Landsat 8 Remote Sensing Image Based on Deep Residual Fully Convolutional Network

Jiaqiang Zhang; Xiaoyan Li; Liyuan Li; Pengcheng Sun; Xiaofeng Su; Tingliang Hu; Fansheng Chen

doi:10.3788/LOP57.102801

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 10 >Page 102801 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 10, 102801 (2020)

Landsat 8 Remote Sensing Image Based on Deep Residual Fully Convolutional Network

Jiaqiang Zhang^1、2、3, Xiaoyan Li^1、2、3, Liyuan Li^1、2、3, Pengcheng Sun^2、4, Xiaofeng Su^1、2、*, Tingliang Hu^1、2, and Fansheng Chen^1、2

Author Affiliations

¹Key Laboratory of Intelligent Infrared Perception, Chinese Academy of Sciences, Shanghai 200083, China

²Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China

³University of Chinese Academy of Sciences, Beijing 100049, China

⁴Key Laboratory of Specialty Fiber Optics and Optical Access Networks, Joint International Research Laboratory of Specialty Fiber Optics and Advanced Communication, Shanghai Institute of Advanced Communication and Data Science, Shanghai University, Shanghai 200444, China

show less

DOI: 10.3788/LOP57.102801 Cite this Article Set citation alerts

Jiaqiang Zhang, Xiaoyan Li, Liyuan Li, Pengcheng Sun, Xiaofeng Su, Tingliang Hu, Fansheng Chen. Landsat 8 Remote Sensing Image Based on Deep Residual Fully Convolutional Network[J]. Laser & Optoelectronics Progress, 2020, 57(10): 102801 Copy Citation Text

show less

Fig. 1. Block diagram of proposed method

Download full size

Fig. 2. Architecture of U-Net based on residual network

Download full size

Fig. 3. ResNet34 residual block diagram. (a) Residual block for feature extraction; (b) residual block for down-sampling

Download full size

Fig. 4. Training procedure. (a) Loss change curves versus the number of epochs; (b) metrics change curves versus the number of epochs

Download full size

Fig. 5. Examples of cloud detection. (a) Input images; (b) ground truth cloud mask; (c) results of proposed method; (d) results of U-Net; (e) results of Otsu method

Download full size

Layer name	Conv1	Conv2_x	Conv3_x	Conv4_x	Conv5_x
Output size	112×112	56×56	28×28	14×14	7×7
Informationof blocks	Conv 7×7,64	$\begin{matrix} \begin{matrix} Max_pool 3 \times 3 \\ [\begin{matrix} Conv 3 \times 3,64 \\ Conv 3 \times 3,64 \end{matrix}] \times 3 \end{matrix} \end{matrix}$	$\begin{matrix} [\begin{matrix} Conv 3 \times 3, 128 \\ Conv 3 \times 3, 128 \end{matrix}] \end{matrix}$ ×4	$\begin{matrix} [\begin{matrix} Conv 3 \times 3,256 \\ Conv 3 \times 3, 256 \end{matrix}] \end{matrix}$ ×6	$\begin{matrix} [\begin{matrix} Conv 3 \times 3, 512 \\ Conv 3 \times 3, 512 \end{matrix}] \end{matrix}$ ×3

Table 1. Structural parameters of ResNet34

Layer name	Input size	Information of blocks	Output size
Encoder1	224×224×3	$\begin{matrix} [\begin{matrix} Conv (7 \times 7) \\ Batc h Norm, ReLU \\ Max_pool (3 \times 3) \end{matrix}] \end{matrix}$	112×112×64
Encoder2	112×112×64	Conv2_x	56×56×128
Encoder3	56×56×128	Conv3_x	28×28×256
Encoder4	28×28×256	Conv4_x	14×14×512
Bridge	14×14×512	Conv5_x	7×7×1024
Decoder1	7×7×1024	$\begin{matrix} [\begin{matrix} Up ⁃ sampling \\ Concat \\ Conv (3 \times 3), ReLU \end{matrix}] \end{matrix}$	14×14×512
Decoder2	14×14×512	$\begin{matrix} [\begin{matrix} Up ⁃ sampling \\ Concat \\ Conv (3 \times 3), ReLU \end{matrix}] \end{matrix}$	28×28×256
Decoder3	28×28×256	$\begin{matrix} [\begin{matrix} Up ⁃ sampling \\ Concat \\ Conv (3 \times 3), ReLU \end{matrix}] \end{matrix}$	56×56×128
Decoder4	56×56×128	$\begin{matrix} [\begin{matrix} Up ⁃ sampling \\ Concat \\ Conv (3 \times 3), ReLU \end{matrix}] \end{matrix}$	112×112×64
Decoder5	112×112×64	$\begin{matrix} [\begin{matrix} Up ⁃ sampling \\ Concat \\ Conv (3 \times 3), ReLU \end{matrix}] \end{matrix}$	224×224×64
Output	224×224×64	Conv(1×1)	224×224×2

Table 2. Parameters of ResNet-based U-Net

Method	PA /%	mPA /%	mIoU /%	Inferencetime /s
Otsu	85.55	78.24	67.04	1.0
U-Net	90.04	84.53	76.03	5.4
Proposed method	93.33	93.43	83.88	6.4

Table 3. Results comparison of different cloud detection methods

Scene ID	Detection algorithm	PA/%	mPA/%	mIoU /%
LC80200462014005LGN00	Otsu	95.00	76.27	73.59
	U-Net	96.44	86.46	82.32
	Proposed	96.59	95.66	85.32
LC80210072014236LGN00	Otsu	88.03	80.99	66.62
	U-Net	93.36	95.87	80.49
	Proposed	87.82	92.49	70.10
LC80310202013223LGN00	Otsu	77.50	78.56	62.50
	U-Net	92.58	92.87	86.19
	Proposed	89.74	89.37	81.18
LC80290372013257LGN00	Otsu	83.29	70.28	60.85
	U-Net	76.28	57.81	45.41
	Proposed	97.56	94.28	92.57
LC81390292014135LGN00	Otsu	90.31	81.99	76.14
	U-Net	92.73	86.98	82.15
	Proposed	95.12	95.48	88.77
LC81590362014051LGN00	Otsu	79.69	76.20	63.12
	U-Net	83.34	80.47	69.21
	Proposed	93.51	92.52	87.35
LC81620432014072LGN00	Otsu	85.01	83.41	66.43
	U-Net	95.54	91.24	86.45
	Proposed	92.99	94.23	81.85

Table 4. Detail results of test set

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information