PAWS framework becomes accessible to the hundreds of parks currently using the SMART (Spatial Monitoring and Reporting Tool) software. Furthermore, the inclusion of dynamic features allows the model to account for seasonal poaching patterns, such as shifts in movement during rainy seasons or changes in vegetation cover.

Despite these advances, the model still requires historical patrol data (labels) to learn. Future research will explore the use of Convolutional Neural Networks (CNNs) to extract even more nuanced spatial patterns directly from raw satellite imagery, potentially reducing the reliance on discretized features and further improving the accuracy of poaching risk maps.

Summary updated

BibTeX
@inproceedings{guo2021using,
  author = {Rachel Guo},
  title = {{Using Remote Sensing Imagery and Machine Learning to Predict Poaching in Wildlife Conservation Parks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17977},
}
Figure