Dynamic Water Surface Detection Algorithm Applied on PROBA-V Multispectral Data

Abstract

Water body detection worldwide using spaceborne remote sensing is a challenging task. A global scale multi-temporal and multi-spectral image analysis method for water body detection was developed. The PROBA-V microsatellite has been fully operational since December 2013 and delivers daily near-global synthesis with a spatial resolution of 1 km and 333 m. The Red, Near-InfRared (NIR) and Short Wave InfRared (SWIR) bands of the atmospherically corrected 10-day synthesis images are first Hue, Saturation and Value (HSV) color transformed and subsequently used in a decision tree classification for water body detection. To minimize commission errors four additional data layers are used: the Normalized Difference Vegetation Index (NDVI), Water Body Potential Mask (WBPM), Permanent Glacier Mask (PGM) and Volcanic Soil Mask (VSM). Threshold values on the hue and value bands, expressed by a parabolic function, are used to detect the water bodies. Beside the water bodies layer, a quality layer, based on the water bodies occurrences, is available in the output product. The performance of the Water Bodies Detection Algorithm (WBDA) was assessed using Landsat 8 scenes over 15 regions selected worldwide. A mean Commission Error (CE) of 1.5% was obtained while a mean Omission Error (OE) of 15.4% was obtained for minimum Water Surface Ratio (WSR) = 0.5 and drops to 9.8% for minimum WSR = 0.6. Here, WSR is defined as the fraction of the PROBA-V pixel covered by water as derived from high spatial resolution images, e.g., Landsat 8. Both the CE = 1.5% and OE = 9.8% (WSR = 0.6) fall within the user requirements of 15%. The WBDA is fully operational in the Copernicus Global Land Service and products are freely available.

1. Introduction

Inland water surface mapping is important to budget freshwater supply for human and animals [1,2,3], and agriculture [4,5,6], but also to monitor ecological issues and to perform ecosystem management [7,8,9,10]. However, water body detection worldwide using spaceborne remote sensing is a challenging task because many objects on the Earth’s surface have similar spectral properties (within the spectral range of the applied sensor) as water bodies. Nevertheless, several spaceborne missions have been used to address this. The Moderate Resolution Imaging Spectrometer (MODIS) sensor was used to create the 250 m land-water mask (MOD44W) for the year 2000 [11]. Recently, MODIS data were used to assess global inland water body dynamics on a daily basis [12]. The SPOT VEGETATION (VGT) sensor was first used to create the dynamic Small Water Bodies (SWB) monitoring product for the African continent [13]. The same sensor is used to construct the global water bodies products for the whole archive, years 1999 to 2014, based on the work presented in this paper and becoming available by the end of 2016. The thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) sensors (Landsat 5 and Landsat 7, respectively), which allow high-resolution mapping of small water bodies, were used for water quality assessment [14] and water bodies extraction [15]. The latter were also used for mapping inland surface water at global scale for the years 2000 and 2010 resulting in the Global Land 30-water 2000 and Global Land 30-water 2010 datasets, respectively [16]. The Global Land Survey (GLS) Landsat data collection was used to produce a global, 30-m-resolution inland surface water body dataset (GIW) for circa-2000 [17], as well as to produce the Global 3 arc-second Water Body Map (G3WBM) [18] and to study Earth’s surface water change over the past 30 years [19]. Progress was made with Landsat 8, containing the nine-band Operational Land Imager (OLI) sensor and the Thermal InfraRed Sensor (TIRS) featuring two spectral bands. The cirrus band (band 9) allows for a better cloud screening while the coastal band might improve water detection [20]. Recently, all Landsat 8 acquisitions between June 2013 and June 2014 were processed using the Google Earth Engine platform to map water surfaces at global scale [21]. Although these sensors have a high spatial resolution (30 m), their revisit time is long (16 days), which makes near real-time water body mapping at global scale impossible.

Medium resolution missions, like PROBA-V, with a near-daily revisit time are more suited for near real-time water body mapping at global scale. The PROBA-V microsatellite was launched in May 2013 as a successor for the SPOT-VEGETATION missions. The instrument has been fully operational since December 2013 and delivers daily near-global synthesis with a spatial resolution of 1 km and 333 m. The high revisit time and the four spectral bands (Blue, Red, NIR and SWIR) make this instrument ideal for climate impact assessments, agricultural monitoring, food security estimates and surface water resources management.

Several techniques are available for water surface extraction. Supervised classification methods make use of ground reference data [22], while unsupervised classification methods first search for suitable endmembers [22,23]. Linear spectral unmixing uses endmembers to unravel the image spectra and, in this case, decide on the fraction of water within each pixel spectrum [24]. Spectral indices are the most frequently used. Well known is the normalized difference water index (NDWI) using the reflectance of the green and NIR bands [25]. Bands thresholding makes use of thresholds that are applied directly on the reflectance bands, on derived spectral indices or on transformed bands [1,26,27]. An overview of global-scale water body products developed in recent years using different techniques is presented by Yamazaki [18].

The objective of the research presented was to develop a global, multi-temporal and multi-spectral image analysis method for near real-time dynamic water body detection using PROBA-V 10-day composite images and which is based on the approach developed by Pekel [28]. This algorithm, which applies bands thresholding, was previously developed for MODIS and, after extensive validation, proved to be very successful [28]. The final output product (i.e., “Area of Water Bodies detected worldwide”) identifies the pixels covered by water on a global scale. The areas of water bodies are understood here with respect to the instrument resolution, i.e., surfaces more or less covered by water with a size of one PROBA-V pixel, about 1 km² [29].

2. Materials and Methods

2.1. Overview

As shown by the general overview in Figure 1, the Water Body Detection Algorithm (WDBA) uses five different sets of input data, which are processed in four different steps. Firstly, the PROBA-V daily Top-Of-Canopy synthesis products (S1-TOC) are assembled into a 10-day mean composite (MC10). During the mean compositing process, the MC10 Status Map (MC10-SM) is constructed, which, in turn, is used as input for the water body detection algorithm. Secondly, the SWIR, NIR and Red bands are transformed to HUE, SATURATION and VALUE using a RGB to HSV color transformation, a technique often used in image processing. Thirdly, the application of specific threshold values on HUE and VALUE applied per pixel allows water body detection. To minimize commission errors, four additional data layers are taken into account, i.e., the MC10 Status Map, Water Body Potential Mask (WBPM), Permanent Glacier Mask (PGM) and the Volcanic Soils Mask (VSM). Finally, to obtain qualitative information about the detected water bodies, an occurrence is calculated which is provided in the output product as a quality layer. The input data and the different steps of the algorithm are described in more detail in the subsequent paragraphs.

2.2. Input Data

2.2.1. Daily Top-Of-Canopy Synthesis Product (S1-TOC)

The PROBA-V instrument ensures a daily coverage between Lat. 35°N and 75°N, and 35°S and 56°S, and a full coverage every two days at the equator. The S1-TOC product is provided at 1 km spatial resolution. Surface reflectance is available for four spectral bands: Blue, 0.447–0.493 µm, Red, 0.610–0.690 µm; NIR, 0.770–0.893 µm; and SWIR, 1.570–1.650 µm. The atmospheric correction is performed using SMAC 4.0 [30]. Standard input data layers includes Normalized Difference Vegetation Index (NDVI), geometric viewing and illumination conditions, reference to date and time of observations, status map (containing identification of snow, ice, clouds, land/sea for every pixel) and four reflectance bands. More info can be found in the Products User Manual [31].

2.2.2. MC10 Status Map (MC10-SM)

At the final stage of the mean compositing step, a status map is created. The status map indicates which observation type was used in the mean compositing step, i.e., un-obscured, cloud or snow. It also has a flag that indicates if the observation was done above land or sea. Finally, it has for each reflectance band an additional quality flag that is set to 0 in case no compositing is done. This is caused when there was no valid observation to be used in the 10-day period. The water bodies detection process only takes un-obscured land pixels into account.

2.2.3. Water Body Potential Mask (WBPM)

Pixels located in mountainous areas often have lowered reflectance values due to shadow or dark vegetation and are as such confused with water bodies because their VALUE drops below the threshold level. Moreover, water bodies cannot exist or appear for pixels located in hilly terrain having steep slopes and a large height difference with their neighboring pixels. To minimize the commission errors in hilly terrain a Water Bodies Potential Mask (WBPM), derived from the Global Land Survey Digital Elevation Model (GLSDEM) [32] (data collected in the early 2000s), is used as an extra input in the water bodies detection process. The GLSDEM has a 90 m spatial and 1 m vertical resolution. Compared to the PROBA-V 1 km product the spatial resolution of the GLSDEM is much higher. To prepare the GLSDEM for use in water body detection, flat areas were considered and extracted as a WBPM. This mask was constructed in three steps.

• First the pixels having the lowest elevation in their local neighborhood are located, i.e., a pixel is indicated as a lowest point when its elevation is lower than or equal to its eight neighbors.
• Next, the detected lowest points are filtered and expanded depending on the topography in order to obtain a 90 m spatial resolution potential water body mask. For each detected lowest point, an imaginary water level is raised in steps of 1 m till the maximum rise of 5 m or the flooding condition is reached. Flooding in this context means detection of a neighboring pixel having an elevation that is lower than the rise level. As long as the edge of the potential water body is not flooded, its area is extended according to the raised level, i.e., all neighboring pixels having the additional elevation are added to the potential water body area. This is schematically shown in a two dimensional representation in Figure 2. The initial detected lowest point contains two pixels. Rising the water level 1 m will expand the potential water body with an additional 5 pixels. Because no flooding occurs the water level is increased again with 1 m. Now, an additional 3 pixels are added to the potential water body. Further rising the water level will start flooding the water body because the first pixel elevation right to the expanded water body is lower. If the size of the final expanded water body is less than nine pixels, it is removed from further processing and therefore not considered as a potential water body. Otherwise, the pixels initially detected as lowest point and surrounded by eight neighbors of equal height are marked as level-1. The other initially detected lowest points and the expanded pixels are marked as level-2. The two levels are used in the last step when resizing the potential water body map to the PROBA -V spatial resolution to obtain the final WBPM.
• In the last step, the 90 m spatial resolution potential water body mask is resampled to the PROBA-V 1 km spatial resolution. For each pixel in the georeferenced PROBA-V image, the corresponding pixels in the georeferenced potential water body mask are located. The 1 km resized pixel is indicated as a potential water body when at least one of the corresponding 90 m pixels is labeled as level-1 or when at least nine of the corresponding 90 m pixels are labeled as level-2.

An example of the final obtained WBPM, for the Rift Valley test area in Ethiopia (

Table A1. The Ethiopian Rift Valley test area used for fixed threshold definition, the upper left corner coordinate and number of samples (ns) and lines (nl) are given.

Area	Dekad	Upper Left Coordinate		PROBA-V
		Lat	Lon	ns	nl
Rift Valley	20131201	8.96429722	37.08032222	400	400

in Appendix A), is shown in Figure 3. A profile according to the red line on the images is shown in the elevation plot. Seven potential water bodies are located along the transect and are indicated by the blue shaded boxes. Two of those potential water bodies are real lakes, Lake Abijato and Lake Langano, and are indicated in Figure 3a.

2.2.4. Permanent Glacier Mask (PGM)

To avoid commission errors due to confusion with permanent glaciers, which often have spectral properties similar to water bodies, a permanent glacier mask was constructed. The most recent permanent glacier data were downloaded from the National Snow and Ice Data Centre [33]. The data were provided as a shape file, which was subsequently rasterized to the PROBA-V pixel size to obtain the permanent glacier mask at global scale.

2.2.5. Volcanic Soil Mask (VSM)

Locations comprised of dark volcanic soils which are spread over wider flat areas are easily confused with water bodies because their pixel VALUE drops below the threshold level. To account for these commission errors, a volcanic soil mask was made. A list of volcanoes worldwide was obtained from the Smithsonian Institution, National Museum of Natural History, Global Volcanism Program. The Holocene Volcano List was used as a reference for the construction of a volcanic soil mask. Volcanic soils could be easily located on Google Earth using the information from the Holocene Volcano List. These areas were manually delineated on Google Earth and were subsequently exported and converted to shape files from which a volcanic soil mask was made by rasterizing the shape files to the PROBA-V pixel size at global scale.

2.3. Mean Compositing

In the first processing step, temporal mean compositing over a 10-day window is performed to obtain (to the degree possible) a cloud free image. The MC10 composites (SWIR, NIR, Red, and Blue) are obtained using the mean compositing method [34] applied on the PROBA-V S1 daily reflectance values over a 10-day window. The mean compositing algorithm improves the radiometric quality of the temporal synthesis by averaging the reflectances and therefore reduces the random component of the noise. The mean composite is calculated, pixel by pixel, across the four spectral band simultaneously. For each pixel location, the input data over a time period of 10 days are handled sequentially. Only valid reflectances of a given pixel for a given date in the 10-day period are considered in the computation. These valid reflectances are determined using the actual reflectance value. The PROBA-V S1 input images define a “No data” reflectance value for each spectral band. All four spectral bands must have a valid reflectance value otherwise the given day does not contribute to the mean composition for that pixel location. To achieve consistency with the SPOT-VGT sensor, and enable the creation of a large time-series, a spectral correction is applied on the different spectral bands before calculating the mean reflectance values. The mean compositing algorithm always tries to return the most optimal result. To achieve this, it only composites the most optimal daily observations. For each day in the 10-day period, a pixel’s observation class is first determined as depicted in Figure 4.

Un-obscured land pixels are preferred before pixels that are classified as snow or cloudy, with cloudy pixels having the least priority. Input pixels not marked as land are not processed. To compute the mean value of all reflectance bands and the SZA, the algorithm keeps track of the reflectance bands and SZA sum of the corresponding pixel values and the number of pixels contributing to these sums, for a given observation class. If a more preferred observation class is detected for a pixel location, the sum and number of observation of the less preferred class are discarded and a new sum and number of observations is calculated, containing only pixel values of the new class. Once all days in the compositing window are computed, the reflectance band sum and SZA sum are divided by the number of observations to obtain the mean reflectance values and an averaged SZA.

2.4. Water Bodies Detection by Decision Tree Classification

An essential step before the actual water bodies detection takes place is the transformation of the Red, NIR and SWIR bands to the HSV (HUE, SATURATION and VALUE) color space as described by Pekel [28]. The detection of water bodies involves three main steps: (i) invalid data filtering; (ii) incompatibility masking; and (iii) water body detection. The decisions made in these three steps were implemented in a Decision Tree Classifier (DTC). Once water bodies are detected, the occurrence of the detection is estimated. An overview of the water body detection algorithm is shown in Figure 5.

2.4.1. Invalid Data Filtering

A low sun elevation causes elongated shadows, which are observed in the remote sensing imagery as darkened Earth surfaces. These might be detected as water bodies when the VALUE drops below the threshold level. A check on the MC10’s averaged Solar Zenith Angle (SZA) prohibits the detection of invalid WBs, i.e., when a MC10 pixel has a SZA larger than 65° no WB detection is done and the pixel is classified as “No Data”.

The next step in the WBDA is filtering the invalid data comprising snow, cloud and cloud shadow, using the information from the MC10 Status Map. Each image pixel has a bit array assigned reflecting its status, pixels indicated as cloud or snow are not considered in the WB detection algorithm. Cloud shadow on the Earth’s surface often has similar spectral properties as WBs and is as such regularly classified as WB. Therefore, to avoid confusion due to cloud shadow in the immediate neighborhood of the initial defined cloud pixel, an additional dilate filtering is performed. In the case a pixel is indicated as cloud, its twelve neighboring pixels (the circular neighborhood with radius 2 pixels) are indicated as cloud as well and are therefore not taken into account for WB detection.

2.4.2. Incompatibility Masking

In addition to glaciers, dark soils and shadows, vegetated surfaces are also often a cause of confusion with WBs. Therefore, the NDVI value, computed as Equation (1) from the Red (ρRed) and NIR (ρNIR) reflectances of MC10, is used to identify the vegetated areas.

$$NDVI=\frac{\rho NIR-\rho Red}{\rho NIR+\rho Red},$$

(1)

Pixels having an NDVI value higher or equal to 0.32 and having a WBPM pixel set (locating lowland areas where holding a WB is possible) are considered to be “Lowland Vegetation”. Some WBs have a high NDVI value (due to algae load or adjacency effect of the surrounding vegetation) and could as such wrongly be classified as “Lowland Vegetation”. To avoid these omission errors an additional check on VALUE is added, this threshold level is denoted as NV. If the VALUE of a pixel which is designated as “Lowland Vegetation”, is less or equal to 0.11 (a value empirically defined) it is classified as “Water”. These thresholds were empirically defined over the Rift Valley test area. The Rift Valley area is unique in the sense that it contains some larger lakes with a wide variety of reflectance values, some are bright blue, others are dark blue while others are almost black (see Figure 3). Besides, these lakes are surrounded with different shades of green vegetation (especially after the rain season). This situation easily allows observing how different thresholds levels influence the detection of WBs.

2.4.3. Water Body Detection

Threshold values on the HUE and VALUE bands are used for detecting WBs worldwide. The thresholds were iteratively and empirically defined by decision tree classifications using different settings for the thresholds levels. First, initial fixed thresholds were obtained by observing the HUE and VALUE for the lakes in the Ethiopian Rift Valley test region shown in Figure 3a. The initial fixed threshold for HUE and VALUE was set to 100 and 0.13, respectively. Subsequently, the threshold levels were iteratively adapted in different steps (VALUE = 0.13, 0.14; HUE = 100, 95; NDVI = 0.32, 0.42; and NV = 0.11, 0.13) while introducing an innovative refinement using quadratic functions, which were empirically defined. The performance of the various cases were evaluated and the thresholds were calibrated over 19 test regions worldwide (

Table A2. Landsat 8 scenes of the 19 test regions selected worldwide were used to refine the threshold levels. As much as possible, cloud free scenes were selected. The corresponding PROBA-V dekad and the upper left corner coordinate are given with the corresponding number of samples (ns) and lines (nl).

Area	Landsat 8 Scene	Dekad	Upper Left Coordinate		PROBA-V
			Lat	Lon	ns	nl
Argentina 1	LC82270862014111LGN00	20140421	−36.40625000	−63.94196429	299	239
Argentina 2	LC82310932014011LGN00	20140111	−46.29910714	−73.83482143	376	257
Australia	LC81110782014131LGN00	20140511	−24.94196429	118.49553571	267	236
Bangladesh	LC81380442014112LGN00	20140421	24.17410714	87.74553571	244	240
Bolivia	LC82330692013310LGN00	20131101	−11.95089286	−66.85267857	232	239
Brazil	LC82220742014028LGN00	20140121	−19.17410714	−51.55803571	248	237
Canada	LC80180272014151LGN00	20140521	48.53125000	−80.06696429	332	246
China	LC81210382014121LGN00	20140501	32.79910714	116.05803571	271	240
Congo	LC81740662014140LGN00	20140521	−7.62053571	25.22767857	235	238
India	LC81460422014072LGN00	20140311	27.04910714	76.05803571	249	240
Mali	LC81950502013332LGN00	20131121	15.51339286	−2.25446429	234	239
Mexico	LC80270422014118LGN00	20140421	27.04017857	−100.12053571	260	238
Poland	LC81870222014087LGN00	20140401	55.60267857	21.60267857	393	250
Rusia	LC81770212014129LGN00	20140501	56.99553571	37.60267857	426	248
South Africa	LC81690802014121LGN00	20140501	−27.77232143	28.12946429	277	246
Spain	LC82010332014121LGN00	20140501	39.95089286	−5.66517857	306	239
Turkey	LC81730332013306LGN00	20131101	39.95982143	37.65625000	299	240
USA	LC80430342014134LGN00	20140511	38.54017857	−121.87946429	281	242
Vietnam	LC81240512014062LGN00	20140301	14.08482143	107.11160714	241	242

of Appendix A) according to the Water Surface Ratio (WSR) and the metrics Commission Error (CE) and Omission Error (OE) as defined by Equations (C1)–(C3) in Appendix C. The choice of the final threshold levels is a balance between omission and commission errors: the best settings in terms of fixed thresholds are VALUE = 0.14, HUE = 100, NDVI = 0.32, and NV = 0.11 while the best quadratic functions (shown by the red lines in Figure 6) are defined as:

• The left part (for the HUE value = 0 till 34) can be written as:

  $$Valu{e}_L=\frac{{\left(34-Hu{e}_L\right)}^2\times Scal{e}_L}{2}+Offse{t}_L,$$

  (2)

  $$Hu{e}_L=\left[0\dots 34\right],$$

  (3)

  $$Scal{e}_L=\frac{0.41}{{34}^2},$$

  (4)

  $$Offse{t}_L=0.14,$$

  (5)
• The right part (for the HUE value = 34 till 100) is a bit more complex because it is based on a parabola rotated over 0.2° for which the rotated coordinates can be written as:

  $$x=Hue\times \cos \left(\alpha \right)-Hu{e}^2\times sin\left(\alpha \right),$$

  (6)

  $$y=Hue\times sin\left(\alpha \right)+Hu{e}^2\times \cos \left(\alpha \right),$$

  (7)
  This equation can be solved using the standard form:

  $$x=\frac{-b\pm \sqrt{b^2-4ac}}{2a},$$

  (8)

From which, only one solution is used to calculate the corresponding VALUE threshold:

$$x=\frac{-\left(-\mathit{cos}\left(\alpha \right)\right)-\sqrt{{\left(-cos\left(\alpha \right)\right)}^2-4\times sin\left(\alpha \right)\times Hu{e}_R}}{2\times sin\left(\alpha \right)},$$

(9)

$$Valu{e}_R=\frac{\left(x\times sin\left(\alpha \right)+x^2\times cos\left(\alpha \right)\right)\times Scal{e}_R}{2}+Offse{t}_R,$$

(10)

with

$$\alpha =0.2°,$$

(11)

$$Hu{e}_R=\left[34\dots 360\right]-28.5,$$

(12)

$$Scal{e}_L=\frac{1}{95000},$$

(13)

$$Offse{t}_R=0.14,$$

(14)

The division by two in both Equation (2) and Equation (10) is done to lower the parabolic curve in order to reduce the commission errors. The value 28.5 in Equation (12) was empirically defined, and is used to broaden the parabolic curve (higher values will broaden the curve). The value 95,000 in the scale factor was empirically deined, and defines the height of the parabolic curve (higher values will lower the curve).

The 2D scatter plot in Figure 6 gives a visual representation of the threshold levels. The fixed threshold levels (H_min = 100, V_max = 0.14) are indicated by the blue lines and the refined threshold levels are indicated by the red parabolic lines. The plot shows the reference WBs for the Vietnam test area indicated by the colored plus symbols and this for three ranges of Water Surface Ratio (WSR), i.e., blue for WSR = 0.95 to 1, cyan for WSR = 0.8 to 0.95 and magenta for WSR = 0.7 to 0.8. The other WSR ranges are not indicated as they would overrule the plot. The non-WB pixels, which are denoted as land, are indicated by the dark grey dot symbols. The refined thresholds allow the detection of WBs outside the fixed threshold levels as indicated by the blue arrow. Although the refined thresholds are tuned to detect as much WBs as possible, not all WBs can be detected because they are intermixed with land pixels as indicated by the red arrow. The 2D scatter plots of the eight first test areas are shown in Appendix B.

2.5. Water Bodies Occurrence Estimation

To qualify the occurrence of the detected WBs, a Water Body Occurrence (WBO) is calculated for each detected WB pixel in each dekad. This measure gives an idea about the permanency or seasonality of the detected WBs. To obtain a WBO map for each dekad, per pixel statistics are calculated by temporal-sequential processing of the available dekads. To get an idea of the WB occurrence, the observation period over which the statistics are calculated has to be long enough. An observation period over two years requires 72 dekads. However, for computational reasons, only 64 dekads are observed; that is, information that can be stored in one “long word” (64 bits). If more than 64cloud free MC10 composite observations are available, only the last 63 cloud free observations and the actual observation are used for calculating the statistics. The per-pixel temporal-sequential statistics are calculated over the available observation dekads (max 64) and can be summarized as:

• Total number of temporal cloud free observations (ntObs);
• Total number of temporal WB detections (ntWBs); and
• Maximum number of continuous temporal WB detections (mctWBs).

From these statistics, the frequency of detected WBs can be calculated as:

$$WBf=\frac{ntWBs}{ntObs}\times 100(\%),$$

(15)

As shown in Figure 7, the calculated WB frequency (WBf) and the maximum number of continuous temporal WB detections (mctWBs) can be visualized in a 2D scatter plot, which, in turn, was used for defining the occurrence threshold levels. The different occurrence levels are indicated by the different colored regions in the figure. The slope and position of the four threshold lines, marked L (Low), M (Medium), H (High) and VH (Very High), define the WBO threshold levels and were empirically defined by observations of WB time series as shown in Section 3. WBs that appear with a high frequency (WBf) or with long continuous observation periods (mctWBs) are most likely stable WBs that exist for a longer period and therefore have a higher occurrence than those with lower frequency (WBf) or shorter continuous observation periods (mctWBs). The defined occurrence threshold levels can be expressed as:

$$VeryLow=\left(mctWBs>0\right)\&\left(mctWBs<\left(2+{\alpha }^L\times WBf\right)\right),$$

(16)

$$Low=\left(mctWBs\ge \left(2+{\alpha }^L\times WBf\right)\right)\&\left(mctWBs<\left(3+{\alpha }^M\times WBf\right)\right),$$

(17)

$$Medium=\left(mctWBs\ge \left(3+{\alpha }^M\times WBf\right)\right)\&\left(mctWBs<\left(4+{\alpha }^H\times WBf\right)\right),$$

(18)

$$High=\left(mctWBs\ge \left(4+{\alpha }^H\times WBf\right)\right)\&\left(mctWBs<\left(5+{\alpha }^{VH}\times WBf\right)\right),$$

(19)

$$VeryHigh=mctWBs\ge \left(5+{\alpha }^{VH}\times WBf\right),$$

(20)

$$Permanent=WBf\ge 95,$$

(21)

where α^x is the slope of the threshold line x, for example,

$${\alpha }^{VH}=\frac{5-0}{0-60},$$

(22)

The red line (o1) in Figure 7 shows the minimum WBf in relation to the mctWBs. The slope depends on the maximum number of observations (ntObs). The plot shows that a certain pixel is classified as a “Very High” WBO if it is identified as WB in at least five consecutive dekads (point “m” in the plot). Note that this number is independent of the ntObs. As an example, if a pixel was detected as a WB in three consecutive dekads (with nObs = 31), its WBf will be 9.7% and therefore will have a WBO “Medium” (point “a”). If this pixel would have subsequently four additional WB detections in separate single dekads, its WBf will rise to 22.6% and its WBO will increase to “High” (point “b”). Note also that if pixels have a WBf of at least 95% their WBO will be “Permanent”.

2.6. Minimum Size of Detectable Water Bodies

A WB has to have a minimum spatial size in order to be detected by the WB detection algorithm. To find the minimum sub-pixel size of a detectable WB, linear spectral mixture analysis was used. Therefore, five water spectra (Figure 8a) with an increasing reflectance level were spectrally mixed with a dark and bright vegetation and soil spectrum (Figure 8b). Only the Red, NIR and SWIR bands were used as these are transformed to HUE and VALUE after the mixture. The mixing was done by increasing the sub-pixel fraction of the WB from 0 to 1 in steps of 0.05.

The WB detection threshold levels were set fixed to HUE = 100 and VALUE = 0.14. By analyzing the HUE and VALUE of the mixed spectra in relation to the detection threshold levels, the minimum fraction of the pixel to be covered by water could be derived. Figure 9 summarizes the results obtained after the mixed pixel analysis. In general, the dark WBs (WB 1, WB 2) needs larger sub-pixel water coverage than the bright WBs (WB 4, WB 5) in order to be detected. E.g., the dark WB 1 needs a water coverage of 45% when mixed with “Dark Veg” and this increases to a water coverage of 85% when mixed with the “Bright Soil”. The bright WB 5 needs a water coverage of 15% when mixed with “Dark Veg” and 40% when mixed with “Bright Soil”. As could be observed on different WB detection results worldwide, small (single pixel) WBs are mostly dark and therefore need a large sub-pixel coverage in order to be detected.

3. Results

Figure 10 shows an example of the Decision Tree Classification (DTC) result using the final threshold levels. The image shows the detection obtained for the 10-days composite of the first dekad of December 2013 (MC10_20131201) over the Rift Valley test area in Ethiopia for which the visual representation is given in Figure 3a. As shown in detail, some smaller lakes with size in the order of one PROBA-V pixel are detected. Figure 11 shows the Water Body Occurrences (WBO) for the same area. Because the larger lakes are detected in each cloud free dekad, they have a detection frequency of more than 95% and are therefore indicated with a “Permanent” occurrence. As shown in the detail window in Figure 11, the smaller lakes have a WBO ranging from “Very High” to “Very Low” due to temporal variability and are, as such, not detected in each dekad. This is indicated in the time series analysis shown in Figure 12.

Figure 12 shows the time series analysis for several smaller lakes, some of them are indicated in detail in Figure 11. Lake Wedecha consists of 2 pixels that have a “Very High” occurrence because they were detected in more than five consecutive dekads. A “Very High” occurrence can also be found for several other lakes which have frequent WBs detected, i.e., K’oftu (pixel b), Chelekleka, Hora and Bishoftu. One pixel (a) of Lake K’oftu has a “Very Low” occurrence because it was detected in two non-consecutive dekads only. Lake Chilotes has a “High” occurrence because it was detected in three and two consecutive dekads. Lake Budamada, indicated by the red arrow in Figure 11, is a small lake of about 1 km diameter. It was detected in only one dekad and therefore has a “Very Low” occurrence. Nevertheless, this crater lake is permanent, i.e., it will not dry up during the drought period. Its irregular detection is due to its small size which is in the order of one PROBA-V pixel. The small size causes mixed pixels, i.e., the water body spectral response is mixed with the spectral response of the surroundings, and therefore it is not detected as WB. This phenomenon accounts for all WBs with size in the order of one PROBA-V km pixel.

The rain season over the Ethiopian Rift Valley peaks in the months July and August while the dry season appears in the period between mid-October and mid-March. This seasonal variation can be observed in the appearance of some WBs as shown in Figure 13. The wetland area near Wererso contains no WB pixels at the end of March, i.e., at the end of the dry season. In the next three dekads, the number of detected WB pixels increases. In the last dekad of July and the first dekad of August, an extremely large number of pixels indicate WBs for this wetland area and its neighborhood. The two other lakes, Lake Boio and the Lake Koka, show a similar behavior, i.e., their size grows over time, which is in accordance with the progression of the rain season.

Figure 14 shows the Water Body Occurrence (WBO) time series for the Rift Valley test area. The plot shows per dekad the cumulated number of pixels for the occurrences ranging from “Very Low” to “Very High”. The WBs with a “Permanent” occurrence are very large and would overrule the plot, therefore they are not indicated. Note that cloud cover may mask WBs with the result that those WBs are not accounted for in the plot and therefore the plot may show some irregularity. Nevertheless, the plot shows some correlation with the seasonal variation as described before. From the first observed dekad, end of October 2013, to mid-February 2014, there is a gradual decrease of detected WB pixels which is in accordance with the dry season. The amount of detected WB pixels with a “Very High” occurrence stays more or less stable. This period is followed by a more random increase of detected WBs as a result of the upcoming rain season. The pixels that have a “Low” and “Very Low” occurrences show the largest variability. In general the number of detected WBs clearly correlates with the seasonal variability, i.e., the rain season, which also correlates with the cloud cover.

4. Assessment of Algorithm Performance

To demonstrate the maturity of the PROBA-V WBDA, reference water bodies were extracted from high spatial resolution Landsat 8 images over 15 regions selected worldwide (

Table A3. Landsat 8 scenes of the 15 test regions selected worldwide were used to assess the water bodies detection algorithm. As much as possible, cloud free scenes were selected. The corresponding PROBA-V dekad and the upper left corner coordinate are given with the corresponding number of samples (ns) and lines (nl).

Area	Landsat 8 Scene	Dekad	Upper Left Coordinate		PROBA-V
			Lat	Lon	ns	nl
Argentina	LC82310872014011LGN00	20140111	−37.83418611	−70.58413611	303	241
Australia	LC80950832013304LGN00	20131021	−32.12027778	141.27138333	288	238
Bolivia	LC82330692014233LGN00	20140821	−11.95757500	−66.85950000	233	239
Canada	LC80180262014295LGN00	20141021	49.94476111	−79.54769167	340	247
Chad	LC81830512014011LGN00	20140111	14.06290278	15.97345556	234	238
China	LC81210392014121LGN00	20140501	31.35326944	115.68581667	270	239
Finland	LC81920132014154LGN00	20140601	68.02264167	21.52997222	686	253
India	LC81420442014028LGN00	20140121	24.16315278	81.55028611	247	239
New Zealand	LC80750912014294LGN00	20141021	−43.52644444	168.42483889	329	243
Russia	LC81100132014252LGN00	20140901	68.03188333	148.26286944	676	255
Swiss	LC81950272014079LGN00	20140311	48.49976111	6.28843889	360	241
Thailand	LC81280482014090LGN00	20140321	18.38631389	101.85821667	247	237
Turkmenistan	LC81570332014133LGN00	20140511	39.95980833	62.34785278	298	242
USA	LC80260372014207LGN00	20140721	34.21517222	−96.81486667	290	237
Zambia	LC81730712014133LGN00	20140511	−14.84697222	25.17750000	240	238

of Appendix A). The workflow and algorithm to derive the reference WBs from Landsat 8 images is described in Appendix C.

The reference water bodies extracted from Landsat 8 imagery might also contain omission and commission errors and might influence the PROBA-V performance analysis. The reference commission errors occur at the Landsat 8 spatial resolution and it concerns groups of pixels ranging from one to several tens. The corresponding calculated WSR is therefore very small (a PROBA-V pixel for which the WSR is calculated comprises a few hundred Landsat 8 pixels). Typically, the WSR for the wrongly detected WB is much lower than the minimum WSR of 0.5 used in the analysis. Omission errors are much more difficult to quantify. Nevertheless, it can be said that these also concern small groups of pixels resulting in a very small reference error not influencing the analysis.

Table 1. Assessment results for the 15 test areas. The user requirements define a maximum omission and commission error of 15%. All commission errors fall within the requirements. The acceptable omission errors for the highest WSR are highlighted.

Area	CE (%)	OE (%)
		Minimum WSR
		0.95	0.9	0.8	0.7	0.6	0.5
Argentina	10.7	0	0	0	0	0.4	0.8
Australia	1.4	0	0	0	0	0.8	1.1
Bolivia	0.7	0	0	0.1	0.7	2.8	9.4
Canada	4.0	0	0	0.2	0.3	1.5	5.0
China	0	0.1	0.2	0.4	1.2	3.3	7.3
Finland	1	13.6	17.3	26.8	36.3	46.6	56.2
India	0	0	0	0	1.3	7.6	16
New Zealand	0	0.2	0.2	1.4	3.9	8.1	12.3
Russia	0.1	5.3	6.9	12.2	19.0	27.6	39.8
Swiss	0	0.6	1.1	3.5	6.5	11.8	18.9
Thailand	0	0	0	0.4	2.2	7.4	16.9
Chad	0	0	0	0	0	0.5	1.7
Turkmenistan	0.3	0	0	0.1	0.8	2.6	6.5
USA	0	0.4	0.9	6.7	14.9	24.7	32.6
Zambia	4.9	0	0	0	0	1.8	7.1
Average	1.5	1.3	1.8	3.5	5.8	9.8	15.4

shows the commission and omission errors on the PROBA-V images for six minimum WSR ranges and this for the 15 test areas.

As seen in Table 1, the commission and omission errors vary for the different test areas. An average commission error of 1.5% was obtained with an average omission error of 9.8% for minimum WSR 0.6 and 15.4% for minimum WSR 0.5. Notice that the user requirements define a maximum omission and commission error of 15% [35]. The acceptable omission errors for the highest minimum WSR are highlighted in the table. Though the Argentina test area shows the best results for the omission errors it has the highest commission error of all test areas (10.7%). This high commission error is due to dark soil pixels, which were wrongly detected as water bodies. The commission error of 4.9% for the Zambia test area is also due to dark soil pixels while the commission error of 4.0% for the Canada test area is due to dark vegetation pixels, which were detected as water bodies. The Finland test area shows the worst performance for the omission error, i.e., an acceptable omission error of 13.6% for minimum WSR 0.95. These high omission errors are due to the presence of many small and narrow lakes, in the order of the size of a PROBA-V 1 km pixel, by which they were not detected. This effect accounts for all other regions as well, i.e., for the Russian test area: 12.2% for minimum WSR 0.8; and, for the USA test area: 14.9% for minimum WSR 0.7. The other test areas show the highest omission errors for minimum WSR 0.6 and 0.5.

5. Discussion

The user requirements define a maximum omission and commission error of 15% [36]. As shown in this work, a mean commission error of 9.8% is obtained for minimum WSR = 0.6 while the mean omission error is 1.5%. Though minimization of omission and commission errors is achieved by the use of additional data layers, WBPM, VSM, PGM and NDVI, they sometimes are inevitably. Small sized WBs in the order of one PROBA-V 1km pixel and WBs with spectral properties that fall outside the defined thresholds are the cause of omission errors. Commission errors are mostly caused by anthropogenic structures having spectral signatures equal to WBs, e.g., some agricultural fields, green houses, build-up areas, etc. In addition, natural surfaces like salt planes, dark soils, mining sites, etc. might cause commission errors. An important source of commission errors are dark surfaces, e.g., due to undetected cloud shadows, dark soils due to heavy industry, dark surfaces due to shadow of high structures, etc. Therefore, to minimize confusion the user should use the water body product with attention to the quality layer which reflects the history of the detected WBs.

Special attention has to be paid to the three layers used during incompatibility masking. The WBPM was made based on the GLSDEM for which data was collected in the early 2000s. Because of the high resolution of the GLSDEM dataset (90 m spatial and 1 m vertical), the derived WBPM (1 km spatial resolution) reveals the finest detail (as shown in Figure 3b). Although the Earth’s topography changes only notably over large geological time scales, natural events (e.g., earthquakes) or anthropogenic activities (e.g., dam building) can influence the topography on shorter time scales. Therefore, when such events take place, the WBPM needs to be updated. Just to give an idea; the PROBA-V 1 km world image contains 40,320 samples × 15,680 lines or 632.22 × 10⁶ pixels. Of those, 29% (183.5 × 10⁶ pixels) consist of landmass (i.e., not ocean), and 50.5% (92.7 × 10⁶ pixels) of the landmass is indicated as potential water body.

Because the extent of glaciers worldwide are strongly influenced by climate (global warming), the PGM needs to be frequently updated. Frequent updates of the permanent glacier data are available at the National Snow and Ice Data Center (NSIDC). The geological situation in the volcanic areas of South America, i.e., Chili and Peru, is very complex and delineating the volcanic soils in these areas was not always easy. It is therefore very well possible that some smaller regions in this area are still missing in the VSM. On the other hand, future volcanic eruptions might produce new dark volcanic soils leading to false detected WBs. In both cases, an update of the VSM is needed.

6. Conclusions

This water bodies detection algorithm is implemented in the operational processing chain of the Copernicus Global Land Service (http://land.copernicus.eu/global/). It delivers 10-day global water bodies detection maps accompanied by a quality layer map.

Although some minor artifacts were introduced during the algorithm assessment using Landsat 8 reference WBs, the assessment gives a good idea about the algorithm performance. The analysis, performed over 15 test regions, shows that the product reaches a mean commission error of 1.5% while the mean omission error for minimum WSR = 0.6 is 9.8%.

The developed methodology can easily be applied to the 333 m and 100 m PROBA-V products and 300 m Sentinel-3 products, by which smaller water bodies will be detected. Applying the methodology on higher resolution sensors as Sentinel-2 is possible, but its dynamics will be hampered by the lower temporal frequency of such sensors. The water bodies detection algorithm is also tuned to detect water bodies on 1 km SPOT-VGT dataset. The whole SPOT-VGT archive is being processed and water body detection maps from 1999 to 2014 will become available by the end of 2016. An exhaustive validation exercise of the PROBA-V and SPOT/VGT 1 km WB products has been performed by independent entity and the report is available on the Copernicus Global Land Service website. The water bodies products can be downloaded from the Copernicus Global Land Service data portal: http://land.copernicus.vgt.vito.be. Additional information is available at: http://land.copernicus.eu/global/products/wb.