Predator Highway Intelligence (PHI)

AI-powered biodiversity intelligence to map mortality niches, eliminate predator-enhanced hotspots, and ensure infrastructure projects protect native species.

Comment

Inspiration

Roads aren't just barriers or physical hazards—they are active "interaction landscapes" that fundamentally alter how species compete.

Recent ecological research highlights that resource-rich road corridors act as selective filters rather than indiscriminate hazards. Human modification and food subsidies along road verges attract both native wildlife and invasive mesopredators like free-ranging cats. However, the spatial usage is starkly asymmetric: synanthropic invasive predators dominate the human-modified core of road corridors. This creates a realised-niche "squeeze", forcing native competitors into less-suitable marginal habitats.

Traditional road mitigation strategies like crossings or fencing are often built reactively without accounting for predator dynamics, inadvertently creating "predator highways" that favor invasive species over native wildlife. We built Predator Highway Intelligence (PHI) to transform over 12,000 verified roadkill records into actionable intelligence—empowering planners to build biodiversity-smart infrastructure.

What it does

PHI is an AI-powered spatial analytics platform designed for infrastructure planners, conservation NGOs, and government ecological departments. It moves road ecology from passive tracking to proactive, AI-driven intervention:

  • Mortality Niche Mapper: Predicts spatial collision risk surfaces for invasive predators versus native competitors using stacked ML models.
  • Hotspot Detector: Pinpoints "predator-enhanced hotspots" where competitive overlap and invasive dominance are most severe.
  • Subsidy & Priority Engine: Identifies anthropogenic food sources along verges and ranks high-biodiversity corridors for targeted invasive control.
  • Mitigation Designer & Simulator: Recommends optimal sites for smart crossings, fencing, and food subsidy reductions, allowing users to model "what-if" mitigation scenarios before breaking ground.
  • Citizen Science Integrator: Allows field teams and citizen scientists to upload roadkill images, utilizing Gemini Vision to auto-identify species and log GPS coordinates.

How we built it

We engineered PHI as a modern, cloud-native geospatial platform optimized for high-throughput spatial inference:

  • Frontend: Built with Next.js 16 (App Router), React 19, and TypeScript 5.9+, styled using Tailwind CSS v4 and Shadcn UI.
  • Geospatial & Visualizations: Map interfaces are powered by MapLibre GL JS with WebGL rendering for fluid heatmap overlays, alongside Recharts and D3.js for ecological metric analytics.
  • Backend Architecture: High-performance Fastify (Node.js) microservices running on Google Cloud Run, managed via Apigee API Gateway.
  • Data Layer: PostgreSQL with PostGIS extensions managed via Prisma ORM for geospatial queries, paired with Cloud Storage for assets and BigQuery for large-scale analytical aggregation.
  • AI & Computer Vision: Integrated Vertex AI for Random Forest predictive modeling using positional encodings and environmental covariates. We leveraged Gemini Vision for automated roadkill species verification and Gemini Pro for automated, EIA-compliant policy report generation.

Challenges we ran into

  1. Geospatial Machine Learning Latency: Training models on over 12,000 spatial records with multiple environmental raster layers (Global Human Modification indices, prey density gradients) caused initial rendering bottlenecks. We solved this by rasterizing risk surfaces into optimized vector tiles and caching tile computations in Redis (Memorystore).
  2. Spatial Asymmetry in Invasive Species: Modeling the niche "squeeze" required capturing delicate competitive overlaps ($D = 0.515$) without overfitting. We implemented spatial cross-validation and stacked spatial positional encodings within Vertex AI to accurately model human-modification gradients.
  3. In-the-Field Species Identification: Field roadkill photos are frequently motion-blurred or degraded. Fine-tuning our vision pipeline with Gemini Vision to parse low-quality imagery while filtering out false reports required strict prompt engineering and confidence scoring thresholds.

Accomplishments that we're proud of

  • High Model Accuracy: Achieved an AUC > 0.85 for predicting invasive cat mortality niches and AUC > 0.75 for native competitors across diverse road networks.
  • Sub-60 Second Pipeline: Developed an end-to-end processing pipeline capable of analyzing a 100 km road alignment file and generating a full risk surface map in under 60 seconds.
  • End-to-End Mitigation Workflow: Successfully linked complex ecological research (Li et al., 2026) directly to actionable engineering outputs—enabling non-ecologist infrastructure planners to design effective wildlife crossings.

What we learned

  • Roads are Complex Systems: Infrastructure planning cannot treat ecological barriers as static line strings; road verges act as artificial ecosystems that heavily favor generalist and synanthropic predators.
  • Interdisciplinary AI Design: Translating scientific literature into software features requires bridging the gap between statistical ecology metrics (e.g., realised-niche overlap) and practical decision-making interfaces for highway engineers.
  • Serverless Spatial Architectures: Modern cloud infrastructure like Cloud Run paired with PostGIS enables scaling heavy spatial ML workloads without maintaining expensive idle servers.

What's next for Predator Highway Intelligence (PHI)

  • Pilot Deployments: Launching field pilots across 3 major infrastructure projects in partnership with regional conservation bodies.
  • Advanced Scenario Simulator: Upgrading our simulation engine to incorporate agent-based modeling (ABM) for multi-year population dynamics post-mitigation.
  • Global Raster Coverage: Expanding our automated environmental covariate ingestion to provide out-of-the-box spatial intelligence across 5 additional countries.
  • Enterprise Integration & API: Exposing PHI’s core prediction models via open REST/GraphQL APIs for seamless integration with existing GIS desktop applications like QGIS and ArcGIS.

Built With

Share this project:

Updates

Submission history