Fig. 1. Whole framework of the proposed algorithm

Fig. 2. Images of several UCM classes. (a) Farmland; (b) airplane; (c) baseball diamond; (d) dense residential; (e) freeway; (f) harbor; (g) storage tanks; (h) tennis court; (i) overpass; (j) golf course

Fig. 3. Images of RSSCN7 classes. (a) Grass; (b) river lake; (c) industry; (d) field; (e) forest; (f) residential; (g) parking

Fig. 4. The t-SNE graphs of Xu and different Zu on GoogLeNet feature. (a) Xu; (b) Zu+I(λ=0, μ=0); (c) Zu+I(λ=1, μ=0); (d) Zu+T(λ=0, μ=1); (e) Zu+T(λ=1, μ=1)

Fig. 5. The t-SNE graphs of Xu and different Zu on VGGNet feature. (a) Xu; (b) Zu+I(λ=0, μ=0); (c) Zu+I(λ=1, μ=0); (d) Zu+T(λ=0, μ=1); (e) Zu+T(λ=1, μ=1)

Fig. 6. OA values of proposed algorithm under different λ and μ

Fig. 7. The ζsum/OA curves of algorithm under GoogLeNet feature

Fig. 8. The ζsum/OA curves of algorithm under VGGNet feature

Fig. 9. Accuracies of different classes of RSSCN7 produced by I type ZSC methods. (a) GoogLeNet; (b) VGGNet

Fig. 10. Accuracies of different classes of RSSCN7 produced by T type ZSC methods.(a) GoogLeNet; (b)VGGNet