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

Table 3

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

Different architectures	Feature size	Training ratios
		40%	60%
Without fusion (CaffeNet(RGB))	4096	95.79 ± 1.37	96.87 ± 0.66
Without fusion (CaffeNet(saliency))	4096	93.21 ± 1.55	95.86 ± 0.50
Without fusion (VGG-Net-16(RGB))	4096	96.09 ± 0.56	96.64 ± 1.08
Without fusion (VGG-Net-16(saliency))	4096	93.75 ± 0.86	95.55 ± 0.89
Without fusion (GoogLeNet(RGB))	1024	93.77 ± 0.79	95.32 ± 1.92
Without fusion (GoogLeNet(saliency))	1024	91.22 ± 0.78	94.10 ± 1.19
Fusion strategy 1 (CaffeNet)	8192	96.78 ± 1.02	98.00 ± 0.59
Fusion strategy 2 (CaffeNet)	4096	97.74 ± 0.98	98.92 ± 0.52
Fusion strategy 1 (VGG-Net-16)	8192	97.28 ± 0.62	97.81 ± 0.87
Fusion strategy 2 (VGG-Net-16)	4096	98.23 ± 0.56	98.79 ± 0.99
Fusion strategy 1 (GoogLeNet)	2048	94.78 ± 0.77	96.34 ± 1.09
Fusion strategy 2 (GoogLeNet)	1024	95.72 ± 0.87	97.29 ± 1.20