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

Table 5

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

Different architectures	Feature size	Training ratios
		20%	50%
Without fusion (CaffeNet(RGB))	4096	87.57 ± 0.32	90.22 ± 0.42
Without fusion (CaffeNet(saliency))	4096	84.45 ± 0.28	87.21 ± 0.48
Without fusion (VGG-Net-16(RGB))	4096	87.24 ± 0.18	90.60 ± 0.31
Without fusion (VGG-Net-16(saliency))	4096	84.25 ± 0.11	87.62 ± 0.56
Without fusion (GoogLeNet(RGB))	1024	84.18 ± 0.53	87.15 ± 0.69
Without fusion (GoogLeNet(saliency))	1024	81.12 ± 0.55	84.28 ± 0.67
Fusion strategy 1 (CaffeNet)	8192	92.26 ± 0.52	94.36 ± 0.29
Fusion strategy 2 (CaffeNet)	4096	92.32 ± 0.41	94.42 ± 0.33
Fusion strategy 1 (VGG-Net-16)	8192	92.04 ± 0.28	94.53 ± 0.18
Fusion strategy 2 (VGG-Net-16)	4096	92.11 ± 0.31	94.58 ± 0.25
Fusion strategy 1 (GoogLeNet)	2048	89.15 ± 0.45	91.25 ± 0.59
Fusion strategy 2 (GoogLeNet)	1024	89.21 ± 0.39	91.31 ± 0.49