SiameseDenseU-Net-based Semantic Segmentation of Urban Remote Sensing Images

Table 2

Performances of the different models.

Models		Impervious surfaces	Building	Low vegetation	Tree	Car	Ave. F1	Overall. Acc
GT	HSN	87.57	92.20	75.03	84.44	75.16	82.88	84.92
	HSN + OI	88.01	92.37	75.83	84.86	76.50	83.51	85.38
	HSN + OI + WBP	88.00	92.34	75.92	84.86	75.95	83.41	85.39
	U-Net + CEloss	85.82	90.51	73.62	83.33	67.24	80.10	83.46
	U-Net + MFB_Focalloss	85.64	90.31	72.86	83.25	76.52	81.72	83.21
	DenseU-Net + CEloss	87.77	92.42	75.89	84.36	83.21	84.73	85.28
	DenseU-Net+ MFB_Focalloss	88.18	92.50	76.23	84.63	83.23	84.95	85.63
	SiameseDenseU-Net + CEloss	88.31	92.49	76.22	84.76	82.77	84.91	85.76
	SiameseDenseU-Net + MFB_Focalloss	88.93	93.48	76.08	85.03	84.15	85.53	86.20
erGT	HSN	90.89	94.51	78.83	87.84	81.87	86.79	88.32
	HSN + OI	91.32	94.66	79.73	88.30	83.60	87.52	88.79
	HSN + OI + WBP	91.34	94.67	79.83	88.31	83.59	87.55	88.82
	U-Net + CEloss	88.92	92.62	77.45	86.70	75.54	84.24	86.75
	U-Net + MFB_Focalloss	88.84	92.40	76.70	86.56	82.68	85.44	86.50
	DenseU-Net + CEloss	90.89	94.57	79.77	87.74	90.83	88.76	88.57
	DenseU-Net + MFB_Focalloss	91.30	94.64	80.17	87.99	90.96	89.01	88.92
	SiameseDenseU-Net + CEloss	91.40	94.59	80.22	88.09	90.49	88.96	89.04
	SiameseDenseU-Net + MFB_Focalloss	92.08	95.57	79.96	88.42	91.33	89.47	89.49

Note: bold font indicates the best results.