*4.3. Tests of Different Attention Modules*

The purpose of the water attention module is to provide the weight information of each pixel that belongs to the glacial lake. Currently, there are many simple but effective water indexes that can extract lake areas, such as NDWI, modified normalized difference water index (MNDWI) [39], and enhanced water index (EWI) [40]. To test whether these water indexes could locate a glacial lake area accurately and be adept at computing the water attention, they were incorporated into our attention module to obtain the pixel weight; then, their ability and importance with regard to mapping glacial lakes were measured. Here, MNDWI and EWI were calculated according to the following formulas:

$$\text{MNDWI} = \frac{\rho\_{\mathcal{G}^{reen}} - \rho\_{SWIR1}}{\rho\_{\mathcal{G}^{reen}} + \rho\_{SWIR1}} \tag{8}$$

$$\text{EWI} = \frac{\rho\_{\text{C}} - \rho\_{NIR} - \rho\_{SWIR2}}{\rho\_{\text{C}} + \rho\_{NIR} + \rho\_{SWIR2}} \tag{9}$$

where *ρgreen*, *ρC*, *ρNIR*, *ρSW IR*1, and *ρSW IR*<sup>2</sup> represent the TOA reflectance values in the green, cirrus, *NIR*, *SWIR*1, and *SWIR*<sup>2</sup> bands measured by the Landsat-8 OLI sensor, respectively.

According to the analysis in Section 4.2, we chose the Attn + ISeg + ResNet-152 structure and used GAN-GL-D as our evaluation data. The accuracy statistical results using different attention modules are listed in Table 4. Using the water index alone achieved low accuracies for mapping glacial lakes, in particular, Recall and IoU. This means that without convolution operations, the water index can misclassify objects when pixels have feature values similar to glacial lakes. Lake areas extracted by NDWI had fewer commission errors and exhibited the highest mapping accuracy when coupled with convolution operations.


**Table 4.** Accuracy evaluation of glacial lake mapping using different attention modules.

Figure 5 shows the visual evaluation of the image weight for the glacial lakes under various environmental conditions using the different water attention modules. Obviously, the use of the water index provided a high weight not only to glacial lakes, but also to melting glaciers and mountain shadows (denoted by the white ellipses in the first and third rows). With added convolution operations in the water attention module, effects from these interferences can be largely avoided. The second row in Figure 5 shows that the weight obtained by EWI is very conservative because its attention tends to the interior of a glacial lake. MNDWI obtained relatively extreme estimates, with attention tending to the

exterior of a lake. Only the NDWI-derived attention was uniform and close to the glacial lake boundary.

**Figure 5.** Weight results using different water attention modules. Input data are from Landsat-8 OLI images (first column, false color composites of bands 7/5/2), covering glacial lakes of various environmental components. Melting glaciers (white ellipses in the first row) and mountain shadows (white ellipses in the third row) also showed high weights using the water index alone.
