Ghosh et al. (2025) Novel R-CNN and transformer models for pollution impacts and land cover changes around iconic heritage sites in developing countries: a case study

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-12-03
• Authors: S. Ghosh, B. Ashok, L. Agilandeeswari, Prabukumar Manoharan, Ariful Rahaman, A. K. Mathur, Abhinav Sudhakar Dubey, Eishani Purohit, Visakh Gangadharan
• DOI: 10.1038/s41598-025-27155-x

Research Groups

• VIT Vellore, Vellore, India
• Leeds University, Leeds, UK

Short Summary

This study quantifies pollution impacts and land cover changes around the Meenakshi Amman Temple in Madurai, India, using novel R-CNN and Transformer models to characterize particulate matter and predict future environmental indicators. It demonstrates the significant benefits of transitioning to electric vehicles and proposes an AI-mediated method for pollution assessment and remediation.

Objective

• To quantify the impact of vehicular pollution (soot, black carbon) and Land Use Land Cover (LULC) changes on the microclimate and heritage structures around the Meenakshi Amman Temple complex.
• To develop and apply novel Mask Recurrent-Convolutional Neural Networks (R-CNN) for rapid characterization of pollution particle Sauter Mean Radii (SMR) from visual greyscale images, circumventing expensive in-situ measurements.
• To utilize a Stacked Transformer model for long-term prediction (1940–2050) of key environmental indicators (Leaf Area Index, Temperature, Precipitation, Cloud Cover, Soil Moisture) under high emissions (RCP 8.5) and electric vehicle transition scenarios.
• To provide actionable insights and quick protocols for government officials and stakeholders regarding environmental impact analyses, cost savings, and smart forecasts for heritage conservation and sustainable urban planning.

Study Configuration

• Spatial Scale: Meenakshi Amman Temple complex in Madurai, India. The Weather Research and Forecasting (WRF) model domain covered 9.86° N and 78.10° E, with a horizontal grid spacing of 3 km across a 100 × 100 grid.
• Temporal Scale: Analysis and predictions span past (1940), extant (2021), and future (2020–2050, with some predictions up to 2060) time periods, contrasting pre-economic liberalization with post-liberalization and high-emission scenarios.

Methodology and Data

• Models used:
  • Weather Research and Forecasting (WRF) model (Advanced Research WRF (ARW) core, version 4.6.1)
  • Mask Recurrent-Convolutional Neural Networks (R-CNN) for particle SMR estimation.
  • Stacked Transformer model for multivariate time series forecasting of environmental parameters.
  • ENVI-met microclimatic model (version 5.7.1) for PM10 deposition assessment.
• Data sources:
  • NCEP FNL Operational Model Global Tropospheric Analyses (for 2021).
  • MESACLIP (MESOSCALE Atmosphere-Ocean Interactions in Seasonal-to-Decadal Climate Prediction) dataset (10-member ensemble of CESM high-resolution RCP 8.5 simulations for 2050).
  • ECMWF (European Centre for Medium Range Weather Forecasts) ERA5 dataset (for 1940).
  • Copernicus Climate Change Service (cds.climate.copernicus.eu), land.copernicus.eu, and open-meteo.com/en/docs/historical-weather-api for Transformer model input data (1940–2024).
  • In-situ experimental data from a four-stroke single-cylinder diesel engine (5 HP) for soot and black carbon collection using Grade-1 micron filter paper (11 µm pore size).
  • Field Emission Scanning Electron Microscope (FE-SEM, Thermo FEI QUANTA 250 FEG, 1.2 nm resolution at 30 kV) and Energy Dispersive X-ray (EDX) for morphological characterization of soot particles.

Main Results

• Diesel engine emissions around the Meenakshi Amman Temple complex release soot and black carbon particles in the accumulation mode (0.2–0.5 µm) that remain airborne for over 7 days, causing soiling of the temple's Gopurams. Natural rain scavenging efficiency for these particles is approximately 20%.
• Over the period 1940–2050, the region experiences a progressive loss of vegetative cover (Leaf Area Index, LAI), decreased precipitation, reduced soil moisture, and an increase in non-precipitating clouds, exacerbated under the RCP 8.5 high emissions scenario.
• A novel AI-mediated method using Mask R-CNN accurately estimates the Sauter Mean Radii (SMR) of emitted particles from visual greyscale images, showing a progressive decrease in SMR and an increase in sub-micron particles with increasing engine loads.
• Transformer model predictions for 2020–2050 (RCP 8.5) indicate a continued decrease in LAI and soil moisture, coupled with an aggravated increase in heat stress.
• A phased transition to electric vehicles (EVs) significantly improves environmental indicators, leading to a 0.6 K temperature decrease, an overall increase in LAI, a 42% reduction in ambient PM10 concentrations, and a 34% decrease in cumulative PM10 deposition on the temple's west gopuram during July.
• The Transformer model provides robust and accurate 1D and 2D forecasts for LAI, Temperature, Cloud Cover, Precipitation, and Soil Moisture, demonstrating its utility for urban planning and decision-making.

Contributions

• Introduces a novel, AI-mediated protocol using R-CNN and Transformer models for rapid, cost-effective characterization of pollution particle size distributions (SMR) and long-term environmental forecasting around heritage sites, eliminating the need for expensive in-situ scanning mobility particle sizer spectrometers (SMPS).
• Provides a comprehensive, multi-decadal assessment (1940–2050) of the combined impacts of urbanization, vehicular pollution, and climate change (RCP 8.5) on the microclimate and structural integrity of an iconic heritage site in a developing country.
• Quantifies the significant environmental and economic benefits (potential savings of millions of USD over 5 years for facade restoration) of transitioning from diesel to electric vehicles, offering strong evidence for policy interventions.
• Develops a practical framework for urban planners and stakeholders to utilize AI-driven forecasts for environmental impact analyses and to inform remediation strategies, such as the design of optimized roadside sprinklers for wet scavenging.

Funding

Open access funding provided by Vellore Institute of Technology. There was no specific project funding for this research.

Citation

@article{Ghosh2025Novel,
  author = {Ghosh, S. and Ashok, B. and Agilandeeswari, L. and Manoharan, Prabukumar and Rahaman, Ariful and Mathur, A. K. and Dubey, Abhinav Sudhakar and Purohit, Eishani and Gangadharan, Visakh},
  title = {Novel R-CNN and transformer models for pollution impacts and land cover changes around iconic heritage sites in developing countries: a case study},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-27155-x},
  url = {https://doi.org/10.1038/s41598-025-27155-x}
}