Select a lead day (0 = today, up to 6 days ahead).
Draw a polygon inside Madhya Pradesh and click "Run Forecast".
Map Tools
Identify: Click any point on a prediction/forecast layer to see the exact value.
Ground Stations: Show the 20 MP CPCB monitoring stations with observed PM2.5/PM10.
Methodology
Scope
This platform analyses PM2.5 and PM10 concentrations exclusively within the Madhya Pradesh state boundary (FAO/GAUL administrative boundary). Any user-drawn area extending outside the state is automatically clipped.
Data Sources
Sentinel-2 L1C (TOA): 10m multispectral imagery, cloud-masked with CloudScore+ (≥0.72)
NASA FIRMS / VIIRS: Active fire radiative power (biomass burning)
SRTM, ESA WorldCover, VIIRS DNB, WorldPop: Static terrain/land features
MP CPCB ground network: Same-day inverse-distance-weighted PM2.5/PM10 from the 20 MP monitoring stations, blended by distance to the queried location (network assimilation)
Open-Meteo (CAMS): 7-day PM2.5/PM10 forecast, downscaled with GFS meteorology
Data Preprocessing
Spatial QC (MP-only stations), physical-range QC, station-topography join, target winsorization at the 99.5th percentile, cyclical + Indian monsoon-calendar seasonal-regime temporal encoding (winter inversion, summer dust, monsoon wet deposition, post-monsoon burning), and training-set-median imputation for structurally sparse bands (AOD, SAR, AER_AI, NO2, CO).
Models
For each of PM2.5 and PM10 independently: Ridge, Lasso, ElasticNet, SVM, Random Forest, Extra Trees, HistGradientBoosting, CatBoost, a stacking ensemble, and a non-negative-least-squares-weighted hybrid ensemble are all hyperparameter-tuned (RandomizedSearchCV with spatial GroupKFold) and compared via honest 5-fold leave-station-group-out cross-validation — every one of the 20 MP CPCB stations is held out exactly once and results are pooled, rather than a single fixed train/test split. The best-performing model is selected using a gap-aware rule that prefers a smaller train–validation gap over a marginally higher but overfit score. PM10 additionally uses a leak-free cascade feature: the PM2.5 model's out-of-fold satellite-only estimate is fed in as an input, mirroring the real physical coupling between the two pollutants.
Current Validation Accuracy
Measured by 5-fold leave-station-group-out cross-validation (every MP station held out once):
PM2.5: R² = 0.53, RMSE ≈ 24 µg/m³, near-zero bias
PM10: R² = 0.38, RMSE ≈ 48 µg/m³, near-zero bias
PM10 accuracy is lower than PM2.5 because coarse dust (its dominant fraction) is driven by very local, short-timescale sources — unpaved roads, construction, wind resuspension — that satellite composites and a 20-station ground network only partially resolve. Per-station diagnosis also found that a small number of stations are genuine regional outliers (actual readings differing sharply from every nearby station and from what the surrounding network would predict, with no distinguishing satellite/meteorology signature) — a station-level data characteristic that additional regional features cannot correct.
Limitations
Cloudy pixels are masked; the system searches up to ±16 days for the nearest cloud-free composite.
Forecasts beyond day 0 rely on CAMS/GFS meteorology rather than observed satellite imagery.
Accuracy depends on the density of the 20-station MP CPCB network, both for satellite-model training and for the live network-assimilation feature.
PM10 estimation accuracy (R² = 0.38) is below the PM2.5 model and should be treated as indicative rather than precise, particularly away from monitoring stations and near stations identified as regional outliers.