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

Table 2

Comparison with the state-of-the-art methods on the UC-Merced dataset.

Methods	Training ratios
	50%	80%
SCK [31]	-	72.52
SPCK [32]	-	73.14
BoVW [33]	-	76.81
BoVW + SCK [31]	-	77.71
BRSP [34]	-	77.80
SIFT + SC [35]	-	81.67 1.23
SSEA [36]	-	82.72 1.18
MCMI [37]	-	88.20
OverFeat [38]	-	90.91 1.19
VLAD [39]	-	92.50
VLAT [39]	-	94.30
MS-CLBP + FV [40]	88.76 0.79	93.00 1.20
CaffeNet [41]	93.98 0.67	95.02 0.81
GoogLeNet [41]	92.70 0.60	94.31 0.89
VGG-VD-16 [41]	94.14 0.69	95.21 1.20
CNN-ELM [27]	-	95.62
salM^3LBP-CLM [42]	94.21 0.75	95.75 0.80
TEX-Net-LF [43]	95.89 0.37	96.62 0.49
Fusion by addition [44]	-	97.42 1.79
Ours	96.97 0.75	98.02 1.03