Fig. 1. Preprocessing of GF-2 multispectral images and panchromatic images

Fig. 2. Spectral curves of vegetation at the same point before and after radiometric calibration（In order to reduce the amount of data storage，the atmospheric correction results are expanded by 10 000 times）.

Fig. 3. Spectral curve of vegetation after atmospheric corrected（In order to reduce the amount of data storage，the atmospheric correction result is expanded by 10 000 times）

Fig. 4. Comparison before and after image fusion

Fig. 5. 512×512 size of the remote sensing image and their corresponding label maps（In the label map，red represents the water body，and black represents the background）

Fig. 6. 64 feature maps of 1×1 with global receptive fields

Fig. 7. Structure diagram of depthwise separable convolution

Fig. 8. Multi-scale fusion attention module

Fig. 9. Atrous convolution receptive field with atrous rate 2

Fig. 10. Feature maps of different sizes output by VGG16

Fig. 11. Feature maps of different sizes output after UNet skip connection

Fig. 12. Structure of A-MSFAM-UNet

Fig. 13. Loss curves of the A-MSFAM-UNet training process

Fig. 14. Comparison of segmentation results of different methods

Fig. 15. Comparison of segmentation results of different methods（In the segmentation image，the white is the water body，and the black is the background）.