Dr Georgios Voulgaris

Dr Georgios Voulgaris

Postdoctoral Researcher in Computer Vision, Deep Learning, and Remote Sensing at the University of Oxford.

IGARSS 2025 — Oral Presentation: Detecting Cement Plants with Landsat-8

Date:

IGARSS 2025 — Oral Presentation (First Author)

I delivered a first author oral presentation at IGARSS 2025 on my work titled “Detecting Cement Plants with Landsat-8: A Physics-Informed, Multi-Temporal, and Multi-Spectral Deep Learning Fusion Approach”. The presentation summarised the development of FusionNet, a deep learning model that integrates thermal and short-wave infrared signatures for enhanced cement plant detection.

Paper: “Detecting Cement Plants with Landsat-8: A Physics-Informed, Multi-Temporal, and Multi-Spectral Deep Learning Fusion Approach”
IEEE IGARSS 2025
DOI: 10.1109/IGARSS55030.2025.11243713

The presentation covered:

  • Challenges of detecting cement plants using remote sensing.
  • Use of the Global Database of Cement Production Assets to build a comprehensive dataset.
  • Extraction of 1 km2 cement and landcover chips using multi-temporal Landsat-8 imagery.
  • Thermal Infrared (Bands 10 and 11) analysis for operational status detection.
  • Short Wave Infrared (Bands 6 and 7) analysis for soil moisture and mineralogical changes.
  • Introduction of the novel SWIR Band 7:6 ratio for enhanced discriminative power.
  • State-of-the-art performance with 90.6 percent accuracy using the Band 7:6 ratio.

FusionNet architecture:

  • A signal-processing-parameterised convolutional layer that improves feature extraction from complex spectral patterns.
  • Mix Pooling (combined average and max pooling) to capture both global and local spatial statistics.
  • Dilated convolutional layers to provide a wider receptive field while preserving fine detail.
  • Five unimodal backbones trained on Bands 11, 10, 7, 6, and the Band 7:6 ratio.
  • Channel attention mechanism to adaptively reweight spectral contributions.
  • CNN5 decoder for final cement/landcover classification.

Key findings:

  • SWIR Band 7:6 ratio provides superior discriminative features compared to thermal bands.
  • FusionNet consistently outperforms baseline models across all spectral combinations.
  • Soil moisture and organic composition changes are more informative than temperature alone.
  • Physics-informed multi-spectral fusion significantly improves cement plant detection accuracy.