Aerosol Classification based on Satellite Data
A Random Forest-based pipeline for classifying aerosol types across Asia using Sentinel-5P TROPOMI and MODIS satellite data, trained on AERONET ground-truth labels.
Aerosol type (dust, smoke, sulfate, black carbon, etc.) strongly affects radiative forcing, climate models, and air quality estimates, but satellite-based classification is hard in cloud-prone, data-sparse regions like South and Southeast Asia. This project builds on Choi et al. (2019/2021) by reducing the input feature set to variables sourced only from TROPOMI (plus static MODIS land-cover layers), cutting dependency on cloud-sensitive MODIS aerosol products and nearly doubling usable training data, at a small cost in raw 7-class accuracy.
- Choi et al. used 11 features (TROPOMI + MODIS AOD/Ångström/TOA reflectance) → 6,268 usable samples, 72% (7-class) / 82% (4-class) accuracy.
- This study uses 6 features (TROPOMI: AI, SZA, CO, NO₂ + MODIS land cover, % urban) → 9,980 usable samples (~1.6x more), 69% (7-class) / 81% (4-class) accuracy.
- Dropping MYD04-derived features (AOD, Ångström exponent, reflectance) removes the main source of missing data and improves spatial coverage substantially, at minimal accuracy cost, especially for the 4-class scheme (1% gap).
| Source | Variables | Resolution |
|---|---|---|
| Sentinel-5P TROPOMI L2 | Aerosol Index (AI), Solar Zenith Angle (SZA), CO column density, NO₂ column density | 7 km × 3.5 km |
| MODIS MYD04_L2 | AOD (550 nm), Ångström exponent (470–660 nm), Deep Blue TOA reflectance (412/470/660 nm) | 10 km |
| MODIS MCD12C1 | Land cover type (IGBP), % urban area (annual) | 0.05° |
| AERONET V3 L1.5 | Depolarization ratio, size distribution, single-scattering albedo (SSA) — ground truth labels | 87 sites across Asia, Jan–Feb 2021 |
Derived from Depolarization Ratio (Rd = δ/0.31 at 1020 nm) and SSA:
7-class scheme
| Type | Threshold |
|---|---|
| Pure Dust (PD) | Rd > 0.89 |
| Dust Dominated Mix (DDM) | 0.53 ≤ Rd ≤ 0.89 |
| Pollution Dominated Mix (PDM) | 0.17 ≤ Rd < 0.53 |
| Non-Absorbing (NA) | Rd < 0.17, SSA > 0.95 |
| Weakly-Absorbing (WA) | 0.90 < SSA ≤ 0.95 |
| Moderately-Absorbing (MA) | 0.85 < SSA ≤ 0.90 |
| Strongly-Absorbing (SA) | SSA < 0.85 |
4-class scheme: PD, DDM, NA (merges NA/WA/PDM), SA (merges MA/SA)
- Label generation — compute Rd from AERONET depolarization ratio, classify into 7/4 types.
- Spatiotemporal filtering — bounding box lat (-11.8, 45.4), lon (34.4, 145.9); ocean pixels removed (geopandas).
- Data merging:
- MCD12C1 (annual land cover, % urban) merged via KD-tree on nearest pixel.
- TROPOMI (AI, SZA, CO, NO₂) merged via Ball Tree within ±5 hours of AERONET measurement.
- MYD04 (AOD, Ångström, reflectance) merged similarly; ~4,000 unmapped rows dropped.
- Train/test split — 60/40, site-based (geographically disjoint sites for train vs. test).
- Model — Random Forest classifier; feature selection via mean-decrease impurity.
- Validation — accuracy, confusion matrix, SSA comparison (AERONET vs. predicted), fine-mode fraction consistency check.
- Full-area application — applied to Jan 1, 2021 satellite data over Asia at 0.1° grid resolution.
| Model | Classes | Accuracy |
|---|---|---|
| Choi et al. | 7 | 72% |
| Choi et al. | 4 | 82% |
| This study | 7 | 69% |
| This study | 4 | 81% |
- Both models reliably detect dust and strongly-absorbing pollution aerosols (SA, DDM, PD).
- Pollution-related subtypes (MA, WA, NA) are frequently confused with each other in the 7-class scheme; merging to 4 classes resolves most of this.
- This study achieves better NA classification than Choi et al. (77% vs. 74%).
- SSA spectral trends (PD/DDM increasing with wavelength, SA decreasing, NA flat) closely match AERONET, confirming physical validity of predictions.
- Full-area maps show broader, more spatially continuous coverage than Choi et al., particularly over Central and East Asia, capturing hotspots like the Thar Desert, Indo-Gangetic Plain, Chinese industrial regions, and Southeast Asian biomass-burning zones.
sentinelhub(v3.11.1) — TROPOMI data retrievalmodis-tools(v1.1.5) — MODIS data retrievalgeopandas— spatial filtering (ocean pixel removal)- KD-tree / Ball Tree — nearest-neighbor spatiotemporal merging
- Random Forest (scikit-learn assumed)
- Temporal expansion beyond Jan–Feb 2021 (multi-season, multi-year)
- Fusion with additional satellite instruments
- Exploration of neural network-based classifiers
Pradnyesh Keluskar*, Aaditya Madaye, Siddh Jain, Sarthi Kanade Computer Science and Engineering, Sardar Patel Institute of Technology, Mumbai, India
*Correspondence: pradnyesh.keluskar23@spit.ac.in
Builds on: Choi, M. et al. (2019/2021) — RF-based global aerosol classifier using AERONET, TROPOMI, and MODIS.