Forest fire and smoke detection using deep learning-based learning without forgetting
A sixth-semester college project became a dataset, a Fire Ecology paper, and eventually a public classifier demo. The research question: can a fire-detection model learn new data without forgetting the old data?
Venue
Fire Ecology 19(1), Springer
Year
2023
Role
Co-author · undergrad research
DOI
10.1186/s42408-022-00165-0
01 · Origin
IndexChapter 01
Origin
The first version began in 2020 as a sixth-semester college project: collect forest imagery, train a Python model, and see whether computer vision could flag fire, smoke, both, or neither. The public v2 demo is a cleaner rewrite of that idea, but this page focuses on the peer-reviewed research layer that came out of it.
Chapter 02
The research problem
Forests are slow to grow and fast to burn, so early detection matters. Deep learning can spot fire and smoke in imagery — but a model fine-tuned on a new fire dataset tends to lose its accuracy on the dataset it originally learned. That failure mode is called catastrophic forgetting, and it makes detectors brittle exactly where you want them robust: on conditions they have seen before but not recently.
Chapter 03
Method
In plain terms: start from models that already know how to see, teach them fire, and stop them from forgetting what they learned first. Step through the pipeline below.
Stage 01 — Imagery
Forest imagery collected and sorted into four classes — fire, nofire, smoke, and smokefire — the corpus published separately as the Forest Fire C4 dataset.
Illustrative pipeline — see the paper for exact architecture and hyperparameters.
Chapter 04
Results, as published
These numbers are taken from the paper’s abstract and kept in the same form for traceability.
| Condition | Dataset | Accuracy |
|---|---|---|
| Best model overall | Original dataset | 98.72% |
| Without LwF | New task (BowFire) | 79.23% |
| With LwF | New task (BowFire) | 91.41% |
| With LwF | Original dataset | 96.89% |
The takeaway: fine-tuning with LwF performed comparably well on the original dataset while still successfully categorising novel, unseen data. Of the three architectures tested, Xception excelled.
Chapter 05
Scope and authorship
Cite this
Sathishkumar, V.E., Cho, J., Subramanian, M., Naren, O.S. (2023). Forest fire and smoke detection using deep learning-based learning without forgetting. Fire Ecology, 19(1).
doi.org/10.1186/s42408-022-00165-0 ↗Lessons
What it taught me
- A model that forgets is a liability, not a curiosity — continual learning is an engineering requirement, not an academic footnote.
- The dataset is half the paper. Building the four-class corpus forced precision about what “detection” even means.
- The public demo matters because people believe what they can test — but honest scope framing matters more.
- Writing research taught me to state limitations plainly. That habit followed me into product work.