Semantic Segmentation of Multispectral Remote Sensing Images Based on Band-Location Adaptive Selection

Zhengyin Liang; Xili Wang

doi:10.3788/LOP222250

Laser & Optoelectronics Progress, Volume. 60, Issue 14, 1410016(2023)

Semantic Segmentation of Multispectral Remote Sensing Images Based on Band-Location Adaptive Selection

Zhengyin Liang and Xili Wang^*

Author Affiliations

School of Computer Science, Shaanxi Normal University, Xi'an 710000, Shaanxi, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Structure schematic of BLASeNet

Download full size

Fig. 2. Diagram of band-location adaptive selection mechanism

Download full size

Fig. 3. Residual blocks. (a) Original residual block; (b) proposed 3D residual block

Download full size

Fig. 4. Segmentation result maps obtained by different methods on the Potsdam dataset

Download full size

Fig. 5. Category space distribution map of Potsdam dataset

Download full size

Fig. 6. Visualisation results of different layer feature maps on the Potsdam dataset

Download full size

Fig. 7. Segmentation result maps obtained by different methods on the Qinghai dataset

Download full size

Fig. 8. Variation of parameter values of the trainable parameter matrix in the ADF₁ and ADF₂ modules. (a) ADF₁; (b) ADF₂

Download full size

Fig. 9. Segmentation result maps obtained by different methods on the Tibet Plateau dataset

Download full size

Table 1. Structure comparison of different methods

View table

Table 1. Structure comparison of different methods

Method	Backbone（layer）	Convolution type	Attention	Feature fusion	Band selection
U-Net^［9］	14 layers	2D Conv		√
ResNet34^［38］	ResNet-34（34 layers）	2D Conv		√
MAResU-Net^［37］	ResNet-34（34 layers）	2D Conv	√	√
DeepLabv3+^［22］	ResNet-101（101 layers）	2D Conv		√
BLASeNet-A	20 layers	3D Conv	√	√	√
BLASeNet-M	20 layers	3D Conv	√	√	√
BLASeNet-C	20 layers	3D Conv	√	√	√

Table 2. Segmentation results on Potsdam dataset

View table

Table 2. Segmentation results on Potsdam dataset

Class name	clutter/ background	impervious surfaces	building	low vegetation	tree	car	A_OA	mean $S_{F 1}$	R_MIoU
U-Net^［9］	26.93	81.82	84.81	76.26	74.81	79.72	75.71	70.73	57.68
ResNet34^［38］	24.94	82.28	86.90	76.73	74.25	79.63	76.44	70.79	58.07
MAResU-Net^［37］	32.82	83.43	87.79	78.82	76.96	82.07	78.31	73.45	61.10
DeepLabv3+^［22］	38.98	82.56	87.47	77.44	76.53	81.23	78.23	74.03	60.96
BLASeNet-A	39.77	84.16	88.40	79.33	79.50	81.64	79.64	75.47	62.90
BLASeNet-M	38.81	83.96	88.19	79.76	79.46	82.21	79.70	75.40	62.89
BLASeNet-C	38.84	84.17	88.17	79.44	77.92	82.00	79.35	75.09	62.47

Table 3. Segmentation results on the Qinghai dataset

View table

Table 3. Segmentation results on the Qinghai dataset

Class name	grass	water	cloud	bare soil	A_OA	mean $S_{F 1}$	R_MIoU
U-Net^［9］	82.37	96.71	64.91	86.05	86.14	82.51	71.80
ResNet34^［38］	82.99	97.41	69.55	86.69	86.89	84.16	73.92
MAResU-Net^［37］	84.06	96.91	69.39	87.34	87.47	84.43	74.29
DeepLabv3+^［22］	84.09	97.02	67.68	86.98	87.29	83.94	73.72
BLASeNet-A	85.37	97.57	77.08	87.89	88.39	86.98	77.71
BLASeNet-M	86.18	97.44	73.28	88.17	88.67	86.27	76.85
BLASeNet-C	85.89	97.37	72.45	88.35	88.63	86.02	76.52

Table 4. Segmentation results on the Tibet Plateau dataset

View table

Table 4. Segmentation results on the Tibet Plateau dataset

Class name	bare ground	construction land	ice/clouds	water	grass	woodland/ shrubs	A_OA	mean $S_{F 1}$	R_MIoU
U-Net^［9］	46.20	16.79	42.79	88.54	78.89	37.97	68.59	51.86	39.07
ResNet34^［38］	44.13	15.43	42.19	88.78	78.59	35.59	67.97	50.78	38.26
MAResU-Net^［37］	42.29	12.88	41.04	89.13	78.50	37.27	68.02	50.19	37.90
DeepLabv3+^［22］	46.62	0.00	32.29	85.35	77.96	39.34	67.57	46.93	35.41
BLASeNet-A	44.47	29.10	44.06	90.03	79.07	39.35	68.68	54.31	40.91
BLASeNet-M	49.84	20.62	42.52	88.25	79.36	43.80	69.60	54.06	40.75
BLASeNet-C	49.51	20.50	43.32	88.53	78.79	36.28	68.92	52.82	39.76

Table 5. Ablation experiments on the Potsdam dataset

View table

Table 5. Ablation experiments on the Potsdam dataset

Class name	clutter/background	impervious surfaces	building	low vegetation	tree	car	A_OA	mean $S_{F 1}$	R_MIoU	Test time /min
Baseline	24.94	82.28	86.90	76.73	74.25	79.63	76.44	70.79	58.07	2.05
+3DRB	35.47	83.07	87.36	77.87	77.56	81.83	78.26	73.86	61.08	2.74
+BLAS	36.72	84.27	88.50	78.58	79.04	82.18	79.29	74.88	62.42	2.75
+ADF	39.32	84.04	88.27	79.29	78.94	82.29	79.49	75.36	62.79	2.75
+CAM	39.77	84.16	88.40	79.33	79.50	81.64	79.64	75.47	62.90	2.76

Tools

Get Citation

Copy Citation Text

Zhengyin Liang, Xili Wang. Semantic Segmentation of Multispectral Remote Sensing Images Based on Band-Location Adaptive Selection[J]. Laser & Optoelectronics Progress, 2023, 60(14): 1410016

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Image Processing

Received: Aug. 5, 2022

Accepted: Sep. 26, 2022

Published Online: Aug. 10, 2023

The Author Email: Wang Xili (wangxili@snnu.edu.cn)

DOI:10.3788/LOP222250

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology