Pollution Intelligence
Air quality varies dramatically with weather, traffic, and time of day. Deterministic forecasts give citizens and regulators no sense of prediction confidence, leading to poorly informed health advisories. Our spatiotemporal model combines weather dynamics, traffic parameters, and satellite observations into localized prediction intervals, giving environmental agencies reliable uncertainty bounds for public health decisions.
Station-Level Forecast Explorer
The explorer enables inspection across a diverse range of pollution patterns, seasonal conditions, and temporal dynamics. Click on a station in the list to fetch its forecast window.
Select a city from the list above or click on a hub marker on the map to explore active stations.
Model Validation Metrics
Summary metrics compiled from out-of-sample evaluations over a 12-month period across all active cities.
Average prediction error across all stations
Model explains 79% of PM2.5 variance
~3 of 4 actuals fall within the 10–90% band
Average uncertainty band width
Across 7 Indian cities
Quantile forecast accuracy at Q0.5
Root mean squared error
Symmetric mean absolute percentage error
Spatiotemporal Error Patterns
Visualizing how model accuracy changes across forecast horizons, diurnal cycles, and different metropolitan centers.
Spatiotemporal Attention Network
Our network uses a message-passing graph architecture over spatial grids, combined with temporal causal convolutions. This constraints prediction limits to conservation matrices.
Spatial-Temporal Node Aggregation
Fuses satellite AOD column densities with continuous traffic profiles and localized wind vectors, aligning them onto a uniform graph mesh.
Graph Neural Network Layer
Passes messages along adjacency matrices computed from topography and road geometry to capture spatial dispersion patterns.
Quantile Output Decoder
Predicts multi-horizon distributions (quantiles 0.1, 0.5, 0.9) using a Pinball-loss objective function rather than simple mean squares.
Conservation Constraint
Forces spatial diffusion values to respect atmospheric mass boundaries, preventing the model from hallucinating pollutant concentration peaks.
7 Cities, 7 Stations
Trained on 72-hour context windows to predict 24-hour windows across 7 Indian metro cities. The model uses a spatiotemporal attention network with AdamW optimiser, cosine decay, and early stopping.
Dataset Composition
7 Indian metro cities with 21 active monitoring stations. 24,393 test windows for PM2.5 prediction. 72-hour sliding context windows with hourly resolution.
Training Configuration
AdamW optimiser with cosine annealing, graph neural network with message passing over spatial adjacency, trained with pinball loss across 3 quantiles (P10, P50, P90).
Context & Prediction Windows
72 hours of historical sensor data to forecast 24 hours ahead. Multi-variate inputs include PM2.5, traffic profiles, and temporal features.
Performance Highlights
MAE of 18.4 ug/m3 across all stations with R2 of 0.882. PICP of 86.7% against a 90% target interval across diverse pollution patterns.
Incorporate predictions
API access for pilot partners is planned as the platform matures.
curl -X POST https://api.purnabodha.ai/v1/predict/pollution \
-H "Authorization: Bearer $API_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"station_id": "bengaluru-city-center",
"traffic_multiplier": 1.0,
"wind_speed_mps": 6.0,
"quantiles": [0.10, 0.50, 0.90]
}'Where We're Taking This Next
Every deployed system raises new questions. These research programs explore the next generation of capabilities for pollution intelligence.
Satellite-Assisted Pollution Forecasting
Fusing columnar NO₂ and aerosol optical depth from Sentinel-5P and MODIS with ground-level monitoring networks for wider spatial coverage.
Large-Scale Representation Learning for Environmental Monitoring
Pretrained on global air quality data, adaptable to local monitoring networks with sparse station coverage and varied sensor configurations.
Spatiotemporal Representation Learning
Learning latent transport dynamics from sparse monitoring stations to reconstruct continuous pollution fields across urban areas.
Extreme Pollution Event Forecasting
Predicting pollution spikes from wildfires, dust storms, and temperature inversions with lead times sufficient for public health advisories.
Interested in piloting this at your facility or city?
We're seeking industrial facilities, municipalities, environmental agencies, and compliance partners interested in evaluating local air quality prototypes.