MSCANet: multiscale context information aggregation network for Tibetan Plateau lake extraction from remote sensing images

Figures & data

Figure 1. MSCANet network structure.

Figure 2. Schematic diagram of the attention module of the inverted bottleneck channel.

Figure 3. Depthwise convolution.

Figure 4. Pointwise convolution.

Figure 5. Inverted Residual Block.

Figure 6. Multi-core Pyramid Pooling Unit (MPPM) network architecture.

Figure 7. The Tibetan Plateau lake semantic dataset.

Table 1. Experimental results of the Tibetan Plateau lake semantic dataset.

	MSCANet	AMS-UNet	U-Net	ResNet	ResUnet	FCN_VGG16	MSCENEet	HR-Seg
Pre/%	94.77	92.06	91.30	92.13	91.09	95.11	95.09	90.55
Rec/%	95.86	94.37	92.49	93.49	91.71	91.11	94.30	91.93
F1/%	95.28	92.46	90.74	91.69	91.40	93.06	94.16	91.24
MIou/%	91.12	87.58	85.07	86.49	84.18	87.24	89.98	83.93

Figure 8. Lake localization for different segmentation models.

Figure 9. Frozen lake localization for different segmentation models.

Figure 10. Lake localization for different segmentation models.

Figure 11. Performance comparison of different models for small lakes extraction.

Table 2. Experimental results of the small lakes dataset.

		MSCANet	AMS-UNet	U-Net	ResNet	ResUNet	FCN_VGG16	MSCENet	HR-seg	NDWI	MNDWI
Image one	Pre Rec F1 Iou	88.98 96.65 92.66 86.32	86.28 71.82 78.39 64.46	93.47 78.68 85.44 74.58	92.59 88.38 90.43 82.54	91.28 71.44 80.15 66.87	80.37 93.03 86.24 75.81	89.49 95.28 92.29 85.69	92.56 60.46 73.14 57.65	92.49 40.84 56.57 39.53	93.91 45.46 61.26 44.15
Image Two	Pre Rec F1 Iou	94.39 93.66 94.02 88.72	95.66 90.37 92.94 86.81	90.34 36.30 51.79 34.94	96.50 68.72 80.27 67.05	94.14 47.07 62.76 45.73	98.11 79.96 88.11 78.75	88.18 96.70 92.24 85.60	95.57 72.67 82.56 70.30	96.43 78.06 86.28 75.86	96.76 42.71 59.26 42.11
Image three	Pre Rec F1 Iou	91.88 97.06 94.40 89.39	92.54 96.95 94.09 89.29	86.84 85.50 86.16 75.69	92.35 90.55 91.44 84.23	86.33 84.06 85.18 74.18	92.26 88.62 90.40 82.49	90.15 94.32 92.19 85.51	90.05 83.14 86.46 76.15	94.52 22.47 36.31 22.19	97.96 41.02 57.83 40.67
Avg. (120)	Pre Rec F1 Iou	93.12 95.19 94.14 88.93	91.39 94.60 92.97 86.88	91.94 75.47 82.90 70.73	92.36 88.99 90.65 82.78	92.47 75.96 83.41 71.25	93.49 91.17 92.32 85.70	90.45 92.91 91.67 84.77	92.79 78.20 84.87 73.36	95.49 81.85 86.31 79.05	97.25 75.73 82.55 74.19

Figure 12. Performance comparison of different models for frozen lakes extraction.

Table 3. Experimental results of the frozen lakes dataset.

		MSCANet	AMS-UNet	U-Net	ResNet	ResUNet	FCN_VGG16	MSCENet	HR-seg	NDWI	MNDWI
Image one	Pre Rec F1 Iou	92.28 95.76 93.99 88.66	87.95 94.89 91.29 83.97	95.15 93.86 94.50 89.58	92.18 95.04 93.59 87.95	92.40 94.93 93.65 88.05	96.35 90.48 93.32 87.48	86.74 93.98 90.21 82.17	98.35 93.40 95.81 91.96	84.88 97.23 90.64 82.88	84.40 99.49 91.32 84.03
Image *two	Pre Rec F1 Iou	98.05 99.86 98.95 97.92	98.17 96.46 97.31 94.76	98.84 95.02 96.89 93.97	98.37 98.76 98.56 97.17	98.39 93.53 95.90 92.12	98.28 93.88 96.03 92.36	99.07 93.45 96.18 92.64	99.06 86.00 92.07 85.31	97.19 22.39 36.40 22.25	99.52 94.48 96.93 94.05
Image three	Pre Rec F1 Iou	97.34 93.98 95.63 91.62	96.75 81.32 88.37 79.16	96.85 50.73 66.58 49.90	98.30 67.98 80.37 67.19	97.08 54.28 69.63 53.41	98.59 74.74 85.02 73.95	97.46 86.71 91.77 84.80	97.82 56.89 71.94 56.18	97.82 93.17 95.44 91.27	98.41 92.26 95.24 90.91
Avg. (123)	Pre Rec F1 Iou	95.74 93.32 94.51 89.56	94.95 93.45 94.19 88.91	96.46 87.84 91.95 85.02	95.84 91.11 93.42 87.53	95.77 89.63 92.60 86.09	96.19 89.42 92.68 86.32	95.09 93.90 94.09 89.69	91.18 90.77 93.87 88.39	97.40 81.34 86.65 79.70	98.24 81.47 87.43 80.12

Figure 13. Performance comparison of different models for background-disturbed lakes extraction.

Table 4. Experimental results of the background-disturbed lakes dataset.

		MSCANet	AMS-UNet	U-Net	ResNet	ResUNet	FCN_VGG16	MSCENet	HR-Seg	NDWI	MNDWI
Image one	Pre Rec F1 Iou	99.54 96.61 98.05 96.17	87.51 97.91 92.42 85.90	88.94 92.74 90.80 83.15	90.87 98.01 94.30 89.22	93.96 94.08 94.02 88.71	96.87 97.80 97.33 94.80	99.54 92.95 96.13 92.55	94.30 56.04 70.31 54.21	97.14 95.84 96.49 93.22	99.50 94.62 96.99 94.17
Image two	Pre Rec F1 Iou	91.82 97.84 94.74 90.00	87.95 94.89 91.29 83.97	91.99 91.44 91.72 84.70	90.83 95.43 93.07 87.04	92.08 93.39 92.73 86.45	90.56 96.99 93.66 88.08	88.94 88.45 88.70 79.69	93.91 84.40 88.90 80.03	88.17 95.59 91.73 84.73	91.59 97.33 94.38 89.35
Image three	Pre Rec F1 Iou	97.36 96.89 97.13 94.41	96.44 96.68 96.55 93.34	96.29 97.54 96.91 94.01	97.33 95.26 96.28 92.83	96.83 91.91 94.31 89.22	97.57 94.45 95.99 92.28	96.06 97.23 96.64 93.50	98.21 95.48 96.83 93.85	98.07 95.67 96.83 93.90	98.64 89.64 93.93 88.55
Avg. (150)	Pre Rec F1 Iou	93.96 94.59 94.27 89.10	91.38 95.89 93.58 87.92	93.65 90.58 92.09 85.33	92.20 94.04 93.11 87.17	93.04 90.63 91.82 84.96	92.84 93.43 93.14 87.14	91.82 92.72 92.26 85.51	94.98 86.08 90.31 82.28	95.67 77.40 83.10 75.13	97.74 80.85 86.62 79.48

Figure 14. Performance comparison of different models for freshwater lakes extraction.

Table 5. Experimental results of the freshwater lakes dataset.

		MSCANet	AMS-UNet	U-Net	ResNet	ResUNet	FCN_VGG16	MSCENet	HR-seg	NDWI	MNDWI
Image one	Pre Rec F1 Iou	98.78 97.02 97.89 95.87	99.74 84.12 91.27 83.94	97.82 26.48 41.67 26.32	99.29 78.40 87.62 77.96	99.54 66.75 79.91 66.54	96.52 75.50 84.72 73.50	98.85 95.25 97.01 94.20	99.53 88.43 93.65 88.06	99.34 96.77 98.04 96.15	99.71 89.40 94.27 89.16
Image two	Pre Rec F1 Iou	98.79 99.68 99.23 98.48	98.99 95.43 97.18 94.51	98.70 62.05 76.20 61.55	98.40 52.32 68.31 51.88	98.66 64.34 77.88 63.78	99.42 96.69 98.03 96.14	99.06 88.54 93.50 87.80	99.09 47.94 64.62 47.73	99.35 99.06 99.21 98.42	99.47 56.11 71.74 55.94
Image three	Pre Rec F1 Iou	96.89 97.42 97.16 94.47	96.73 93.28 94.97 90.43	97.36 80.24 87.98 78.53	97.66 70.39 81.81 69.22	96.52 80.10 84.55 77.85	94.61 97.74 96.15 92.58	97.57 80.54 88.24 78.95	98.96 51.85 68.05 51.57	97.64 97.66 97.65 95.41	97.87 71.76 82.81 70.66
Avg. (183)	Pre Rec F1 Iou	98.23 92.43 95.24 90.93	97.75 85.04 90.95 83.46	97.31 66.39 78.93 65.46	97.94 69.29 81.16 68.18	97.50 72.00 82.83 70.90	97.93 85.71 91.41 84.19	98.34 85.75 91.62 84.55	97.97 76.38 85.84 74.83	98.36 90.38 92.87 89.14	98.52 79.02 86.26 78.30

Figure 15. Performance comparison of different models for salt lakes extraction.

Table 6. Experimental results of the salt lakes dataset.

		MSCANet	AMS-UNet	U-Net	ResNet	ResUNet	FCN_VGG16	MSCENet	HR-seg	NDWI	MNDWI
Image one	Pre Rec F1 Iou	95.07 96.73 95.89 92.11	92.49 97.84 95.09 90.64	95.37 92.47 93.90 88.50	94.96 96.32 95.64 91.64	95.52 93.88 94.69 89.92	95.71 94.68 95.19 90.82	90.79 94.74 92.73 86.44	98.16 91.90 94.93 90.34	98.94 81.66 89.47 80.95	99.48 87.54 93.13 87.14
Image two	Pre Rec F1 Iou	96.87 93.17 94.98 90.45	75.59 91.92 82.96 70.88	95.38 91.08 93.18 87.23	93.69 89.47 91.53 84.38	95.63 81.51 88.01 78.58	93.06 88.30 90.62 82.84	78.47 94.93 85.92 75.31	93.59 89.32 91.41 84.18	96.12 88.33 92.06 85.29	98.59 87.89 92.93 86.79
Image three	Pre Rec F1 Iou	95.88 93.90 94.88 90.26	91.82 92.23 92.20 85.22	89.72 94.67 92.13 85.40	91.17 95.75 93.40 87.62	93.36 93.22 93.29 87.43	95.07 93.54 94.30 89.22	91.12 92.88 91.99 85.17	91.12 92.88 91.99 85.17	92.61 96.80 94.66 89.86	92.45 97.32 94.82 90.15
Avg. (96)	Pre Rec F1 Iou	96.35 94.63 95.48 91.38	94.55 95.11 94.83 90.20	95.82 88.92 92.24 85.66	95.58 93.33 94.44 89.54	96.25 86.66 91.20 83.77	95.21 93.46 94.33 89.29	94.37 94.02 94.19 88.99	95.97 84.06 89.62 81.00	97.76 91.04 93.76 89.17	98.74 88.25 92.61 87.20

Figure 16. Performance comparison of different models for Sentinel-2A dataset extraction.

Table 7. Experimental results of the Sentinel-2A dataset.

		MSCANet	AMS-UNet	U-Net	ResNet	ResUNet	FCN_VGG16	MSCENet	HR-seg
Region one	Pre Rec F1 Iou	99.26 98.89 99.08 98.17	99.32 94.52 96.86 93.92	99.44 81.79 89.76 81.42	99.76 92.75 96.12 92.54	99.09 88.15 93.30 87.44	99.88 91.67 95.60 91.57	99.66 95.66 97.62 95.35	99.51 88.48 93.67 88.10
Region two	Pre Rec F1 Iou	98.57 99.99 99.28 98.56	98.24 99.16 98.70 97.43	94.81 97.00 95.89 92.11	97.24 99.98 98.59 97.23	94.83 98.20 96.49 93.21	96.88 99.33 98.09 96.25	97.77 98.35 98.06 96.20	95.09 93.57 94.32 89.26
Region three	Pre Rec F1 Iou	96.47 95.88 96.17 92.62	96.60 90.30 93.34 87.51	82.38 87.51 84.86 73.71	97.54 94.71 96.11 92.51	93.49 89.86 91.64 84.57	67.40 91.57 77.64 63.46	98.24 96.32 97.27 94.69	98.38 76.65 86.17 76.70
Region four	Pre Rec F1 Iou	99.59 97.83 98.70 97.44	99.50 95.51 97.46 95.05	98.88 89.59 94.01 88.69	99.24 87.12 92.78 86.54	99.06 88.86 93.68 88.12	99.63 97.38 98.49 97.03	99.69 72.50 83.95 72.34	98.33 74.72 84.92 73.79
Region five	Pre Rec F1 Iou	99.39 98.89 99.14 98.30	99.63 97.30 98.45 96.95	97.80 94.43 96.09 92.47	98.57 93.72 96.08 92.40	98.04 93.85 95.90 92.12	98.93 98.40 98.67 97.37	99.24 99.28 99.26 98.53	96.34 93.54 94.92 90.33

Figure 17. Lake prediction results for different textures and spectral characteristics of the Tibetan Plateau.

Table 8. Experimental results of the Google Earth Dataset.

		MSCANet	AMS-UNet	U-Net	ResNet	ResUNet	FCN_VGG16	MSCENet	HR-seg
Region one	Pre Rec F1 Iou	99.44 89.40 94.15 88.95	99.64 89.35 93.98 88.70	99.48 90.62 94.10 90.19	99.47 90.26 94.04 88.83	99.44 91.27 94.12 90.80	99.53 91.02 94.08 90.63	99.81 88.07 93.81 86.92	99.90 90.10 93.74 90.01
Region two	Pre Rec F1 Iou	98.51 99.29 98.90 97.82	97.85 98.96 98.40 96.86	97.94 98.92 98.42 96.90	97.98 99.02 98.50 97.04	97.80 98.97 98.38 96.81	96.99 99.03 98.00 96.07	98.51 97.87 98.19 96.45	98.59 99.12 98.85 97.73
Region three	Pre Rec F1 Iou	98.80 99.41 99.10 98.22	98.44 99.21 98.98 98.66	99.41 98.53 98.97 97.96	98.51 98.11 98.81 97.63	99.52 98.77 99.15 98.33	99.46 98.72 99.00 98.20	99.61 97.02 98.79 96.64	99.07 99.16 99.02 98.25
Region four	Pre Rec F1 Iou	96.43 97.99 97.21 94.56	91.09 97.77 94.31 89.24	87.65 97.18 92.17 85.48	90.74 97.67 94.08 88.82	86.69 97.40 91.73 84.73	92.20 97.62 94.84 90.18	94.21 98.06 96.10 92.49	89.59 96.97 95.61 91.43
Region five	Pre Rec F1 Iou	95.10 99.37 97.19 94.54	89.19 99.46 94.04 88.75	83.94 99.60 91.10 83.66	84.95 99.77 91.73 84.78	82.01 99.20 90.04 81.88	87.42 99.70 93.16 87.19	95.41 98.38 96.87 93.93	84.38 99.46 91.30 84.00

Table 9. Ablation experiment results.

Model	The Tibetan Plateau lake semantic dataset
	Pre(%)	Recall(%)	F1(%)	MIou(%)
CANet	93.44	94.26	93.19	88.44
MSANet	92.87	93.63	93.25	87.42
MSIRNet	94.50	94.53	93.86	89.70
MSCNet	94.33	95.78	95.05	90.70
MSCANet	94.77	95.86	95.28	91.12

Data availability statement

The Tibetan Plateau lake semantic dataset is provided by National Cryosphere Desert Data Center. (http://www.ncdc.ac.cn). The code used in this study are available by contacting the corresponding author.