India PM2.5 Prediction

User Guide

Getting Started

1. Select a date using the calendar or date input. Green dates indicate Sentinel-2 satellite availability.
2. Use the drawing tools on the map to define your area of interest (polygon or rectangle).
3. Select resolution: 60m gives Sentinel-2 native pixel-level detail; 500m or 1km are faster for larger areas.
4. Click "Predict PM2.5" to generate a wall-to-wall PM2.5 map covering the entire drawn region.
5. Click "Load Layers" to view environmental indicator layers at native resolution.

Map Tools

• Layers: Toggle the left panel showing calendar, controls, basemap selection, and indicator layers.
• Dashboard: Toggle the right panel showing prediction statistics and indicator interpretations.
• Identify: Click any point on the PM2.5 prediction to see the exact value and coordinates.
• Ground Stations: Fetch and display real-time CPCB station data from WAQI API.

Basemaps

Choose between OpenStreetMap, ESRI Satellite with labels, OpenTopoMap terrain, or Dark basemap using the selector in the left panel.

Indicator Layers

Each layer is displayed at its native resolution from Google Earth Engine. Toggle layers on/off and view their color legends. Use the opacity slider to adjust transparency.

Methodology

Data Sources

• Sentinel-2 L1C (TOA): 10m multispectral imagery, cloud-masked using Google CloudScore+ (threshold >= 0.72). Only pixels with high confidence clear-sky are used.
• Landsat 8/9 TOA: 30m backup, cloud-masked using QA_PIXEL band (cloud, cloud shadow, and cirrus bits). Only cloud-free pixels are retained.
• MODIS MCD19A2: 1km Aerosol Optical Depth (MAIAC algorithm)
• ERA5-Land: 11km meteorological variables (temperature, humidity, wind, pressure)
• ERA5: 31km boundary layer height
• SRTM: 30m elevation data
• VIIRS DNB: 500m nighttime lights as urbanization proxy
• WorldPop: 100m population density
• ESA WorldCover: 10m land use/land cover

Model

A stacking ensemble (HistGradientBoosting + ExtraTrees + Ridge meta-learner) trained on CPCB ground-truth PM2.5 measurements paired with satellite-derived features. The model predicts 24-hour average PM2.5 concentrations at each grid point (60m spatial resolution) within the selected region, providing wall-to-wall coverage matching Sentinel-2 native pixel size.

AQI Standard

PM2.5 categorization follows India's National Ambient Air Quality Standards (NAAQS) / CPCB breakpoints for 24-hour average concentrations.

Limitations

• Cloudy pixels are masked (CloudScore+ >= 0.72 for S2, QA_PIXEL for Landsat) — prediction gaps appear where clouds persist.
• On heavily overcast dates, the system searches up to +/-16 days for the nearest cloud-free composite.
• The model performs best for dates and regions similar to training data coverage.
• Accuracy degrades in areas with sparse ground-truth monitoring.