A Two-Stream Deep Fusion Framework for High-Resolution Aerial Scene Classification

Table 7

Classification performance of the proposed method on the NWPU-RESISC45 dataset using different feature extractors and fusion strategies.

Different architectures	Feature size	Training ratios
		10%	20%
Without fusion (CaffeNet(RGB))	4096	77.34 ± 0.32	80.54 ± 0.22
Without fusion (CaffeNet(saliency))	4096	75.06 ± 0.51	78.20 ± 0.33
Without fusion (VGG-Net-16(RGB))	4096	77.10 ± 0.14	80.45 ± 0.31
Without fusion (VGG-Net-16(saliency))	4096	74.94 ± 0.23	78.09 ± 0.48
Without fusion (GoogLeNet(RGB))	1024	76.87 ± 0.45	79.12 ± 0.23
Without fusion (GoogLeNet(saliency))	1024	74.67 ± 0.52	77.04 ± 0.19
Fusion strategy 1 (CaffeNet)	8192	80.15 ± 0.23	83.08 ± 0.21
Fusion strategy 2 (CaffeNet)	4096	80.22 ± 0.22	83.16 ± 0.18
Fusion strategy 1 (VGG-Net-16)	8192	79.95 ± 0.12	82.96 ± 0.19
Fusion strategy 2 (VGG-Net-16)	4096	80.03 ± 0.19	83.02 ± 0.14
Fusion strategy 1 (GoogLeNet)	2048	79.69 ± 0.47	81.46 ± 0.22
Fusion strategy 2 (GoogLeNet)	1024	79.75 ± 0.41	81.52 ± 0.28