Seeing methane: notes on the 3D-CNN behind the ICoICT paper

Created on August 21, 2026

2026   ·   computer-vision   deep-learning   ·   research

Draft. This is a placeholder outline, not a finished piece. Full version coming soon.

Methane is invisible. In the right infrared band it is not quite invisible, and that gap is the whole problem.

Outline to fill in:

  • The setup. Thermal camera, gas released under varying conditions, video rather than stills. Why the temporal dimension matters and why a plain 2D CNN was not enough.
  • The preprocessing that did the heavy lifting. Gaussian Mixture Models, running average, adaptive background subtraction, and why classical methods still earned their place in front of the network.
  • Getting it small. Input-resolution scaling and quantization cut inference time by about 60 percent while accuracy stayed above 99 percent. Worth being honest about what that number does and does not mean.
  • What I would do differently. Mostly around the dataset.

The paper is on IEEE Xplore if you want the formal version.