Cloud Detection Based on Multi-Scale Dilation Convolutional Neural Network for ZY-3 Satellite Remote Sensing Imagery

Lin Gao; Weidong Song; Hai Tan; Yang Liu

doi:10.3788/AOS201939.0104002

Acta Optica Sinica, Volume. 39, Issue 1, 0104002(2019)

Cloud Detection Based on Multi-Scale Dilation Convolutional Neural Network for ZY-3 Satellite Remote Sensing Imagery

Lin Gao^1,2、*, Weidong Song^1、*, Hai Tan², and Yang Liu^1,2

Author Affiliations

¹ School of Mapping and Geographical Science, Liaoning Technical University, Fuxin, Liaoning 123000, China

² Satellite Surveying and Mapping Application Center, National Administration of Surveying, Mapping and Geoinformation, Beijing 100048, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(8)

Fig. 1. Convolution kernel with different dilation rates. (a) r=1; (b) r=6; (c) r=12; (d) r=18

Download full size

Fig. 2. Schematic of deep multiscale dilation fully convolutional neural network architecture

Download full size

Fig. 3. Visual analysis on cloud area for different pooling layers

Download full size

Fig. 4. Diagram of edge protection of training image

Download full size

Fig. 5. Comparison of cloud detection results at different areas using different algorithms. (a) Area covered by thick cloud; (b) area covered by middle-thick cloud; (c) area covered by thick and thin clouds; (d) area covered by thick cloud and haze with the a complex scene; (e) area covered by a large range of thick cloud

Download full size

Table 1. Multi-spectral image parameters of ZY-3 satellite

View table

Table 1. Multi-spectral image parameters of ZY-3 satellite

Index	Parameter
Resolution /m	5.8
Wavelength /nm	Band 1: 450-520; Band 2: 520-590
	Band 3: 630-690; Band 4: 770-890
Width /km	52
Single scene /km²	2704
Latitude range	upper left:30.5633N; upper right: 30.4678N
	lower left: 30.1186N; lower right: 30.0234N
Longitude range	upper left: 113.7162E; upper right: 114.2382E
	lower left: 113.6103E; lower right: 114.1299E

Table 2. Comparison of the accuracy of different network structures
View table
Table 2. Comparison of the accuracy of different network structures
Method Accuracy /% Training time /h
FCN-8S 86.9 36
Proposed 96.8 14

Table 3. Quantity evaluation parameters of different algorithms

View table

Table 3. Quantity evaluation parameters of different algorithms

Fig.	Detection algorithm	OA	D_OA	F1-measure	Kappa	D_Kappa
	Otsu	0.9444	-0.0458	0.9668	0.7990	-0.1687
Fig. 5(a)	FCN-8S	0.9661	-0.0241	0.9793	0.8867	-0.081
	Proposed	0.9902	—	0.9936	0.9677	—
	Otsu	0.8413	-0.0895	0.8412	0.6869	-0.1705
Fig. 5(b)	FCN-8S	0.9096	-0.0212	0.9150	0.8189	-0.0385
	Proposed	0.9308	—	0.9416	0.8574	—
	Otsu	0.8580	-0.1075	0.8837	0.7075	-0.2106
Fig. 5(c)	FCN-8S	0.8841	-0.0814	0.9076	0.7544	-0.1637
	Proposed	0.9655	—	0.9755	0.9181	—
	Otsu	0.9731	-0.0146	0.9855	0.8100	-0.0686
Fig. 5(d)	FCN-8S	0.9789	-0.0069	0.9886	0.8440	-0.0346
	Proposed	0.9858	—	0.9928	0.8786	—
	Otsu	0.7649	-0.2275	0.1277	0.0987	-0.6152
Fig. 5(e)	FCN-8S	0.9767	-0.0157	0.5386	0.5279	-0.0186
	Proposed	0.9924	—	0.7175	0.7139	—

Tools

Get Citation

Copy Citation Text

Lin Gao, Weidong Song, Hai Tan, Yang Liu. Cloud Detection Based on Multi-Scale Dilation Convolutional Neural Network for ZY-3 Satellite Remote Sensing Imagery[J]. Acta Optica Sinica, 2019, 39(1): 0104002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites