Tile-Based Semisupervised Classification of Large-Scale VHR Remote Sensing Images

Table 3

Classification results for the Riyadh dataset using one hidden layer of size 256.

	OA	AA	Time (s)	Z score	Class 1	Class 2	Class 3	Class 4	Class 5
VGG-m
NN	87.96	81.98	0.22	—	99.09	71.49	86.26	96.53	56.55
NN + RW	88.07	82.21	0.05	−0.49	99.15	71.37	87.27	96.56	56.67
Spatial graph in the hidden layer	81.52	71.48	24.00	6.25	81.90	74.57	78.79	94.33	27.82
Spectral graph in the hidden layer	93.12	88.98	30.36	−5.30	99.22	85.03	91.01	98.61	71.04
NN + spectral graph + RW	93.18	89.07	0.05	−5.38	99.22	85.09	91.52	98.64	70.92
GoogleNet
NN	89.03	88.63	0.37	—	94.51	85.89	82.22	89.48	91.03
NN + RW	89.43	89.05	0.08	−1.06	94.51	85.43	83.64	90.09	91.61
Spatial graph in the hidden layer	80.97	75.54	23.97	5.76	96.21	77.77	70.61	85.50	47.59
Spectral graph in the hidden layer	93.45	91.03	55.17	−4.80	97.91	90.34	87.98	96.26	82.64
NN + spectral graph + RW	93.84	91.40	0.09	−5.20	97.91	90.40	88.18	96.83	83.68
ResNet
NN	89.58	86.33	0.35	—	88.30	81.09	87.78	95.40	79.08
NN + RW	89.72	86.44	0.07	−0.50	89.22	81.26	86.97	95.46	79.31
Spatial graph in the hidden layer	73.64	65.76	39.26	8.61	99.68	30.91	90.20	87.19	20.81
Spectral graph in the hidden layer	95.13	93.56	45.94	−5.86	98.76	89.54	96.36	97.50	85.63
NN + spectral graph + RW	95.29	93.80	0.07	−6.00	99.09	89.43	96.47	97.61	86.44
Very-deep-16
NN	90.97	88.09	0.63	—	95.36	75.60	94.85	96.81	77.82
NN + RW	93.13	91.84	0.11	−1.71	97.39	82.29	95.96	96.22	87.36
Spatial graph in the hidden layer	79.73	67.40	69.42	6.74	80.65	62.29	83.64	97.19	13.22
Spectral graph in the hidden layer	94.72	92.21	91.27	−4.72	98.00	88.85	95.12	98.29	80.79
NN + spectral graph + RW	95.26	92.91	0.11	−5.40	98.50	90.06	94.45	98.55	82.99