Multi-Scale Sea-Land Segmentation Method for Remote Sensing Images Based on Res2Net

Hui Gao; Xiaodong Yan; Heng Zhang; Yiting Niu; Jiaqi Wang

doi:10.3788/AOS202242.1828004

Acta Optica Sinica, Volume. 42, Issue 18, 1828004(2022)

Multi-Scale Sea-Land Segmentation Method for Remote Sensing Images Based on Res2Net

Hui Gao, Xiaodong Yan, Heng Zhang^*, Yiting Niu, and Jiaqi Wang

Institute of Geospatial Information, Strategic Support Force Information Engineering University, Zhengzhou 450001, Henan, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Architecture of the proposed network

Download full size

Fig. 2. Architectures of Res_Block and Res2_Block. (a) Res_Block; (b) Res2_Block (s=4)

Download full size

Fig. 3. Architectures of attention module. (a) Architecture of SE module; (b) architecture of SA module

Download full size

Fig. 4. Comparison of experimental results of various networks on Data1. (a) Original image; (b) original label; (c) U-Net; (d) Deeplabv3+; (e) U²-Net; (f) RAUNet; (g) MSRNet；(h) water line superposition result

Download full size

Fig. 5. Comparison of experimental results of various networks on Data2. (a) Original image; (b) original label; (c) U-Net; (d) Deeplabv3+; (e) U²-Net; (f) RAUNet; (g) MSRNet; (h) water line superposition result

Download full size

Fig. 6. Ablation experimental results on Data1. (a) En_Decoder+ResNet; (b) En_Decoder+Res2Net; (c) En_Decoder+Res2Net+DS; (d) MSRNet

Download full size

Table 1. Input and output parameters of encoder layer

View table

Table 1. Input and output parameters of encoder layer

Encoder layer	Input	Convolution type	Kernel size	Stride	Output
Encoder_1	512×512×3 256×256×64	Conv maxpooling	7×7 3×3	2 2	256×256×64 128×128×256
Encoder_2	128×128×256	3×Res2_Block	3×3	1	128×128×256
Encoder_3	128×128×256	4×Res2_Block	3×3	2	64×64×512
Encoder_4	64×64×512	6×Res2_Block	3×3	2	32×32×1 024
Encoder_5	32×32×1 024	3×Res2_Block	3×3	2	16×16×2 048

Table 2. Input and output parameters of decoder layer

View table

Table 2. Input and output parameters of decoder layer

Layer	Input	Convolution type	Output	Side-output
Concat	512×512×10	Conv	512×512×2
Decoder_5	16×16×2 048 16×16×1 024	SA_5 Res2_Block+Conv+upsample	16×16×1 024 32×32×1 024	512×512×2
Decoder_4	32×32×1 024 32×32×512	SA_4 Res2_Block+Conv+upsample	32×32×512 64×64×512	512×512×2
Decoder_3	64×64×512 64×64×256	SA_3 Res2_Block+Conv+upsample	64×64×256 128×128×256	512×512×2
Decoder_2	128×128×256 128×128×64	SA_2 Res2_Block+Conv+upsample	128×128×64 256×256×64	512×512×2
Decoder_1	256×256×64 256×256×32	SA_1 Res2_Block+Conv+upsample	256×256×32 512×512×2	512×512×2

Table 3. Details of two datasets

View table

Table 3. Details of two datasets

Parameter	Data1	Data2
Image source	Gaofen-1	Landsat-8
Bands	Red，green，blue，near-infrared	Red，green，blue
Spatial resolution /m	8	30
Image size	256×256	512×512
Number of training samples	1544	1950
Number of validation samples	178	1411
Number of test samples	192	-

Table 4. Segmentation results of each method

View table

Table 4. Segmentation results of each method

Method	Data1				Data2
Method	F1-score /%	MIOU /%	MAE	F1-score-b /%	F1-score /%	MIOU /%	MAE	F1-score-b /%
U-Net	97.92	95.92	0.021	51.90	98.13	96.33	0.979	52.84
Deeplabv3+	98.23	96.53	0.018	48.11	98.53	97.10	0.709	55.48
U²-Net	97.92	95.92	0.023	60.68	98.30	96.65	0.762	52.05
RAUNet	98.38	96.82	0.016	69.25	98.39	96.84	0.694	49.77
MSRNet	99.09	98.19	0.009	80.14	98.95	97.93	0.507	72.86

Table 5. MSRNet ablation experiments
View table
Table 5. MSRNet ablation experiments
Architecture of network F1-score-b /%
En_Decoder+ResNet 63.63
En_Decoder+Res2Net 74.08
En_Decoder+Res2Net+DS 75.11
En_Decoder+Res2Net+DS+SA（MSRNet） 80.14

Tools

Get Citation

Copy Citation Text

Hui Gao, Xiaodong Yan, Heng Zhang, Yiting Niu, Jiaqi Wang. Multi-Scale Sea-Land Segmentation Method for Remote Sensing Images Based on Res2Net[J]. Acta Optica Sinica, 2022, 42(18): 1828004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites