CO₂ Reconstruction Explorer · ISRO NRSC

Pavan AI 2.0
CO₂ Reconstruction Explorer · ISRO NRSC

OCO‑2 · OCO-3 · ERA5 · HGB v2

Query Filters

Date

Satellite

Search radius (km)

100 km

Day Explorer

Select date

Show as CO₂ heatmap

Bulk Reconstruction

Select a point on the map + date range → runs 24h model on every satellite obs within the radius → download one CSV with all days reconstructed.

From

To

Spatial radius (km) 50 km

Matching obs —

Date range —

Satellites —

Output rows —

Generating reconstruction…

Waiting…

Click on the map to begin

Select any point over India to query the nearest satellite observation and run the 24h CO₂ reconstruction.

Click anywhere on India to select a point

Charts appear here after querying a point

◎ Choose a mode, then select a region

🗺️

Select a region to begin

Choose Point (click map + radius), City, or State — then set a date range and click Analyse.

XCO₂ Time-Series

Each dot = one obs · coloured by satellite · line = 7-day rolling mean

Mean Diurnal Profile

Averaged across all days · BLH (blue, left) · CO₂ (red, right) · shading = ±1 std

BLH

CO₂

BLH Time-Series (Anchor Hour)

Daily BLH at 13:00 UTC — shows seasonal convective variation

Daily CO₂ Spread

Min / Mean / Max reconstructed CO₂ per day — diurnal amplitude over time

BLH vs Satellite CO₂

Each obs · colour = date · anti-phase confirms mixing dynamics

CO₂ Hour × Date Heatmap

X = hour of day · Y = date · colour = reconstructed CO₂ (ppm) · diurnal pattern evolution over time

lo

hi ppm

Monthly Mean XCO₂

Average of all obs per calendar month

Diurnal Amplitude Over Time

Daily (max − min) reconstructed CO₂ range — proxy for mixing strength

All Observations in Region

Sorted by date

Date	Sat	Lat	Lon	Dist km	Anchor CO₂	Unc	Min	Max	Mean	Amplitude

🧠 Model Architecture & Inference Pipeline

HistGradientBoosting · 24 per-hour models · ERA5 BLH + OCO-2/3 satellite anchor · physics-informed smoothing

Data Flow Pipeline

🛰️

OCO-2/3

Satellite XCO₂
anchor @ hr 13

→

🌤️

ERA5 NetCDF

24h boundary
layer heights

→

⚙️

Feature Eng.

11 features
per hour

→

📐

StandardScaler

Per-hour
normalisation

→

🤖

HGB model_h

24 models
→ Δ CO₂

→

〰️

Smoothing

Gaussian blur
+ BLH physics

→

📈

Output

24h CO₂
profile (ppm)

Input Features (11)

#	Feature	Type	Description
1	`blh`	Meteo	ERA5 BLH at target hour (m)
2	`blh_ratio`	Meteo	blh / daily mean BLH
3	`log_blh`	Meteo	log(blh + 1) — log-scale mixing
4	`sin_hour`	Time	sin(2π·h/24) — cyclic encoding
5	`cos_hour`	Time	cos(2π·h/24) — cyclic encoding
6	`co2_13`	Anchor	Satellite XCO₂ at hr 13 (ppm)
7	`blh_13`	Anchor	ERA5 BLH at hr 13 (m)
8	`sin_doy`	Time	sin(2π·doy/365) — seasonal
9	`cos_doy`	Time	cos(2π·doy/365) — seasonal
10	`season`	Calendar	0=Winter 1=Spring 2=Summer 3=Autumn
11	`is_weekend`	Calendar	0/1 — weekend emission flag

Model Details

🤖 Algorithm

HistGradientBoostingRegressor (sklearn) — histogram-based gradient boosting, fast on large datasets, native missing-value support.

Count24 models (one / hour)

🎯 Target Variable

Predicts Δ CO₂ = CO₂_h − CO₂_13. Final output = co2_13 + Δ. Anchoring to satellite prevents accumulative drift.

Unitsppm (deviation)

📐 Scaler

Per-hour StandardScaler fitted on named pd.DataFrame. Must receive column-named input — avoids sklearn UserWarning on wrong feature order.

TypeX-only (not target)

〰️ Smoothing

Gaussian blur σ=1.8h (circular wrap) + 35% BLH physics blend (CO₂ ∝ −BLH). Hard re-anchor at hr 13.

Alpha0.35 physics / 0.65 model

🔐 No Leakage

Temporal split by date (not row). Optuna tunes on validation only. Test set touched once. Zero future CO₂ leakage — only co2_13 anchor used.

Split70 / 15 / 15

⚡ Inference Speed

24 model.predict() calls with pre-loaded pkl files. BLH NetCDF read is the bottleneck (~0.3s). Total query response typically <1s.

BottleneckERA5 NetCDF I/O

Active hour models

Live Inference Simulation Watch the pipeline execute step-by-step for a sample query

// Click ▶ Run to simulate the full inference pipeline

0 / 24 hours