Assessment of the Performance of Image Fusion for the Mapping of Snow Using Euclidean Distance

The assessment of the performance of multi-resolution image fusion, or image sharpening methods, is difficult. In the context of binary classification of snow targets in mountainous terrain, fusion methods were applied to help achieve more accurate mapping. To quantify objectively the gain of information that can be attributed to an increase in spatial resolution, we investigate the Mean Euclidean Distance (MED) between the snowline obtained from the classification, and a reference snowline (or a ground truth line), as a relevant indicator to measure both the discrepancy between datasets at different spatial resolutions, and the accuracy of the mapping process. First, a theoretical approach based on aggregating detailed reference images from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) showed that the MED has a linear relationship with the pixel size that makes it suitable to assess images of different resolutions. Secondly, we tested the MED to snow maps obtained ‘with’ or ‘without’ applying a fusion method to the MODerate Resolution Imaging Spectroradiometer (MODIS). We demonstrated that the MED identified a significant value added, in terms of mapping accuracy, which can be attributed to the fusion process. When the fusion method was applied to four different images, the MED overall decreased by more than 30%. Finally, such a ‘feature based’ quality indicator can also be interpreted as a statistical assessment of the planimetric accuracy of natural pattern outlines.