5.4 MODIS and the AO in flood forecasting
High interannual variability in timing and rate of peak snow cover depletion appear to modulate flooding occurrence, intensity and timing in eastern Kazakhstan. Specific flood reports for the study area are lacking, but figures from the Emergency Events Database (Guha-Sapir et al., 2018) and Global Active Archive of Large Flood Events (Dartmouth Flood Observatory, 2018) point to snowmelt-induced floods occurring throughout Kazakhstan in 2008, 2010, 2011, 2015 and 2017 (Fig.5). Comparison with peak PSCDR shows a general agreement, with the highest peak snowmelt rates in the 18 year series occurring in 2011 and 2017. In particular, severe flooding affecting the Uba basin in mid-April 2011 was reported on the official government news website (Dixinews, 2011), corresponding with the timing of the highest PSCDR in the Uba basin in the study period (Figs. 4 and 5). However, the third highest PSCDR in 2006 has no correspondence in flooding, although this might be due to a bias towards better reporting in more recent years (both databases are based on local news). Broad agreement is also seen with respect to timing, with floods occurring relatively late in 2011, 2015 and 2017 (end of March –beginning of spring) and early in 2008 (20th February, Dartmouth Flood Observatory, 2018), corresponding with late and early values of DSCD (Fig. 7). Flood occurrence in 2010 however is reported as early as March, in spite of very late peak snow cover depletion found in our study. 2010 was a particularly cold winter with record snowfalls (Cohen et al, 2010), and temperature increase above the snow melting point at low elevations might have led to abundant snow melt and subsequent flooding.
In spite of this general agreement, quantitative assessment of the predictive ability of snow cover depletion maps based on MODIS and of the possible use of the AO in flood forecasting in the Upper Irtysh catchment is complicated by a lack of in situ observations in the study area. To improve the ground-based hydrological monitoring network in Kazakhstan, runoff should ideally be measured at stations, located in the smaller, non-transboundary basins upstream of Bukhtarma reservoir. In principle, the timing of flooding could could also be investigated via remote sensing, by detecting flooded areas through optical (Revilla-Romero et al., 2015) or SAR (Brown et al., 2016) satellite images, and algorithms have also been devised to estimate discharge from MODIS images based on the reflectance difference between water and land pixels in near infrared bands (Tarpanelli et al., 2017). Validation of these approaches in sparsely gauged catchments however remains a non trivial task. Since our approach is based on readily available data, in areas where ground based observations do exist, it could be more easily validated, providing a means to evaluate MODIS DPSCD and PSCDR products with respect to water management strategies and possibly implement long-term forecasts based on the AO.