Deep Adversarial Domain Adaptation Method for Cross-Domain Classification in High-Resolution Remote Sensing Images

Wenxiu Teng; Ni Wang; Taisheng Chen; Benlin Wang; Menglin Chen; Huihui Shi

doi:10.3788/LOP56.112801

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 11 >Page 112801 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 11, 112801 (2019)

Deep Adversarial Domain Adaptation Method for Cross-Domain Classification in High-Resolution Remote Sensing Images

Wenxiu Teng^1、**, Ni Wang^2、3、*, Taisheng Chen^2、3, Benlin Wang^2、3、4, Menglin Chen^2、3, and Huihui Shi³

Author Affiliations

¹ College of Forest, Nanjing Forestry University, Nanjing, Jiangsu 210037, China

² School of Geographic Information and Tourism, Chuzhou University, Chuzhou, Anhui 239000, China

³ Anhui Engineering Laboratory of Geographical Information Intelligent Sensor and Service, Chuzhou, Anhui 239000, China

⁴ School of Earth Sciences and Engineering, Hohai University, Nanjing, Jiangsu 210098, China

show less

DOI: 10.3788/LOP56.112801 Cite this Article Set citation alerts

Wenxiu Teng, Ni Wang, Taisheng Chen, Benlin Wang, Menglin Chen, Huihui Shi. Deep Adversarial Domain Adaptation Method for Cross-Domain Classification in High-Resolution Remote Sensing Images[J]. Laser & Optoelectronics Progress, 2019, 56(11): 112801 Copy Citation Text

show less

Framework of proposed method. (a) High spatial resolution remote sensing image dataset; (b) source deep convolutional neural network; (c) unsupervised adversarial domain adaptation; (d) remote sensing image scene classification

Fig. 1. Framework of proposed method. (a) High spatial resolution remote sensing image dataset; (b) source deep convolutional neural network; (c) unsupervised adversarial domain adaptation; (d) remote sensing image scene classification

Download full size

Fig. 2. Sample images of UC-Merced dataset

Download full size

Fig. 3. Sample images of WHU-RS dataset

Download full size

Fig. 4. Classification confusion matrix of UC-Merced dataset. (a) Classification accuracy of source domain; (b) classification accuracy of domain adaptation

Download full size

Fig. 5. Classification confusion matrix of WHU-RS dataset. (a) Classification accuracy of source only; (b) classification accuracy of domain adaptation

Download full size

Large class(six)	Subclass(seventy-one)
Construction land	city_building, container, storage_room, pipeline, town, baseball_diamond, basketball_court, golf_course, tennis_court, ground_track_field, church, commercial_area, industrial_area, mobile_home_park, palace, stadium, thermal_power_station, dense_ residential, medium_residential, sparse_residential, Square, Center, Park, Resort, School, Playground
Ultivated land	green_farmland, dry_farm, bare_land, circular_farmland, rectangular_ farmland, terrace
Transportation	airplane, airport_runway, avenue, highway, harbor, parkinglot, crossroads, bridge, airport, overpass, railway, railway_station, ship, roundabout
Water area	beach, dam, hirst, lakeshore, river, sea, stream, island, lake, sea_ice, wetland
Woodland	artificial_grassland, sparse_forest, forest, mangrove, sapling, river_protection_forest, shrubwood, chaparral, meadow
Other	desert, snow_mountain, mountain, sandbeach, cloud

Table 1. Dataset category for remote sensing image scene classification of source domain

Layer name	Layer type	Output size /(pixel×pixel×pixel)
Source /targetDCNN	Input	224×224×3
	Convolution×2	224×224×64
	Map pooling	112×112×64
	Convolution×2	112×112×128
	Map pooling	56×56×128
	Convolution×3	56×56×256
	Map pooling	28×28×256
	Convolution×3	28×28×512
	Map pooling	14×14×512
	Convolution×3	14×14×512
	Map pooling	7×7×512
	Fully connected	1×1×1024
	Softmax	1×1×18/1×1×21