Wildfire Smoke Detection.

Convolutional neural network model based on the architecture of the Faster-RCNN for wildfire smoke detection. For this project we used a pretrained model on ImageNet dataset, from detectron2's Model Zoo, and fine-tuned it for the task of wildfire smoke detection from optical image data.

Dataset Properties.

This dataset is released by AI for Mankind in collaboration with HPWREN (High Performance Wireless Research and Education Network), and is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License.

AI For Mankind is a 501(c)(3) nonprofit organization with the mission of mobilizing the tech community to work on world challenging problems using AI and Data. We organize tech talks, workshops, and hackathons. We want to build a grassroot community of volunteers creating solutions using AI and Data to bring positive impacts to society at large. (https://aiformankind.org/)

The High Performance Wireless Research and Education Network (HPWREN), a University of California San Diego partnership project led by the San Diego Supercomputer Center and the Scripps Institution of Oceanography's Institute of Geophysics and Planetary Physics, supports Internet-data applications in the research, education, and public safety realms. (http://hpwren.ucsd.edu/)

The above dataset is available in 2 different versions in Pascal VOC annotation format:

Bounding Box Annotated Wildfire Smoke Dataset Version 1.0 with 744 annotated images.

↪️ BBAWS Dataset v1.0 - Pascal VOC

Bounding Box Annotated Wildfire Smoke Dataset Version 2.0 with 2192 annotated images.

↪️ BBAWS Dataset v2.0 - Pascal VOC

The first version of the dataset is also available in COCO annotation format by Roboflow:

↪️ BBAWS Dataset v1.0 - COCO

For this project we used the dataset in COCO annotation format provided by Roboflow.

Model Architecture

The model architecture is based on the general architecture of the Faster-RCNN, which includes the main modules of Feature Pyramid Network, Region Proposal Network as well as the model of Fast-RCNN. For the bottom-up pathway of the FPN network the architecture of the ResNet50 was used.

Image source: https://miro.medium.com/max/2000/1*Wvn0WG4XZ0w9Ed2fFYPrXw.jpeg

Model Performance

Metric	IoU	Area	maxDets*	Score
Average Precision (AP)	0.50:0.95	all	100	0.551
Average Precision (AP)	0.50	all	100	0.921
Average Precision (AP)	0.75	all	100	0.582
Average Precision (AP)	0.50:0.95	small	100	0.333
Average Precision (AP)	0.50:0.95	medium	100	0.495
Average Precision (AP)	0.50:0.95	large	100	0.660
Average Recall (AR)	0.50:0.95	all	1	0.604
Average Recall (AR)	0.50:0.95	all	10	0.608
Average Recall (AR)	0.50:0.95	all	100	0.608
Average Recall (AR)	0.50:0.95	small	100	0.429
Average Recall (AR)	0.50:0.95	medium	100	0.568
Average Recall (AR)	0.50:0.95	large	100	0.700

*Maximum number of detections per image.

Results

Results_video_1.mp4

Original video at: https://www.youtube.com/watch?v=q07TBd5o1HQ&t=35s

Results_video_2.mp4

Original video at: https://www.youtube.com/watch?v=5cEr5ZXGUYA

Download full results HERE (≈589 MB)

Requirements

torch == 1.9.0+cu102                           numpy == 1.19.5                           json == 2.0.9
torchvision == 0.10.0+cu102                    yaml == 5.1                               fiftyone == 0.12.0
pyyaml == 5.1                                  pandas == 1.3.2                           IPython == 5.5.0
detectron2 == 0.5                              cv2 == 4.1.2