Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comparison of Lightweight Deep Learning Models for Particulate-Matter Nowcasting in the Indian Subcontinent & Surrounding Regions

Ansh Kushwaha, Kaushik Gopalan

TL;DR

This work targets practical 6-hour nowcasting of PM$_1$, PM$_{2.5}$, and PM$_{10}$ over the Indian subcontinent using CAMS Global Analysis fields. It introduces three lightweight architectures—ConvGRU, ConvLSTM, and U-Net—that operate on 256×256 input patches to predict a central 128×128 India-centric domain, with separate models for each PM species. Across PM types, these models consistently surpass the Aurora foundation model in RMSE, SSIM, and bias, while using three orders of magnitude fewer parameters, and exhibit robust seasonal and spatial performance. The findings support the viability of regionally optimized, resource-efficient deep learning approaches for operational air-quality nowcasting, with clear avenues for extending inputs, resolution, and architectures in future work.

Abstract

This paper is a submission for the Weather4Cast~2025 complementary Pollution Task and presents an efficient framework for 6-hour lead-time nowcasting of PM$_1$, PM$_{2.5}$, and PM$_{10}$ across the Indian subcontinent and surrounding regions. The proposed approach leverages analysis fields from the Copernicus Atmosphere Monitoring Service (CAMS) Global Atmospheric Composition Forecasts at 0.4 degree resolution. A 256x256 spatial region, covering 28.4S-73.6N and 32E-134.0E, is used as the model input, while predictions are generated for the central 128x128 area spanning 2.8S-48N and 57.6E-108.4E, ensuring an India-centric forecast domain with sufficient synoptic-scale context. Models are trained on CAMS analyses from 2021-2023 using a shuffled 90/10 split and independently evaluated on 2024 data. Three lightweight parameter-specific architectures are developed to improve accuracy, minimize systematic bias, and enable rapid inference. Evaluation using RMSE, MAE, Bias, and SSIM demonstrates substantial performance gains over the Aurora foundation model, underscoring the effectiveness of compact & specialized deep learning models for short-range forecasts on limited spatial domains.

A Comparison of Lightweight Deep Learning Models for Particulate-Matter Nowcasting in the Indian Subcontinent & Surrounding Regions

TL;DR

This work targets practical 6-hour nowcasting of PM, PM, and PM over the Indian subcontinent using CAMS Global Analysis fields. It introduces three lightweight architectures—ConvGRU, ConvLSTM, and U-Net—that operate on 256×256 input patches to predict a central 128×128 India-centric domain, with separate models for each PM species. Across PM types, these models consistently surpass the Aurora foundation model in RMSE, SSIM, and bias, while using three orders of magnitude fewer parameters, and exhibit robust seasonal and spatial performance. The findings support the viability of regionally optimized, resource-efficient deep learning approaches for operational air-quality nowcasting, with clear avenues for extending inputs, resolution, and architectures in future work.

Abstract

This paper is a submission for the Weather4Cast~2025 complementary Pollution Task and presents an efficient framework for 6-hour lead-time nowcasting of PM, PM, and PM across the Indian subcontinent and surrounding regions. The proposed approach leverages analysis fields from the Copernicus Atmosphere Monitoring Service (CAMS) Global Atmospheric Composition Forecasts at 0.4 degree resolution. A 256x256 spatial region, covering 28.4S-73.6N and 32E-134.0E, is used as the model input, while predictions are generated for the central 128x128 area spanning 2.8S-48N and 57.6E-108.4E, ensuring an India-centric forecast domain with sufficient synoptic-scale context. Models are trained on CAMS analyses from 2021-2023 using a shuffled 90/10 split and independently evaluated on 2024 data. Three lightweight parameter-specific architectures are developed to improve accuracy, minimize systematic bias, and enable rapid inference. Evaluation using RMSE, MAE, Bias, and SSIM demonstrates substantial performance gains over the Aurora foundation model, underscoring the effectiveness of compact & specialized deep learning models for short-range forecasts on limited spatial domains.

Paper Structure

This paper contains 9 sections, 2 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Results
  4. PM$_{1}$ Forecasting Performance
  5. PM$_{2.5}$ Forecasting Performance
  6. PM$_{10}$ Forecasting Performance
  7. Seasonal Performance Variability
  8. Analysis of Spatial Bias Patterns
  9. Conclusion

Figures (2)

  • Figure 1: Monthly averages of normalized RMSE (NRMSE = RMSE/mean) and SSIM for PM$_{1}$, PM$_{2.5}$, and PM$_{10}$ across all models. Each row shows a PM type with NRMSE on the left and SSIM on the right.
  • Figure 2: Spatial mean bias (Prediction $-$ Target) for PM$_{1}$, PM$_{2.5}$, and PM$_{10}$, averaged over the 2024 test period. Positive (red) values indicate overestimation and negative (blue) values indicate underestimation.