Wildfire Smoke Detection using CNN

Wildfires grow from a wisp of smoke to an uncontrollable front in minutes, and early detection is the difference between a contained incident and a disaster. Forest camera networks stream thousands of frames a day, but nobody can watch them all — an automated first pass that flags suspicious smoke plumes for a human reviewer is exactly what these networks need. Building that first pass is genuinely hard: smoke is amorphous and changes shape constantly, and clouds, fog, dust and glare are classic false alarms that look deceptively similar from a distance. This project frames the task as binary frame classification: a transfer-learned ResNet or EfficientNet scores each sampled frame for smoke, and an alert candidate is raised only when confident detections persist over a time window — single-frame noise never triggers an alert. Grad-CAM overlays show which image regions drove each alert so the reviewer judges reasoning, not just a number, and the evaluation includes an explicit false-alarm analysis over hard negatives.

Dataset & model:
Dataset name: Public wildfire smoke imagery from forest camera networks such as HPWREN (High Performance Wireless Research and Education Network, UC San Diego).
Source: Publicly released camera imagery from the HPWREN network, used in fire-detection research.
Task: Binary image classification (smoke vs no-smoke) per sampled frame, with a time-window alert rule.
Classes: Smoke / no-smoke — the binary labeling scheme defined by the build over the public imagery (there is no single official labeled smoke-detection benchmark used here).
Model: Transfer-learned CNN — pre-trained ResNet or EfficientNet backbone fine-tuned on the smoke imagery.
Input: Forest camera frames (JPEG/PNG), sampled and normalized.
Prediction: Per-frame smoke probability.
Output: Alert candidates on sustained detections, with Grad-CAM overlays explaining each alert.
Evaluation metrics: Precision, recall, F1 on the validation split, plus a dedicated false-alarm analysis over clouds, fog and dust — computed by the notebook during the buyer's build. No metrics claimed.

Working:

1. Frames are sampled from forest camera feeds and normalized with OpenCV.
2. Training phase: a pre-trained ResNet or EfficientNet backbone is fine-tuned on public smoke imagery for binary smoke / no-smoke classification.
3. Inference phase: the fine-tuned classifier scores each sampled frame and outputs a smoke probability.
4. Frames staying above the threshold across a time window become an alert candidate, suppressing single-frame noise.
5. Grad-CAM overlays visualize the attended regions so a human reviewer can judge whether the model saw a real plume or was fooled by a cloud.
6. Evaluation phase: the notebook computes precision, recall and F1 on the validation split and runs the false-alarm analysis over hard negatives.

Specifications:
Model | Transfer-learned CNN (ResNet or EfficientNet backbone, ImageNet pre-trained, fine-tuned)
Data | Public HPWREN fire and smoke camera imagery
Task | Binary frame classification (smoke / no-smoke) with time-window alerting
Input | Forest camera frames (JPEG/PNG), sampled and normalized
Output | Alert candidates on sustained detections + Grad-CAM overlays
Alert rule | Configurable probability threshold + persistence across a time window
Evaluation | Precision, recall, F1 + false-alarm analysis — computed by the notebook during the buyer's build
Demo app | Flask app with camera-frame upload interface
Scope | Daytime forest camera frames from the evaluated imagery sources

[Transfer-Learned Smoke Classifier] (implemented) — ResNet or EfficientNet backbone, pre-trained on ImageNet and fine-tuned on public wildfire smoke imagery for binary classification.
[Per-Frame Smoke Probability] (implemented) — Each sampled frame gets a smoke probability score with a configurable alert threshold.
[Time-Window Alert Rule] (implemented) — Alert candidates trigger only on sustained high-confidence detections across a time window, not single frames.
[Grad-CAM Alert Explanations] (implemented) — Overlays show which image regions drove each alert for human review.
[False-Alarm Analysis] (implemented) — The notebook explicitly evaluates clouds, fog and dust as hard negatives alongside precision, recall and F1.
[Evaluation Notebook] (implemented) — Computes precision, recall and F1 on the validation split with the false-alarm study during the build.
[Camera-Feed Demo App] (implemented) — Flask demo app for uploading and scoring camera frames.
[Night Infrared Inputs] (future-scope) — Night-capable infrared inputs and video-sequence modeling are future work, not included features.

Complete source code (training, inference, alert logic, Grad-CAM, demo app)
Jupyter training and evaluation notebook (precision, recall, F1, false-alarm analysis)
Project report PDF (background, smoke imagery, methodology, evaluation, error analysis)
PPT presentation for final review
Viva Q&A preparation document (CNNs, transfer learning, Grad-CAM, false alarms, evaluation)
Setup guide (environment, dataset download, training, running on your own camera frames)

Clouds, fog, dust and glare are the classic false-alarm sources: the notebook includes a false-alarm analysis, but the system cannot reliably separate distant haze from smoke in all conditions — a human must review every alert.
Public HPWREN imagery is the only evaluated source; other terrains, camera types and night conditions fall outside the documented coverage.
Scope covers daytime forest camera frames from the evaluated imagery sources only.
This is an early-warning research prototype, not a certified fire-alarm system — it must never replace official fire-detection infrastructure or be presented as safety-grade equipment.
The smoke / no-smoke labeling scheme is the build's own; there is no single official labeled benchmark used here, so cross-source generalization is not claimed.