Terrestrial surface water bodies such as lakes, reservoirs, and rivers cover approximately 3 % of the global land mass. They play a crucial role in the global hydrological cycle, biodiversity conservation, and climate process (Chahine, 1992; Tranvik et al., 2009). Detailed knowledge on surface water distribution, and its seasonal, inter-annual, and long-term variability can serve as an important source for water management (Cole et al., 2007), ecosystem assessment, and biodiversity conservation (Turak et al., 2017). Remote-sensing data have increasingly been used to monitor surface water changes, and powerful methods and tools have been developed for analyzing Earth observation data. However, existing approaches for monitoring the surface water extent are limited either in geographic scope, temporal extent of the record, or with respect to the temporal frequency of observations.

At the global scale, several static datasets exist that provide information on the spatial extent of water bodies and wetlands. For example, the Global Lakes and Wetlands Database (GLWD: Lehner and Doll, 2004) was based on historical maps and has a spatial resolution of 30 arcsec (approx. 1 km). Carroll et al. (2009) combined the Shuttle Radar Topography Missions (SRTM) Water Body Data (SWBD) with 250 m Moderate Resolution Imaging Spectrometer (MODIS) reflectance data to produce a global static map of surface water for circa 2000–2002. Global Landsat-based static surface water datasets include the 3 arcsec (∼90 m) Water Body Map (G3WBM: Yamazaki et al., 2015) and the Global Land Cover Facility (GLCF) inland surface water dataset at a 30 m resolution for 2000 (Feng et al., 2015).

Even though static water maps are adequate for some applications there is an increasing demand for information on the spatiotemporal variability of inland water bodies and their long-term evolution (Belward, 2016). Dynamic mapping and monitoring of the surface water extent have been explored using optical sensors featuring fine (10–30 m) to medium (250–500 m) spatial resolutions. At fine spatial resolution, several studies have recently presented interesting results on long-term variability of surface water with the entire Landsat archive at regional (Halabisky et al., 2016; Heimhuber et al., 2016), continental (Mueller et al., 2016), and global scales (Donchyts et al., 2016; Pekel et al., 2016). The European Commission Joint Research Centre's (JRC) Global Surface Water (GSW) dataset (Pekel et al., 2016) quantifies changes in global surface water over the past 32 years with a monthly time interval. This product allows for the analysis of surface water dynamics over long time periods at fine spatial resolution, but only provides information on monthly changes in surface water. Moreover, the Landsat archive also contains data gaps and temporal discontinuities depending on the geographical location (Pekel et al., 2016). This is due to both the limited number of acquisitions during specific time intervals, and the location- and time-dependent persistency of cloud cover. These data gaps affect the accuracy of the seasonality information (Yamazaki and Trigg, 2016). To better represent water bodies with short hydroperiods and short-duration flooding, it is critical to account for such gaps when monitoring surface water. In recent years, the revisit time of fine-resolution sensors has increased (e.g., Sentinel-2 has offered a 5 d repeat since March 2017: Du et al., 2016). However, these data cannot yet be used to create long-term (>10 year) consistent time series at short time intervals.

Moderate resolution imagery derived from satellite sensors such as MODIS provides daily observations over long time-spans and as such has the potential to construct long-term and dense time series of surface water over large regions. Many studies have explored the use of MODIS in mapping water body dynamics at regional to continental scales (Kaptue et al., 2013; Pekel et al., 2014; Sharma et al., 2015) using binary classification methods. At the global scale, Khandelwal et al. (2017) used MODIS multispectral data to map the global extent and temporal variations of 94 large reservoirs at a 500 m resolution and at an 8 d interval from 2000 to 2015. The recent Global Climate Observing System (GCOS) report states that essential climate variables (ECVs) need to be established for water extent and lake ice cover products, ideally with daily temporal resolution (Belward, 2016). To address this requirement, the first daily global dataset of inland water bodies at a 250 m spatial resolution from 2013 to 2015 was developed by Klein et al. (2017). This work advanced surface water mapping using remote sensing, due to its dense temporal resolution, and enhanced our understanding of rapid water changes caused by extreme climate change and human activities. However, like other surface water mapping efforts based on binary classification methods (e.g., Khandelwal et al., 2017; Mohammadi et al., 2017), this product omits lakes and narrow rivers that only cover a portion of a MODIS resolution cell.

To overcome this limitation and incorporate small water bodies, several researchers have attempted to predict sub-pixel surface water estimates of MODIS by providing the water fraction in each pixel using techniques like linear spectral mixture modeling (e.g., Hope et al., 1999; Li et al., 2013; Olthof et al., 2015) and machine learning (e.g., Li et al., 2018; Rover et al., 2010; Sun et al., 2012) for small areas. However, the utility and efficiency of these methods have rarely been explored for the estimation of the surface water fraction for larger areas. In our previous work (Li et al., 2018), we explored the use of rule-based regression models over two small areas on the Iberian Peninsula and concluded that a single global regression model can provide accurate surface water estimates across areas with different environmental conditions as long as it is fed with training data that comprise these various conditions. Consequently, we concluded that this approach has the potential to be applied over much larger areas. Therefore, the aim of this paper is to explore the utility and efficiency of a rule-based regression model for the estimation of the surface water fraction for the Mediterranean region, and to develop a new surface water fraction dataset for the Mediterranean region using fine temporal resolution MODIS data as input for the effective assessment of seasonal, intra-annual, and long-term surface water dynamics inclusive of small water bodies. Our specific objectives are as follows:

to develop an approach for the estimation of the surface water fraction for the Mediterranean region at a fine temporal resolution from MODIS data;

We loosely defined the Mediterranean in this study as the region that is contained within 10 MODIS grid tiles, which together cover all coastal areas of the Mediterranean and the Black Sea, including a significant portion of their inland areas (Fig. 1). This boundary is defined based on the combination of (1) the definition of the Mediterranean region by the Mediterranean Wetland Observatory (MWO) project, i.e., 27 Mediterranean countries are included by MWO; (2) the inclusion of areas with a large amount of Ramsar wetlands; and (3) the exclusion of southern parts of north African countries (i.e., Morocco, Algeria, Libya, and Egypt) that comprise few water bodies according to the maximum water extent over 32 years from JRC's GSW product.

The study region covers 13 climate zones as defined by Peel et al. (2007) (Fig. 1). Numerous water bodies of different types are found in this region, including large coastal lagoons, fresh, brackish or salt marshes, riverine forests and reed beds, flood plains and wet meadows, mountainous lakes and surrounding wetlands, salted lakes, temporary marshes, and streams (Costa et al., 1996). Our study area accounts for 25 % of the world's Ramsar sites that contain a great ecological, social, and economic value, especially as they provide habitat, reproduction, and migration stopover sites for numerous bird species (Galewski, 2012). A good number of the water bodies and wetlands in the region are small, shallow, and highly variable between seasons and years due to weather effects and human activities (Costa et al., 1996).

Many Mediterranean water resources are degraded mainly due to urbanization, agricultural reclamation, increasing water use for irrigation, and hydraulic works such as dams, dikes, river channeling, and drainage and irrigation networks (Batalla et al., 2004). A number of projects and programs have performed monitoring of surface water and wetlands in the Mediterranean region, such as MWO (http://medwet.org/, last access: 10 July 2019) and the GlobWetland initiative (http://webgis.jena-optronik.de/, last access: 10 July 2019), which highlighted the importance of protecting Mediterranean water resources. However, these projects either performed wetland mapping for a few moments in time (e.g., GlobWetland only covered 1975, 1990, and 2005), or were limited to specific water bodies and wetlands instead of the whole landscape. Although surface water dynamics in the Mediterranean can be analyzed at fine spatial resolution with JRC's GSW, it has large spatial and temporal gaps. Figure 2 shows the percentage of pixels with valid observations in JRC's GSW monthly water history dataset for each month between January 2000 and October 2015 calculated over the entire Mediterranean area. The figure illustrates that no valid observation exists for December in the years 2000–2015 in Mediterranean areas according to the GSW monthly water history map. In addition, less than 10 % of the Mediterranean area has observations for January.

Table 1 summarizes all datasets used in this study. They include a number of sources used to derive model input variables, training data for building the model, validation data for model accuracy assessment, and other existing surface water products against which we compared our products. Details are provided in the following sections.

We estimated the surface water fraction for the Mediterranean region from MODIS data, and improved on previous efforts to estimate surface water fraction from medium-resolution imagery (MODIS or similar). The prediction accuracy of our model (R2=0.91, RMSE =11.41 %, and MAE =6.39 %) is higher than the R2 of 0.625 reported by Weiss and Crabtree (2011), who used a linear regression model, and the R2 of 0.7 reported by Guerschmann et al. (2011), using a logistic regression model. This research successfully expanded our previous work (Li et al., 2018) by upscaling it from a relatively small region to the whole Mediterranean while retaining a similar high accuracy (both achieved an R2 of 0.91). This attributes to the feasibility and robustness of Cubist regression modeling with respect to dealing with different environmental conditions when training data are collected across a wide geography resulting in varying spectral characteristics.

Secondly, we generated surface water fraction maps at high temporal frequency, which is an advantage over existing fine-resolution datasets. Our MODIS-derived surface water fraction product accurately displays the spatiotemporal variability of surface water. The comparison with the GSW 30 m product and water level data (Figs. 7–9) reveal that our product can efficiently monitor the seasonal, intra-annual, and long-term surface water dynamics with good spatial and temporal accuracy. For example, it complements the GSW by allowing for the detection of inundation and recession processes over short time periods and for the better representation of seasonality changes and temporal trends over longer periods. The MODIS surface water fraction can also accurately detect the spatial distribution of surface water inclusive of small water bodies (less than one MODIS pixel) and narrow rivers, which are missing in other coarse-resolution products using binary classification method such as MOD44W (Salomon et al., 2004). Metrics derived from the MODIS surface water fraction and GSW time series can reflect different facets of surface water dynamics. The accuracy of some metrics, such as seasonality, relies on the number of valid observations and the temporal interval. Landsat-derived metrics might be problematic in areas with large temporal gaps caused by persistent cloud cover, as shown in this study (Fig. 8) and previously by Pekel et al. (2016) and Klein et al. (2017). Although cloud coverage also limits MODIS observations, the probability of obtaining cloud-free observations is higher (Table 6) due to the daily acquisitions and consequent temporal compositing possibilities. With these advantages, we expect that our MODIS surface water fraction product could fill in important information on surface water for areas and time periods for which cloud-free Landsat acquisitions are few or non-existent.

Our MODIS-derived surface water fraction product also has limitations. Firstly, it is designed to detect only open surface water; therefore, it may not effectively capture water bodies covered by dense vegetation, such as swamps, lakes with considerable coverage of aquatic vegetation, and inundated dense forests. Secondly, the MODIS surface water fraction product overestimates small surface water fractions of less than 20 % (Table 5), which has also been found in previous studies on surface water fraction mapping (Li et al., 2018; Parrens et al., 2017). This overestimation might be attributed to the mixed spectral response of pixels with different land cover types, as already demonstrated by many studies (e.g., Guerschmann et al., 2011; Klein et al., 2017). Further work could consider reassigning pixels with less than 20 % surface water to 0 % water fraction for locations where water is never present according to the GSW maximum water extent. Thirdly, the auxiliary layers utilized for the identification of potential areas of commission errors also appear to have some limitations. For example, some urban areas might be not mapped in MCD12Q1, and areas of cloud cover and cloud shadow might not be completely removed from the MCD43A4 product. The importance of these limitations may diminish as the quality of these auxiliary layers improves or dynamic datasets rather than static layers are incorporated (e.g., Global Human Settlement Layer: Pesaresi et al., 2016). Fourthly, the MCD43A4 product also suffers from many missing values, especially in regions with large amounts of precipitation, aerosol concentrations, or snow and ice coverage (Klein et al., 2017). Future research should focus on combining other moderate-resolution data (e.g., MOD09) for areas with missing data to ensure a gap-free reconstruction of inland water development for the past and future.

This work can be further scaled up over much larger regions and for shorter (e.g., daily) time intervals. A high-quality training dataset is crucial for the effective application of the rule-base regression model and the collection of such training data is time consuming (Sun et al., 2012). In this paper, time series of the GSW monthly water history dataset proved to be an efficient basis for building a reliable training dataset. Considering that GSW is globally available, we are confident that our approach can be scaled to monitor the surface water fraction globally with MODIS data. Although the surface water fraction maps were produced with an 8 d time step, the model developed in this paper was actually trained using daily MODIS data and could be directly applied to that temporal resolution. The resulting daily surface water fraction maps could be of high interest for ecological and hydrological research. The recent GCOS report requires water extent and lake ice cover with a daily temporal resolution and a 20 m and 300 m spatial resolution, respectively (Belward, 2016). Our approach makes this requirement for the coarser of the two resolutions within reach.

Our MODIS surface water fraction dataset may benefit a large number of applications. For example, it could be used as a monitoring tool for analyzing hydrologic extremes such as floods and droughts, detecting abnormal changes of wetland hydrology, capturing short-duration events, identifying newly formed and disappearing water bodies, and estimating global water loss. The MODIS surface water fraction dataset may help to improve the calibration and validation of hydrological models. For example, the water area can help to estimate a series of hydrological parameters such as water discharge (Huang et al., 2018) and water volume (Busker et al., 2019; Cael et al., 2017; Duan and Bastiaanssen, 2013; Tong et al., 2016). This would be particularly useful for areas where in situ measurements are sparse or inaccessible. Closely monitoring hydrological variability is important for understanding how climate change, meteorological variability, and human activities affect the dynamics of surface water and human livelihoods (Tulbure and Broich, 2019; Zhang et al., 2019). It may also provide new insights for understanding how surface water dynamics further influence climate. For example, lake expansion and the creation of new dams can alter local and regional precipitation patterns (Ekhtiari et al., 2017; Hossain et al., 2009; Mohamed Degu et al., 2011). Similarly, our long-term records and derived metrics have the potential to contribute to the management and conservation of biodiversity and other ecosystem services associated with terrestrial surface water and wetlands.

We derived an 8 d, 500 m resolution surface water fraction product over the Mediterranean for the period from 2000 to 2017 by applying a global Cubist regression tree model to MODIS and SRTM data. We validated the results with JRC's Landsat-derived GSW dataset, which resulted in a high overall accuracy (R2=0.91, RMSE =11.41 %, and MAE =6.39 %). The MODIS-derived surface water fraction showed a good spatial and temporal correspondence with JRC's GSW. Comparison with satellite altimetry and in situ water level data for selected lakes demonstrated the ability of MODIS surface water fraction to effectively monitor seasonal and inter-annual changes in surface water extent. Our dataset provides a consistent, long-term record (18 years) of 8 d water fraction dynamics for the Mediterranean region, and complements fine spatial resolution surface water products, especially in regions where such products have long temporal and spatial data gaps due to both the limited number of acquisitions and persistent cloud cover. Our approach is also promising for monitoring surface water fraction at the global scale and at a daily interval.