Forest Fire Detection from UAV Images Using Fusion of Pre-trained Mobile CNN Features
Forest Fire Detection from UAV Images Using Fusion of Pre-trained Mobile CNN Features
- New search for: Mohan, Bhuma Chandra
- New search for: Jain, Kamal
- New search for: Mishra, Vishal
- New search for: Pradhan, Biswajeet
- Further information on Pradhan, Biswajeet:
-
https://orcid.org/https://orcid.org/0000-0001-9863-2054
- New search for: Mohan, Bhuma Chandra
In this work, a Convolutional Neural Network based approach is presented for accurate classification of forest areas with fire from UAV images. In general, the deeper the CNN architecture, the classification of ‘fire’ versus ‘no fire’ is more accurate. However, deeper architectures consume lot of battery power and impose constraints on the processor used in UAV. It is time taking too. Hence, architectures like ResNet50 are not suitable as 23 million parameters are required to train a ResNet50 model. In this regard, mobile CNN architectures are quite handy and they require very few parameters of typical 1–7 millions. They are faster also and take very less time for inference. In this work, the features from selected pre-trained mobile CNN architectures i.e., Squeezenet, MobileNetv1, MobileNetv2, MnasNet, MobileNet v3, SqueezeNext, ShuffleNet, CondenseNet, DiCENet, FBNet, MixNet, and EfficientNet Lite-0, EfficientNet Lite-1 are used in the classification process. All the architectures are pre-trained on ‘imagenet’ dataset with 1000 classes and 14 millions of images. Features from the last pooling layer of each network are obtained. Feature fusion (concatenation) from the selected mobile CNN architectures is considered for classifying the images with ‘fire’ and ‘no fire’. SVM classifier is applied to the fused feature vector. In general, as the size of the fused feature vector increases, the classification accuracy increases. A wildfire image dataset with 2096 images is chosen with balanced classes of ‘fire’ and ‘no fire’. With a 80% train and test split, the mean classification accuracy obtained is in excess of 98%. Various other performance metrics are also given to emphasize the merit of the proposed approach.
Document information
Forest Fire Detection from UAV Images Using Fusion of Pre-trained Mobile CNN Features
Lecture Notes in Civil Engineering
International Conference on Unmanned Aerial System in Geomatics ; 2021 ; Roorkee, India April 02, 2021 - April 04, 2021
2023-03-16
12 pages
Article/Chapter (Book)
Electronic Resource
English
Similar titles
Forest fire detection and estimation of forest fire risk index from SEVIRI synthetic images
| British Library Conference Proceedings | 2000
Forest Fire Monitoring Using Envisat-AATSR and MERIS Images
| British Library Conference Proceedings | 2008
Forest Fire Burned Biomass Estimation Using Satellite Images and Reference Data
| British Library Conference Proceedings | 2014