Optical remote sensing road extraction network with directional guidance and topological awareness

Step 1： $d_{c o n}$ ← $σ (W_{2} δ (W_{1} \frac{1}{H_{e} \times W_{e}} \sum_{i = 1}^{H_{e}} \sum_{j = 1}^{W_{e}} E_{5} (i, j)))$ #计算多种方向之间的依赖关系，得到方向向量 $d_{c o n}$ ， $W_{1}$ 表示卷积， $δ$ 与 $σ$ 分别表示ReLu与Sigmoid激活函数

Step 2：按照道路八向关系将 $d_{c o n}$ 与 $E_{5}$ 进行通道切片，得到 $d_{c o n}^{j}$ ， $E_{5}^{j}$ ， $j \in (1,2, \dots, 8)$

Step 3：For $j$ to 8 #为各方向构建特定学习路径，给予各方向不同的关注度

Step 3.1： $E_{5}^{j}$ ← $C o n v (C P A M (E_{5}^{j}) + C C A M (E_{5}^{j}))$ #捕获空间、通道维度依赖关系

Step 3.2： $E_{c o n}^{j}$ ← $W_{3}^{j} F_{c} (d_{c o n}^{j}, E_{5}^{j'}) + E_{5}^{j}$ #根据 $d_{c o n}^{j}$ 中该方向的重要性，引导网络给予各方向不同的关注度， $F_{c}$ 表示通道乘法， $W_{3}^{j}$ 表示卷积

Step 4： $E_{M D G}$ ← $= C o n v (c o n c a t (E_{c o n}^{1}, E_{c o n}^{2}, \dots E_{c o n}^{8}))$ #将各方向输出 $E_{5}^{j}$ 按通道进行拼接，得到方向感知特征图 $E_{M D G}$ ，算法结束

Table 3. Calculation process of TALoss

View table

View in Article

Table 3. Calculation process of TALoss

算法2：拓扑感知损失函数步骤

Input：预测结果O，标签G

Output：TALoss

Step 1：计算尺度参数 $ε$ 变化期间的复形序列 $O_{N}$ ， $G_{N}$

Step 1.1：寻找O与G中不同维度的拓扑特征点 $(ε_{b}^{i}, ε_{d}^{i})$ ，并表示为持续条形码 $P H_{i}^{k} (G)$ ， $P H_{i}^{k} (O)$ # $ε_{b}^{i}$ ， $ε_{d}^{i}$ 分别代表第 $i$ 个特征点的出生时间与死亡时间， $k \in 0,1, 2$ ，表示特征维度

Step 1.2：利用 $P H_{i}^{k} (G)$ ， $P H_{i}^{k} (O)$ ，计算持久图 $P D_{O}^{k}$ ， $P D_{G}^{k}$

Step 1.3： $L_{P D}^{k}$ ←Hausdorff（ $P D_{O}^{k}$ ， $P D_{G}^{k}$ ） #利用平均Hausdorff距离计算预测路网与真实路网的拓扑相似度

Step 2： $L_{P D}^{}$ ← $L_{P D}^{0} + L_{P D}^{1}$ #统计两个维度的拓扑差异

Step 3：TALoss← $L_{P D}^{} + L_{B C E}$ #联合像素级损失与拓扑损失，得到TALoss

Table 4. Confusion matrix
View table
View in Article
Table 4. Confusion matrix
样本预测为道路预测为背景
真实道路 TP FN
真实背景 FP TN

Table 5. Quantitative evaluation results on the DeepGlobe dataset

View table

View in Article

Table 5. Quantitative evaluation results on the DeepGlobe dataset

Methods	Precision	Recall	F1-score	IoU
FCN	70.53	73.65	72.06	56.32
SegNet	69.60	78.17	73.64	58.28
U-Net	72.70	77.66	75.10	60.12
PSPNet	66.27	79.48	72.28	56.47
D-LinkNet	74.64	80.03	77.24	64.55
SDUNet	78.40	74.20	79.40	66.80
RADANet	/	/	73.67	/
CoANet	/	/	81.22	68.37
BDTNet	84.18	76.77	80.30	67.09
CR-HR-RoadNet	76.47	77.12	76.79	62.33
SASwin Transformer	80.97	77.56	79.23	65.60
CFRNet	81.03	79.77	80.40	67.22
RDTA-Net	80.69	83.26	81.95	69.35

Table 6. Quantitative evaluation results on the Massachusetts dataset

View table

View in Article

Table 6. Quantitative evaluation results on the Massachusetts dataset

Methods	Precision	Recall	F1-score	IoU
FCN	68.52	68.94	68.73	52.16
SegNet	74.33	77.14	75.71	60.32
U-Net	75.81	78.73	77.24	62.54
PSPNet	75.92	77.33	76.61	61.49
D-LinkNet	76.99	78.22	77.60	63.45
CR-HR-RoadNet	80.34	76.15	78.19	64.19
SASwin Transformer	80.02	77.65	78.82	65.04
CADUNet	79.45	76.55	77.89	64.12
SGCN	84.82	73.91	78.99	65.28
DDU-Net	82.54	73.99	78.03	63.98
CRAENet	81.21	77.48	78.48	65.72
D-FusionNet	/	/	78.11	64.08
RDTA-Net	79.91	80.05	79.98	66.38

Table 7. DeepGlobe ablation experiment results

View table

View in Article

Table 7. DeepGlobe ablation experiment results

Model	Baseline	MDG	FGC	$L_{B C E}$	$L_{T A}$	Precision/%	Recall/%	F1-score/%	IoU/%
Model1	√			√		75.08	79.11	77.04	63.08
Model2	√	√		√		76.67	80.22	78.40	65.28
Model3	√		√	√		77.18	79.82	78.48	65.36
Model4	√	√	√	√		78.83	81.53	80.16	67.04
Model5	√	√	√		√	80.69	83.26	81.95	69.35

Table 8. Massachusetts ablation experiment results

View table

View in Article

Table 8. Massachusetts ablation experiment results

Model	Baseline	MDG	FGC	$L_{B C E}$	$L_{T A}$	Precision/%	Recall/%	F1-score/%	IoU/%
Model1	√			√		76.11	79.03	77.54	63.18
Model2	√	√		√		77.23	79.42	78.31	64.26
Model3	√		√	√		77.62	79.12	78.36	64.33
Model4	√	√	√	√		78.80	79.54	79.17	65.79
Model5	√	√	√		√	79.91	80.05	79.98	66.38

Table 9. Impact of different loss functions on road extraction

View table

View in Article

Table 9. Impact of different loss functions on road extraction

Methods	Loss	Precision/%	Recall/%	F1-score/%	IoU/%	$β_{0}$ error	$β_{1}$ error
FCN	$L_{B C E}$	70.53	73.65	72.06	56.32	1.816	1.738
	clDice	71.21	74.30	72.72	56.96	1.715	1.639
	$L_{T A}$	72.14	75.74	73.90	58.60	1.522	1.396
SegNet	$L_{B C E}$	69.60	78.17	73.64	58.28	1.584	1.483
	clDice	70.13	78.82	74.22	58.69	1.489	1.317
	$L_{T A}$	70.92	79.25	74.85	59.26	1.329	1.218
U-Net	$L_{B C E}$	72.70	77.66	75.10	60.12	1.274	1.143
	clDice	73.56	78.49	75.95	61.07	1.183	1.085
	$L_{T A}$	74.35	79.38	76.79	62.32	1.077	0.942
PSPNet	$L_{B C E}$	66.27	79.48	72.28	56.47	1.764	1.681
	clDice	67.06	79.93	72.93	57.43	1.692	1.590
	$L_{T A}$	68.22	80.52	73.86	58.96	1.451	1.264
RDTA-Net	$L_{B C E}$	78.83	81.53	80.16	67.04	0.708	0.691
	clDice	79.28	82.37	80.80	67.85	0.643	0.617
	$L_{T A}$	80.69	83.26	81.95	69.35	0.541	0.581

Table 10. Time consuming comparison of loss function calculation
View table
View in Article
Table 10. Time consuming comparison of loss function calculation
Methods Loss IoU/% Training time/s
RDTA-Net BCE 67.04 0.226
RDTA-Net clDice 67.85 0.279
RDTA-Net TALoss 69.35 0.305

Table 11. Quantitative experimental analysis in different scenarios

View table

View in Article

Table 11. Quantitative experimental analysis in different scenarios

Methods	遮挡情况		噪声干扰		光照变化
Methods	F1-score	IoU	F1-score	IoU	F1-score	IoU
FCN	62.25	48.83	61.14	47.26	61.22	48.07
SegNet	67.19	54.41	66.23	54.03	65.69	53.82
U-Net	68.91	55.02	67.68	54.87	67.34	54.71
D-LinkNet	71.51	57.17	71.25	56.54	70.61	55.58
RDTA-Net	73.42	61.63	72.86	60.38	72.47	60.26

Table 12. Complexity Analysis
View table
View in Article
Table 12. Complexity Analysis
Methods Params FLOPs FPS Training time
FCN 15.31 80.47 85 0.156
SegNet 18.82 97.11 78 0.173
U-Net 31.04 150.85 64 0.195
PSPNet 53.32 218.61 31 0.379
D-LinkNet 31.1 157.99 69 0.218
SDUNet 80.24 314.54 32 0.462
RDTA-Net 47.26 194.72 53 0.305

Tools

Get Citation

Copy Citation Text

Yuebo MENG, Xinyu HUANG, Shilong SU, Heng WANG. Optical remote sensing road extraction network with directional guidance and topological awareness[J]. Optics and Precision Engineering, 2025, 33(10): 1638

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites