Tree Species Recognition Using Combined Attention and ResNet for Unmanned Aerial Vehicle Images

Zhiyang Xu; Qiao Chen; Yongfu Chen

doi:10.3788/LOP212527

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 2 >Page 0210004 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 2, 0210004 (2023)

Tree Species Recognition Using Combined Attention and ResNet for Unmanned Aerial Vehicle Images

Zhiyang Xu^1,2,3, Qiao Chen^1,2,*, and Yongfu Chen^1,2

Author Affiliations

¹Research Institute of Forest Resource Information Techniques, Chinese Academy of Forestry, Beijing 100091, China

²Key Laboratory of Forestry Remote Sensing and Information System, National Forestry and Grassland Administration, Beijing 100091, China

³East China Inventory and Planning Institute, National Forestry and Grassland Administration, Hangzhou 310019, Zhejiang , China

show less

DOI: 10.3788/LOP212527 Cite this Article Set citation alerts

Zhiyang Xu, Qiao Chen, Yongfu Chen. Tree Species Recognition Using Combined Attention and ResNet for Unmanned Aerial Vehicle Images[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0210004 Copy Citation Text

show less

Fig. 1. ECA mechanism block

Download full size

Fig. 2. Basic unit of ECA-ResNet

Download full size

Fig. 3. Overall structure of proposed network

Download full size

Fig. 4. Results of single tree crown segmentation and recognition in five test circular samples

Download full size

Scheme	Optimizer	Training data			Validation data			Independent test data
Scheme	Optimizer	Accuracy /%	Kappa	Loss	Accuracy /%	Kappa	Loss	Accuracy /%	Kappa
I（variable size）	SGD	97.98	0.9638	0.0973	95.73	0.9449	0.1929	77.27	0.7056
II（32×32 pixel）	SGD	93.25	0.9131	0.2711	86.75	0.8293	0.4920	68.18	0.5966
III（64×64 pixel）	SGD	98.98	0.9869	0.1012	96.60	0.9595	0.1678	85.61	0.8140
IV（96×96 pixel）	SGD	97.98	0.9769	0.0553	95.14	0.9388	0.1838	81.82	0.7629
V（128×128 pixel）	Adam	95.19	0.9381	0.2000	96.15	0.9505	0.1453	79.55	0.7332

Table 1. Training and independent test results of the model in single-tree crown image with different patch sizes

Recognized species	Ground true species
Recognized species	Alnus	Other broad-leaves	Cunninghamia	Liriodendron	Pinus
Alnus	26	1	0	0	0
Other broad-leaves	0	18	4	2	2
Cunninghamia	1	3	38	1	1
Liriodendron	0	0	1	15	0
Pinus	2	0	1	0	16
PA /%	89.66	81.82	86.36	83.33	84.21
UA /%	96.30	69.23	86.36	93.75	84.21

Table 2. Confusion matrix of independent test dataset

Network	Training data			Validation data			Independent test data
Network	Accuracy /%	Kappa	Loss	Accuracy /%	Kappa	Loss	Accuracy /%	Kappa
VGG16	95.38	0.9405	0.1783	94.23	0.9258	0.2090	75.00	0.6804
ResNet18	94.36	0.9251	0.1529	93.80	0.9203	0.3176	77.27	0.7036
ResNet34	95.84	0.9464	0.2126	93.80	0.9201	0.2722	79.55	0.7356
ResNet50	96.35	0.9442	0.1505	94.80	0.9301	0.2024	80.30	0.7452
ResNet101	96.69	0.9602	0.1278	94.44	0.9284	0.2348	69.70	0.6130
ResNet152	96.41	0.9567	0.1229	94.87	0.9339	0.2264	65.91	0.5641
resnext50_32x4d	96.77	0.9583	0.0999	94.44	0.9285	0.1845	72.73	0.6501
densenet121	96.40	0.9536	0.1544	93.16	0.9119	0.2911	73.48	0.6557
MobileNetV2	86.41	0.8254	0.4771	88.89	0.8571	0.3599	65.91	0.5555
SqueezeNet	84.01	0.7939	0.5726	86.75	0.8286	0.4139	68.94	0.5948
ECA-ResNet	98.98	0.9869	0.1012	96.60	0.9595	0.1678	85.61	0.8140

Table 3. Performance comparison of different models (single tree crown image clip dataset with 64×64 pixel)

Scheme

Operation

ECA

Training accuracy /%

Validation accuracy /%

Test accuracy /%

FLOPs /Gbit

Parameter

Speed /

（frame·s^-1）

I（ResNet50）

Before reduced

96.35

94.80

80.30

3.827

23518277

4.59

Before reduced

√

96.57

94.51

83.09

3.832

23518325

4.99

III（ECA-ResNet）

After reduced

√

98.98

96.60

85.61

3.015

19886697

5.45

Table 4. Performance comparison of CNN model before and after improvement

Zhiyang Xu, Qiao Chen, Yongfu Chen. Tree Species Recognition Using Combined Attention and ResNet for Unmanned Aerial Vehicle Images[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0210004

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享