5.1 Data and methods uncertainty
The main uncertainties in the analysis of snow cover variability and the association between snow cover variables and atmospheric indices are related to 1) the use of 8-day snow cover area instead of daily products, which also implies the derived DSCD and DPSCD are in fact 8-day products, and 2) cloud-cover interpolation. Drawing links between these variables and flood hazards in East Kazakhstan also requires an assumption that snow covered area can be used to represent SWE and snow cover depletion can be a proxy for snow cover melt and river discharge.
The choice of the 8-day composite product was made to reduce the impact of cloud cover on the study area. While the main misclassification in the daily MODIS product occurs between clouds and snow, and can propagate to the 8-day product (Hall & Riggs, 2007), the accuracy of the composite product might actually be higher than that of the daily products (93% on average, Hall & Riggs, 2007) in areas of high cloud cover, also leading to a higher correlation of the composite product with streamflow (Zhou et al., 2005). In a study conducted in northern China, Wang et al. (2009) found an accuracy of 94% for snow mapping for the composite product at snow depths greater than 4 cm, decreasing to 39% for lower depths and patchy snow cover. These are however likely to represent a small amount of SWE and therefore have a small impact on the assessment of water availability. The use of a cloud-interpolation scheme can also help reduce the cloud-cover related uncertainty (Dietz et al., 2011; Dong & Menzel, 2016), although our algorithm might introduce biases by enhancing elevation dependency (see S1, point 3) and in case of snow melt and subsequent snowfall (see S1, point 4); a further improvement on cloud cover interpolation accuracy is the use of in situ temperature and precipitation data (Dong & Menzel, 2016), although these are not often readily available in remote areas. Small uncertainties in snow cover variability might also stem from data gap filling in the original MODIS product in 2001 and 2002 at lower and higher elevations, respectively. Higher elevations in our study area might be affected by late spring snow falls after snow melt, which will not be recorded as DSCD, calculated as the first week of snow disappearance. However, these events are expected to provide a low contribution to the SWE.
The association between snow covered area, SWE and stream discharge has been demonstrated in several studies using the daily or 8-day MODIS composite product and other SCA datasets. Delbart et al. (2015) found an uncertainty of 15% in predicting water discharge from SCA between April and September in four Andean catchments. Further still, Gurung et al. (2017) found high correlations (> 0.78) at elevations between 4000 and 6000 m a.s.l. in the Gandaki basin in the Hindu Kush Himalayas. Using the composite product, Tong et al. (2009) identified a correlation of 0.84 between cloud-filtered snow cover extent and streamflow in the Quesnel river basin, Canada.
In this study, the relationship between snow covered area and SWE cannot be assessed due to the lack of in situ observations in our study area (Mashtayeva et al., 2016). The regulated nature of the Irtysh River, i.e. the presence of reservoirs and channels in China and Kazakhstan also causes dampening of peak flow and decreased runoff variability downstream (Huang et al., 2012; 2014), preventing attempts at correlating snow covered depletion rates and timing and discharge.