Title Page
VITA
Abstract
Contents
Chapter 1. Introduction 12
1.1. Overview of forest fire detection 12
1.2. Related work 12
1.3. Evaluation methodology 13
1.4. Thesis Outline 13
Chapter 2. Background 15
2.1. Convolution neural networks 15
2.1.1. Convolution 15
2.1.2. Pooling 16
2.2. Forest fire detection model 17
2.3. Motivation 19
Chapter 3. Survey of Forest Fire Detection 22
3.1. Traditional approaches 22
3.2. CNN-based approaches 22
Chapter 4. Methodology 24
4.1. Backbone architecture 24
4.1.1. Stem Block 24
4.1.2. Residual block 26
4.1.3. Transition block 28
4.1.4. Attention block 29
4.2. Neck architecture 32
4.3. Head architecture 33
4.4. Loss function 34
Chapter 5. Experiments 36
5.1. Dataset 36
5.2. Experimental setup 36
5.3. Evaluation metrics 37
5.4. Experimental results 38
5.5. Ablation study 43
Chapter 6. Contribution Summary and Further Work 44
6.1. Contribution summary 44
6.2. Future work 44
Bibliography 45
Table 5.1. Performance Comparison of our model and the other models 38
Table 5.2. Performance Comparison of proposed backbone and other backbones 39
Table 5.3. Ablation study on backbone modules with different techniques 43
Figure 2.1. An example of the convolution 15
Figure 2.2. An example of the max-pooling 16
Figure 2.3. The architecture of the forest fire detection model 18
Figure 2.4. The same receptive field of using of three 3x3 kernels and one 7x7 kernel 19
Figure 2.5. Depth-wise convolutions of multi kernel size 20
Figure 4.1. Backbone architecture 24
Figure 4.2. Stem block 25
Figure 4.3. Residual block 27
Figure 4.4. Transition block 28
Figure 4.5. CBAM architecture 29
Figure 4.6. Neck architecture 32
Figure 4.7. Head architecture 33
Figure 5.1. The qualitative results for forest fire detection on our dataset. 41
Figure 5.2. Grad-CAM visualization 42