land cover land use change, understanding climate change impacts, as well as for tracking water in lakes and reservoirs. Since 1978, a series of AVHRR instruments has been carried aboard the National Oceanic and Atmospheric Administration’s (NOAA) family of polar orbiting platforms (POES). Compared to Landsat, which has a spatial resolution of 15–60 m, and a temporal resolution of 16 days, the 1.1 km pixel size of AVHRR is much coarser. The AVHRR has daily temporal coverage, however, which enables more cloud free images to be collected. Since the year 2000, MODIS sensors on board Terra (launched in 2000) and Aqua (launched in 2002) have been collecting VIS/NIR/SWIR images twice a day at 250–500 m resolution (which is a major improvement over AVHRR). In general, a low‐resolution sensor such as AVHRR can be used for estimating the area variations of a lake with an area (at capacity) of larger than 200 km2, while a high‐resolution sensor such as Landsat can monitor a lake smaller than 1 km2. Over the very large lakes, the area estimated from sensors at different resolutions would be similar in most cases. When the reservoir has long shorelines surrounded by seasonal vegetation, however, the area estimated by AVHRR or MODIS will tend to have an exacerbated seasonality, as the effective reflectance of the mixed pixels along the shoreline will be affected by the changing vegetation reflectance (Gao et al., 2012). It is worth noting that these are only a few examples of the many similar sensors that have existed over the years (e.g., Medium Resolution Imaging Spectrometer, Visible Infrared Imaging Radiometer Suite, etc.).
Based on the various properties of VIS, NIR, and SWIR, many approaches have been developed to extract surface water data from satellite imagery. In Gao (2015), the algorithms were summarized into three categories: threshold‐based, image‐classification‐based, and multiple‐step hybrid approaches. These algorithms typically use single or multiple variables as their classification criteria, ranging from the reflectance of a particular water sensitive band (e.g., SWIR) to indices from multiband reflectance data. A detailed comparison of commonly used Normalized Difference Water Indices (NDWIs) is available from Ji et al. (2009). While thresholding approaches were commonly used to extract water area in earlier studies (Bryant, 1999; Harris & Mason, 1989; Mason et al., 1992), supervised and unsupervised classification techniques are now more commonly used, as they produce more reliable and consistent results in general (Lu et al., 2011; Maulik & Saha, 2010; Seeber et al., 2010). Multistep hybrid approaches usually estimate optimal area values through methods such as combining thresholding and supervised classification (Wang et al., 2014), and then conducting postclassification image enhancement (Zhang et al., 2014).
With the evolutional development of the Google Earth Engine, a new cloud‐computing platform for hosting global scale Earth observation data, the mapping of global surface water is no longer plagued by the need to download and process large volumes of imagery data. Pekel et al. (2016) generated a first of its kind global surface water data set from Landsat (monthly at a 30 m resolution) from 1984 to 2015. Although this data set has a very high classification accuracy, the reservoir surface area values are often underestimated due to contaminations from clouds, cloud shadows, and terrain shadows. Zhao & Gao (2018) developed a novel algorithm that automatically corrected these contaminated classifications in the Pekel et al. (2016) data set. This led to a time series of area values for 6817 global reservoirs (with an integrated capacity of 6099 km3). As shown in Figure 3.4, these area estimations agree well with observed elevation/storage values.
Here we introduce a new hydrological drought index, the Reservoir Area Fraction Index (RAFI), using remotely sensed water area information calculated with equation 3.2:
where A RS is the remotely sensed reservoir area, and A 95%_max is 95% of the maximum water area from 1984 to 2015. The use of a 95 percentile maximum remotely sensed area will allow us to exclude the possibility that the flooded area is larger than the area at capacity. Figure 3.5 shows that, except under flooding conditions, there is a strong agreement between the Landsat based RAFI and U.S. Geological Survey observed elevation values at Lake Whitney (Texas). By developing drought indices from the remotely sensed reservoir area, true global monitoring capability can be achieved. Furthermore, the decades of available Landsat data can offer insightful information for long‐term planning.
3.3. ADOPTING REMOTELY SENSED RESERVOIR DATA TO SUPPORT DROUGHT MODELING APPLICATIONS
Since the early 1990s, a number of large‐scale hydrological models (LHMs) have been developed to address the global water scarcity issue (Bierkens, 2015). During the early stage, most LHMs focused on solving the issues related to water balance and the routing of runoff water without considering reservoir flow regulations (Vörösmarty et al., 1989). In the early 2000s, more LHMs started to consider the human impacts on the hydrological cycle. Haddeland et al. (2014) found that the impacts from reservoir management are often as large, or larger than, those from global warming. This was based on results from seven global LHMs, H08 (Hanasaki et al., 2008), LPJmL (Bondeau et al., 2007), MPI‐HM (Stacke & Hagemann, 2012), PCR‐GLOBWB (Wada et al., 2011), VIC (Liang et al., 1994), WaterGAP (Döll et al., 2001), and WBMplus (Wisser et al., 2008). However, because reservoir operation rules are usually not shared, they have virtually always been represented in a rather simplistic manner within these global models. This has prevented these models from supporting purposes related to the monitoring of hydrological droughts.
Figure 3.4 Comparison of remotely sensed surface area with observed storage/elevation for nine reservoirs. The two numbers in parentheses indicate the values of R2 between observed storages (or elevations) and water areas before and after enhancement.
(Source: From Zhao, G., & Gao, H. (2018). Automatic Correction of Contaminated Images for Assessment of Reservoir Surface Area Dynamics. Geophysical Research Letters. 45(12), 6092–6099. © 2018, John Wiley & Sons.)
In contrast, the physically based, distributed LHMs, which do not have a reservoir component, have been well adopted for monitoring agricultural drought at the continental and global scales. Examples include (but are not limited to) the North American Land Data Assimilation System (NLDAS) Drought Monitor (http://www.emc.ncep.noaa.gov/mmb/nldas/drought/), the Princeton United States and Global Drought Monitor (http://hydrology.princeton.edu/forecast/current.php), and the Global Integrated Drought Monitoring and Prediction System (GIDMaPS) (http://drought.eng.uci.edu/). Although some of these drought monitors also use modeled streamflows as indicators, they are often biased because they do not consider the effects of reservoir flow regulations in their routing scheme. The most reliable streamflow based monitor is the U.S. Geological Survey Water Watch (https://waterwatch.usgs.gov), which collects national scale observed gauge data. However, it is impossible to set up monitors of this kind over most of the world due to the lack of gauge data (Kugler & De Groeve, 2007). Indeed, as pointed out by Wada et al. (2017) and other studies (Fekete et al., 2015; Lawford et al., 2013), there is a critical need for comprehensive data for purposes of calibrating and evaluating hydrological models over continental to global scales.
Remotely sensed reservoir data can be used to support drought‐modeling applications in two ways. First, the models can be calibrated and validated using remotely sensed
