SAIP2026

Detection of Wildfire Smoke Plumes using Machine Learning Systems

7 Jul 2026, 17:20

1h 20m

Great Hall (University of the Western Cape)

Poster Presentation Track F - Applied Physics Poster Session 1

Speaker

Mr Mbongeni Mncube (University of Kwazulu-Natal)

Description

Triggered false alarms are one of the biggest problems for smoke detection systems. Cloud movement in the sky, cloud reflections on hills, cars, animals eating, and trees moving about in a forest can easily trigger a false alarm. This poses a critical issue because forest wildfires are constantly on the rise. This affects the global economy substantially and forest ecosystem, therefore reliable detection systems are on demand. This study investigates different preprocessing techniques in conjunction with deep learning for image classification and object detection, to create an AI system with an accuracy of 90% or greater for identifying smoke plumes in the forest.

The image classifier was built around a cumulative difference image derived from four sequential camera frames. Three pairwise differences were computed and summed into a single representation capturing overall scene motion. This image was evaluated in three forms as classifier input: grayscale, optical flow, and colour. Among the representations explored, the colour-based input demonstrated promising performance because its accuracy was about 0.83 with loss of about 0.62, Grayscale results were 0.78 accuracy and 0.79 loss. the optical flow performed comparatively weaker.

There's notable divergence between training loss and validation loss, to rule out whether they were from the model built or data used to build the model, k-cross fold was undertaken with a ratio of 80% training data and 20% evaluation data on all 5 folds, this was done then compared with the results from test data, the outcome proved that the model was consistent across folds but not as much during testing due to the fact that the results from testing were vastly different from k cross folds. This translates to the quality of the data not being good enough or is a representation of the dataset.

on the object detection side images were bounded to focus the learning on the region of interest, optical flow was employed as means to analyse smoke movement, with optical flow, colour contrast changes and heatmap vision this was done to provide the classifier with richer and more discriminative visual information.
while results do show promising potential, expanding data, to overcrowd the camera shaking error trees cars and animals. would prove to be great.