Local Evidence Base
Grounded in Ugandan visual realities with validated edge-AI performance.
Grounded in Ugandan Visual Realities
Urban infrastructure AI cannot be trained on synthetic, imported assumptions. The Drain platform operates on a localized, ontology-driven evidence base. To construct our baseline architecture, our engineering team manually collected and annotated a highly specialized dataset of 320 high-resolution images covering drainage channels across Kampala under diverse weather and lighting conditions.
Dataset Overview
Localized Urban Drainage Dataset
Drainage Channel Classifications
Validation Results
Rigorous Multi-Model Evaluation
Model Performance Comparison
YOLOv8-Nano vs. YOLOv8-Small on urban drainage classification
Implementation Methodology
How to Pilot the Drain Platform
For municipal authorities (such as KCCA) and private infrastructure contractors seeking to transition from reactive patrols to predictive analytics, we provide a structured, low-risk implementation pathway. The Drain platform is deployed through a comprehensive 12-week institutional pilot program.
Integration & Baseline Mapping
Secure API integration and deployment of lightweight YOLOv8-Nano models onto the municipality's existing edge devices (drones or stationary CCTV). We map the specific drainage grid topology into our Spatio-Temporal Graph database.
Shadow Deployment
The Drain Web-GIS dashboard runs in a shadow capacity. The model processes live visual feeds, classifies sub-classes (e.g., Blocked-silt vs. Blocked-plastic), and generates silent alerts, allowing engineers to compare AI detection rates against manual field reports without interrupting standard dispatch protocols.
Forensic Auditing & ST-GNN Tuning
City engineers and environmental analysts review the dashboard outputs, validating the predictive flood propagation paths generated by the Spatio-Temporal GNN.
Full Deployment & Scale Recommendation
Synthesis of validation data into a formal impact report. The MLOps pipeline transitions to active prediction, empowering the platform to deliver real-time, color-coded alerts directly to emergency response teams.
Product Roadmap
Disciplined Municipal Scaling
Visual Edge Detection
- YOLOv8-Nano MVP operational
- Docker containerization
- Hugging Face Spaces sync
Sensor API Integration
- Rainfall intensity APIs
- Hydrodynamic flow data
- Real-time flood prediction
Graph Transformers
- Spatio-Temporal Attention Networks
- Long-range city-wide reasoning
- Massive drainage network scaling
Understanding System Limitations
In civic infrastructure AI, overpromising is a severe operational risk. We maintain total transparency regarding our model's current constraints. Due to the severe scarcity of annotated urban drainage datasets, our model relies on Generative Adversarial Networks (GANs) and Diffusion Models to synthesize extreme-weather blockage scenarios. Furthermore, our current deployed prototype focuses heavily on visual prediction; the Spatio-Temporal Graph Neural Network (ST-GNN) is still being fine-tuned to seamlessly ingest live IoT sensor flow data and rainfall intensity APIs.
Computational Bottlenecks
Initial model training severely constrained by limited cloud hardware (Kaggle 30-hour GPU expirations)
Environmental Data Scarcity
Extreme weather conditions obscured blockages; delays in acquiring geospatial datasets from municipal authorities
Immediate Next Steps
- Secure dedicated cloud infrastructure (Crane Cloud)
- Expand dataset across dry and wet seasons
- Integrate live rainfall and flow APIs
Ready to Pilot Drain in Your City?
Join our 12-week municipal pilot program and experience proactive urban flood management.