metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype:
        class_label:
          names:
            '0': cloud
            '1': other
            '2': smoke
  splits:
    - name: train
      num_bytes: 85556006
      num_examples: 14318
    - name: validation
      num_bytes: 22137739
      num_examples: 3671
    - name: test
      num_bytes: 11026374
      num_examples: 1843
  download_size: 132474880
  dataset_size: 118720119
tags:
  - climate
task_categories:
  - image-classification
task_ids:
  - multi-label-image-classification
license: mit

COMPARING SIMPLE DEEP LEARNING MODELS TO A COMPLEX MODEL FOR SMOKE DETECTION

Homepage: Sage Continuum
Author: Jakub Szumny, Math and Computer Science Division, University of Illinois at Urbana-Champaign
Mentors: Bhupendra Raut, Seongha Park
Repository: GitHub Repository

Motivation

Forest fires are a major problem, and have detrimental effects on the environment. Current solutions to detecting forest fires are not efficient enough, and other machine learning models have far too long computational speeds and poor accuracies. This study is a continuation of the work done by UCSD, and their SmokeyNet deep learning architecture for smoke detection.
My goal is to compare many different deep learning models, in order to find the best model for this issue, and to find if a simple model can compare to a complex model. The models which I compared are: VGG16, UCSD SmokeyNet, Resnet18, Resnet34, and Resnet50.

Major Accomplishments

Created a large dataset of 41,000 images, comprised of many different wildfire events from HPWREN. I split the images into 5 different classes: sky, ground, horizon, cloud, and smoke.
Tested in many different ways, and found that the best results are when the classes: sky, ground, and horizon, are grouped together as other, and smoke and cloud are left separate. The major issue with this, is that smoke and clouds often look very similar.
On my dataset, created with HPWREN images, each model performed rather well, having about the same accuracy at around 90%.
Found that the VGG16 model with 3 features (smoke, cloud, other), was the best performing model on the testing dataset from ARM, and all the other models performed quite poorly.
Must keep in mind that the burning event was not very obvious in the ARM testing data, but it won’t always be cut and clear, so it is a great test to see which model perform best with the least.
With a FPR of about 13%, a TPR of about 96%, a FNR of about 4%, and a TNR of about 88%, the VGG16 model had the best results, on the ARM Data.
Created a plugin application to be able to test and use my model and algorithm on wild sage nodes, taking images and detecting smoke in real time.

Impact

The impact my research has made, is having created a large dataset for future research, and for better model creation.
Found that a simple model is very accurate and can compare to a complex model.
An algorithm which can compute and classify an entire image in a very short period of time.
This research can greatly help the fight against forest fires, in order to at one point solve the problem of forest fires, by being able to attend to them before they get out of control.

Future Direction

More work is needed on creating a more efficient model. There may be a different model which can perform even better on detecting smoke.
It is helpful as a dataset is already created, and through my Github repository, anyone can replicate my work, and try to improve on it.
Need to explore more ways to augment the images, by scaling the contrast levels, etc, as I believe this would be a good way to separate smoke from cloud from other.

Citation

Dewangan A, Pande Y, Braun H-W, Vernon F, Perez I, Altintas I, Cottrell GW, Nguyen MH. FIgLib & SmokeyNet: Dataset and Deep Learning Model for Real-Time Wildland Fire Smoke Detection. Remote Sensing. 2022; 14(4):1007. https://doi.org/10.3390/rs14041007