Building Extraction from Remote Sensing Imagery Based on Scale-Adaptive Fully Convolutional Network

Fan Feng; Shuangting Wang; Jin Zhang; Chunyang Wang; Bing Liu

doi:10.3788/LOP202158.2428006

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 24 >Page 2428006 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 24, 2428006 (2021)

Building Extraction from Remote Sensing Imagery Based on Scale-Adaptive Fully Convolutional Network

Fan Feng¹, Shuangting Wang¹, Jin Zhang¹, Chunyang Wang^1、*, and Bing Liu²

Author Affiliations

¹School of Surveying and Land Information Engineering, Henan Polytechnic University, Jiaozuo, Henan 454000, China

²PLA Strategic Support Force Information Engineering University, Zhengzhou, Henan 450001, China

show less

DOI: 10.3788/LOP202158.2428006 Cite this Article Set citation alerts

Fan Feng, Shuangting Wang, Jin Zhang, Chunyang Wang, Bing Liu. Building Extraction from Remote Sensing Imagery Based on Scale-Adaptive Fully Convolutional Network[J]. Laser & Optoelectronics Progress, 2021, 58(24): 2428006 Copy Citation Text

show less

Fig. 1. Flow chart of building extraction based on SA-Net

Download full size

Fig. 2. Schematic diagram of the SA-Net

Download full size

Fig. 3. Schematic diagram of the RSPP

Download full size

Fig. 4. Schematic diagram of the AFR module

Download full size

Fig. 5. Two sets of example images. (a) WHU dataset; (b) Massachusetts dataset

Download full size

Fig. 6. Diagram of the overlap strategy

Download full size

Fig. 7. Segmentation results of WHU dataset by different models. (a) Image; (b) label; (c) U-Net; (d) MultiResUNet; (e) Res-UNet; (f) S-UNet; (g) USPP; (h) SA-Net

Download full size

Fig. 8. Segmentation results of the Massachusetts dataset by different models. (a) Image; (b) label; (c) U-Net; (d) MultiResUNet; (e) Res-UNet; (f) S-UNet; (g) USPP; (h) SA-Net

Download full size

Model	U-Net	USPP	S-UNet	Res-UNet	SA-Net	MultiResUNet
Parameter number /10⁶	7.76	4.82	7.97	4.73	7.13	7.26
Max batch size	37	22	23	19	25	6

Table 1. Number of parameters and the maximum number of batches of different models

Model	Batch size	Image size(pixel×pixel)	Graphic card	Video memory /G	IOU /%
U-Net (random cropping training)	16	256×256	RTX 2070	8	88.58
U-Net (Ref. [15])	8	512×512	Nvidia P6000	24	84.08
U-Net (Ref. [4])	6	512×512	Nvidia Titan XP	12	86.80

Table 2. Random sampling training and regular training results of different models in the WHU dataset

Parameter	WHU ariel dataset	Massachusetts dataset
Training image number (size)	4736 (512 pixel×512 pixel)	8631 (512 pixel×512 pixel)
Validation image number (size)	4144 (256 pixel×256 pixel)	144 (256 pixel×256 pixel)
Training epoch	300	200
Steps per epoch	296	540
Batch size	16(6 for MultiResUNet)	16(6 for MultiResUNet)
Iteration number	296×300	540×200
Padding size	0	64

Table 3. Experimental settings of WHU ariel dataset and Massachusetts dataset

Dataset	Model	Precision	Recall	IOU	F1 score
WHU	U-Net	94.37	93.52	88.58	93.94
	USPP	94.50	94.35	89.44	94.42
	MultiResUNet	97.00	90.01	87.57	93.37
	S-UNet (Ref. [14])	95.20	93.00	88.80	94.09
	SR-FCN (Ref. [4])	94.40	93.90	88.90	94.15
	S-UNet	94.74	93.77	89.14	94.25
	DeepLab V3+ (Ref. [4])	91.60	94.60	87.10	93.08
	Res-UNet	92.71	93.90	87.44	93.30
	SA-Net	95.27	93.80	89.62	94.53
Massachusetts	U-Net	85.84	81.18	71.60	83.44
	MultiResUNet	93.22	66.84	63.74	77.86
	USPP	88.50	79.37	71.95	83.69
	S-UNet	86.05	81.50	71.99	83.71
	Res-UNet	87.08	77.66	69.64	82.10
	Res-UNet (Ref. [11])	86.21	80.26	71.14	83.13
	JointNet (Ref. [11])	86.21	81.29	71.99	83.68
	SA-Net	86.78	82.70	73.45	84.69

Table 4. Quantitative evaluation results of different models on the WHU and Massachusetts datasets unit: %

Model	U-Net	USPP	S-UNet	Res-UNet	SA-Net	MultiResUNet
Training time	10.7	11.8	12.8	13.4	13.3	35.1

Table 5. Training time of different models on the WHU dataset unit: h

Dataset	Index	U-Net (base-line)	MIMO	RSPP	AFR	IOU	F1 score
WHU	1	Ö				88.58	93.94
	2	Ö	Ö			89.37	94.38
	3	Ö	Ö	Ö		89.67	94.55
	4	Ö	Ö	Ö	Ö	89.62	94.53
Massachusetts	1	Ö				71.60	83.44
	2	Ö	Ö			73.02	84.41
	3	Ö	Ö	Ö		73.06	84.44
	4	Ö	Ö	Ö	Ö	73.45	84.69

Table 6. Evaluation results of ablation experiments unit: %

Dataset	Model	Precision	Recall	IOU	F1 score
WHU	U-Net	83.03	85.87	73.05	84.43
	USPP	87.42	86.60	77.00	87.01
	S-UNet	87.11	86.64	76.80	86.87
	SA-Net	88.92	86.23	77.86	87.55
Massachusetts	U-Net	86.30	73.46	65.79	79.36
	USPP	86.64	75.79	67.86	80.85
	S-UNet	84.56	79.21	69.20	81.80
	SA-Net	87.49	79.28	71.21	83.18

Table 7. Experimental results of small sample conditions unit: %

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information