Building Change Detection in High-Resolution Remote-Sensing Images Based on Deep Learning

Fig. 5. Comparison of detection results of building changes with different scales. (a) First-phase remote sensing image; (b) second-phase remote sensing image; (c) ground truth; (d) UNet; (e) ChangeNet; (f) CSCDNet; (g) proposed method

Download full size

Table 1. Network structure of feature extraction stage

View table

Table 1. Network structure of feature extraction stage

ResNet50	Size
7×7， 64， stride2	112×112
3×3， max pooling， stride2	56×56
$[\begin{matrix} \begin{array}{l} 1 \times 1 \\ 3 \times 3 \end{array} & \begin{array}{l} 64 \\ 64 \end{array} \\ 1 \times 1 & 256 \end{matrix}]$ ×3	56×56
$[\begin{matrix} \begin{array}{l} 1 \times 1 \\ 3 \times 3 \end{array} & \begin{array}{l} 128 \\ 128 \end{array} \\ 1 \times 1 & 514 \end{matrix}]$ ×4	28×28
$[\begin{matrix} \begin{array}{l} 1 \times 1 \\ 3 \times 3 \end{array} & \begin{array}{l} 256 \\ 256 \end{array} \\ 1 \times 1 & 1024 \end{matrix}]$ ×6	14×14
$[\begin{matrix} \begin{array}{l} 1 \times 1 \\ 3 \times 3 \end{array} & \begin{array}{l} 512 \\ 512 \end{array} \\ 1 \times 1 & 2048 \end{matrix}]$ ×3 （dilated convolution）	14×14

Table 2. Lab environment

View table

Table 2. Lab environment

Lab environment	Configuration
CPU	6×Intel（R）Xeon（R）CPU E5-2678 v3@2.50 GHz
GPU	NVIDIA GeForce RTX 2080 Ti
Memory	62 GB
Operating system	Ubuntu 18.04
Deep learning framework	Pytorch1.6
Programming language	Python 3.7
GPU processing framework	CUDA 10.0， CUDNN 7.6

Table 3. Hyperparameters’ optimization of neural network
View table
Table 3. Hyperparameters’ optimization of neural network
Initial learning rate Maximum number of iterations F1/%
0.1 500 69.76
0.01 80.01
0.001 88.89
0.0001 90.38
0.00001 88.17
0.0001 50 85.64
100 88.58
200 90.17
500 90.12

Table 4. Ablation experiment
View table
Table 4. Ablation experiment
Method Precision /% Recall /% F1 /%
Baseline 80.362 82.946 81.273
+CBAM 81.261 84.594 82.884
+PPM 80.567 83.339 81.460
Proposed method 82.508 85.232 83.324

Table 5. Evaluation of change detection results of different methods
View table
Table 5. Evaluation of change detection results of different methods
Method Precision /% Recall /% F1 /% Time /h
UNet 72.507 69.421 70.864 8.7
ChangeNet 74.659 70.215 72.337 9.4
CSCDNet 80.295 82.347 81.362 11.3
method
Proposed
82.508 85.232 83.324 10.1

Tools

Get Citation

Copy Citation Text

Xing Han, Ling Han, Liangzhi Li, Huihui Li. Building Change Detection in High-Resolution Remote-Sensing Images Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2022, 59(10): 1001003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites