LEM-YOLO-Based Lightweight Multi-Scale Detection of Forest Fire Smoke in UAV Imagery

Fig. 9. Comparison of detection results. （a1）‒（a3） Flame with drastic scale changes；（b1）‒（b3） flame obscured by smoke；（c1）‒（c3） small-target flame obscured by trees；（d1）‒（d3） smoke with indistinct features；（e1）‒（e3） smoke confused with background

Download full size

Table 1. Experimental parameters
View table
Table 1. Experimental parameters
Parameter name Parameter value
Epoch 200
Batch size 16
Optimizer SGD
Learning rate 0.01
Weight decay 0.0005
Momentum 0.937
Image size /（pixel×pixel） 640×640

Table 2. Ablation experiment

View table

Table 2. Ablation experiment

No.	Model	AP /%		mAP@0.5 /%	mAP@0.5：0.95 /%	Params /10⁶	FLOPs /10⁹	FPS /（frame/s）	Size /MB
No.	Model	Fire	Smoke	mAP@0.5 /%	mAP@0.5：0.95 /%	Params /10⁶	FLOPs /10⁹	FPS /（frame/s）	Size /MB
0	YOLOv8n	85.9	95.7	90.8	66.2	3.01	8.1	68.81	6.3
1	+A	86.7	97.3	92.0	68.5	3.15	8.6	62.59	6.6
2	+A+B	86.5	97.3	91.9	68.3	2.70	7.5	61.31	5.7
3	+A+B+C	87.2	97.2	92.2	68.6	2.06	5.9	61.25	4.8
4	+A+B+C+D	88.1	97.2	92.7	68.7	2.06	5.9	57.82	4.8

Table 3. Comparative experiments of different attention mechanisms

View table

Table 3. Comparative experiments of different attention mechanisms

Attention	mAP@0.5 /%	mAP@0.5：0.95 /%	Params /10⁶	FLOPs /10⁹	FPS /（frame/s）
None	90.8	66.3	3.05	8.3	64.45
SE	90.9	66.8	3.11	8.5	61.65
CA	91.5	67.9	3.11	8.4	59.50
CBAM	91.3	67.5	3.45	10.9	50.83
ECA	91.3	67.5	3.12	8.1	57.15
EMA	91.8	68.1	3.15	8.5	62.59

Table 4. Comparative experiments of different loss functions
View table
Table 4. Comparative experiments of different loss functions
Lossfunction Precision /% Recall /% mAP@0.5 /% mAP@0.5：0.95 /%
SIoU 89.1 87.5 92.1 68.4
EIoU 89.5 86.1 91.9 68.6
Shape-IoU 89.6 85.8 92.3 68.4
MDPIoU 88.8 86.4 91.9 68.1
PIoU 89.5 87.5 92.7 68.7

Table 5. Comparative experiments of different object detection models

View table

Table 5. Comparative experiments of different object detection models

Model	mAP@0.5 /%	mAP@0.5：0.95 /%	Params /10⁶	FLOPs /10⁹	FPS /（frame/s）
Faster R-CNN	74.5	54.7	193.83	171.4	27.55
Sparse R-CNN	81.2	59.9	105.92	64.2	28.12
TOOD	81.5	60.2	31.79	180.9	35.65
YOLOv3-tiny	75.5	55.2	12.13	18.9	51.45
YOLOv5n	89.7	64.7	2.50	7.1	55.90
YOLOv6n	88.2	64.1	4.23	11.8	47.87
YOLOv7-tiny	91.4	67.7	6.01	13.1	49.55
YOLOv8n	90.8	66.2	3.01	8.1	68.81
YOLOv9s	91.1	66.5	7.20	26.7	50.51
YOLOv10n	90.4	65.7	2.32	6.7	70.15
YOLOv11n	90.9	66.3	2.58	6.3	67.25
RT-DETR r18	92.7	68.6	19.80	56.9	50.35
LEM-YOLO	92.7	68.7	2.06	5.9	57.82

Table 6. Comparison of false positive and false negative rates after detection
View table
Table 6. Comparison of false positive and false negative rates after detection
Model FPR of fire /% FPR of smoke /% FNR of fire /% FNR of smoke /%
YOLOv8n 20.4 2.0 19.5 11.7
LEM-YOLO 19.2 1.8 16.3 10.4

Tools

Get Citation

Copy Citation Text

Ruijie Kuang, Xiang Li, Yu Liu, Bingying Hu, Xianshun Wang. LEM-YOLO-Based Lightweight Multi-Scale Detection of Forest Fire Smoke in UAV Imagery[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1437003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites