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

Table 1

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

Different architectures	Feature size	Training ratios
		50%	80%
Without fusion (CaffeNet(RGB))	4096	94.60 0.63	95.69 0.91
Without fusion (CaffeNet(saliency))	4096	92.62 0.74	94.04 0.88
Without fusion (VGG-Net-16(RGB))	4096	94.77 0.73	95.91 1.41
Without fusion (VGG-Net-16(saliency))	4096	92.82 0.91	94.31 0.99
Without fusion (GoogLeNet(RGB))	1024	93.31 0.71	94.99 0.78
Without fusion (GoogLeNet(saliency))	1024	91.32 0.98	93.30 0.55
Fusion strategy 1 (CaffeNet)	8192	95.79 0.52	96.83 0.91
Fusion strategy 2 (CaffeNet)	4096	96.74 0.49	97.80 0.88
Fusion strategy 1 (VGG-Net-16)	8192	96.02 0.77	97.05 1.00
Fusion strategy 2 (VGG-Net-16)	4096	96.97 0.75	98.02 1.03
Fusion strategy 1 (GoogLeNet)	2048	94.46 0.60	96.17 0.90
Fusion strategy 2 (GoogLeNet)	1024	95.41 0.58	97.12 0.96