During the first pre-processing steps the Level 1A Radiance Counts (MOD 01), the Geolocation dataset (MOD 03), and finally the Level 1B Calibrated Geolocated Radiances (MOD 02) are calculated. The Level 1A Radiance contains counts for the 36 MODIS channels, along with raw instrument engineering and spacecraft ancillary data. The data are used as input for geolocation, calibration, and processing. The Level 1B calibrated Geolocation dataset consists of the calibrated and geolocated at-aperture radiances (W/(m² μm sr)) for 36 bands generated from MODIS Level 1A sensor counts (Figure 2a). The Geolocation dataset contains geodetic coordinates, ground elevation, solar and satellite zenith and azimuth angles for each MODIS 1 km sample. These data are provided as a companion dataset to the Level 1B calibrated radiances and the Level 2 datasets to enable further processing [18]. The processing for these three products occurs within the SeaWiFS Data Analysis System [19] and includes the product levels MOD 01–03.

During this pre-processing step MODIS Level 2 corrected reflectances are calculated from the MODIS Level 1B calibrated radiances (see Figure 2b). The processing is performed by the MODIS Corrected Reflectance Science Processing Algorithm (CREFL_SPA), provided by the NASA Direct Readout Laboratory (DRL). CREFL_SPA creates a MODIS Level 2 Corrected Reflectance product. A simple atmospheric correction of MODIS visible, near infrared, and short-wave infrared bands (Bands 1–16) is used. It requires no real-time input of ancillary data. The Corrected Reflectance products are comparable to the MODIS Land Surface Reflectance product (MOD 09) in clear atmospheric conditions [20].

The swath dimensions of MODIS data are 2,330 km (across track) by 10 km (along track at nadir), giving it the capability to cover the entire globe every one to two days. In contrast to other scanning sensors, MODIS observes, within one scan, ten lines of 1 km spatial resolution (40 lines of 250 m resolution and 20 lines of 500 m resolution, respectively). The projection of the swath to grid data is accomplished by using the MODIS Swath Reprojection Tool (MRTSwath). MRTSwath provides the capability to transform MODIS Level 1B and Level 2 data from swath format to a uniformly gridded image that is geographically referenced according to user-specified projection and resampling parameters [21].

The three result datasets (reflectances of the 250 m and 500 m resolution bands and the geolocation file) are post-processed with the GDAL software. This process encompasses the stacking of the layers, assigning “nodata” values, and the calculation of the layer statistics.

In order to enable the distinction between persistent water bodies and flooded areas, the Global Raster Water Mask (MOD44W) [22], with a 250 m spatial resolution, is used. As the MOD44W product was generated using MODIS data from 2000 to 2007, it is temporally static and therefore does not provide an indication of normal seasonal water fluctuations [23]. As the proposed pre-operational service focuses mainly on Europe, seasonal influences are comparably low. The ASTER Global Digital Elevation Model Version 2 (GDEM V2) [24], with a spatial resolution of one arc second, is employed to refine the flood mask. Slope information s_(x,y) in degrees for each pixel (x,y) are computed (local steepness of terrain) according to (1) s ( x , y ) = arctan ( ( Δ x ( x , y ) n ⋅ r x ) 2 + ( Δ y ( x , y ) n ⋅ r y ) 2 ) ⋅ 180 πwhere Δx and Δy are the result of a standard Sobel edge filter [25] applied on the DEM taking into account n = 8 pixel values (resulting from a 3 × 3 kernel). r_x, r_y denotes the pixel resolution of the DEM in the x and y directions. The auxiliary datasets are resampled and clipped with respect to the pixel size, extent and location of each MODIS scene.

The spectral indices EVI (Enhanced Vegetation Index), LSWI (Land Surface Water Index) and DVEL (Difference value between EVI and LSWI) are computed from MODIS spectral Bands 1–3 and 6. These indices have been successfully applied for water detection based on MODIS data in several studies [26–28]: (2) EVI = G × ρ NIR − ρ RED ρ NIR + C 1 × ρ RED − C 2 × ρ BLUE + L (3) LSWI = ρ NIR − ρ SWIR ρ NIR + ρ SWIR (4) DVEL = EVI − LSWIG is a given gain factor, L is a canopy background adjustment factor, C₁ and C₂ are coefficients of the aerosol resistance term, which uses ρ_BLUE to correct for aerosol influences in ρ_RED. The values of the parameters in the computation of the EVI are defined as G = 2.5, L = 1, C₁ = 6, C₂ = 7.5 according to [29,30].

A preliminary threshold based classification is performed built on an approach originally developed by Sakamoto et al. [26] and revised by Islam et al. [27], who propose a decision tree procedure to assign each image element to the temporary class “Water-related” and the output classes “Non-flood”, “Flood”, and “Mixture”. The thresholding can be summarized as follows (Figure 3):

Determination of cloud-cover areas based on a threshold of ≥0.27 from the blue reflectance band. The cloud positions are subsequently used for a geometry-based detection of cloud shadows.

The following post-processing steps are integrated into the classification process to improve the accuracy of the initial classification result derived by the approach described in [26,27].

Information derived from the ASTER GDEM V2 is employed to reduce the number of misclassified water-related pixels in areas where the plausibility for a flood occurrence is low due to topographic considerations. Areas of steep incline (>10°), or significant height (>2,000 m and a slope of >8.0°), are removed from the flood mask.

Two region growing steps are applied for a refinement of the classification accuracy by relaxing the thresholds in the neighborhood of the initial flood classification result. In the first region growing step image elements of the classes “Flood” or “Mixture” are used as seeds. Neighboring image elements of the class “Non-flood” are assigned to the class “Water-related” by using the following criteria: EVI ≤ 0.31 and DVEL ≤ 0.07 or EVI ≤ 0.06 and LSWI ≤ 0.1. The grown area is subsequently separated into the classes “Mixed” (0.1 < EVI ≤ 0.3) and “Flood” (EVI ≤ 0.1). The second region growing step is used to increase the spatial homogeneity of flooded areas by dilating preliminary extracted flood surfaces and assigning pixels of the class “Mixture” with a LSWI ≤ 0.08 to the class “Flood”. Flood pixels with neighboring mixed pixels are used as seeds.

Cloud shadows are a major problem when monitoring water surfaces using optical remote sensing data, especially in coarse/moderate resolution satellite data. As cloud shadows and water have similar spectral properties, the classification by means of spectral information alone could lead to a significant overestimation of water surfaces near clouded areas. An overview of methods that improve the classification in optical satellite data by detecting cloud shadows can be found in [31,32]. We implemented an approach for separating water and cloud shadows by combining geometrical, spectral, as well as temporal information.

In a first step, a geometry-based technique described in [33] is employed to estimate the position of cloud shadows based on 2D-location (x_img, y_img) and height (top and bottom) above the surface h_c of the cloud, the solar and viewing zenith angles, θ_s and θ_v, respectively, and the solar and viewing azimuth angles, θ_s and θ_v, respectively. Assuming a cloud pixel position (x_img, y_img), the projection of the cloud on the earth surface is defined by (x_nadir, y_nadir): (5) x nadir = x img + h c tan θ v sin ( Φ v + γ ) (6) y nadir = y img + h c tan θ v cos ( Φ v + γ )

The projection of the cloud shadow on the ground is then determined by (x_shadow, y_shadow): (7) x shadow = x nadir − h c tan θ s sin ( Φ s + γ ) (8) y shadow = y nadir − h c tan θ s cos ( Φ s + γ )where γ is the azimuth angle of the true North from the y axis. The solar and sensor zenith and azimuth angles are captured along with the MODIS data at a 1 km spatial resolution. Equations (7) and (8) approximate a flat surface. A more complex formulation considering the Earth curvature can be found in [32]. For a detailed identification of the cloud shadow positions the availability of information on the cloud top and bottom height is necessary [34]. To be independent from the availability of external data sources during rapid flood mapping activities, simplified assumptions for cloud heights are made [31,33,35].

The elevation of clouds may vary considerably depending on the cloud type and also the geographic latitude. The maximal cloud top height decreases with increasing latitude from a mean height of the tropopause from 16 km in tropical regions to 8 km in polar regions [35]. Similar to [31], the upper limit of cloud top height is divided into three classes according to the geographical latitude. For lower latitudes between 30°N–30°S the maximum cloud altitude is set to 16 km, for the mid-latitude between 30°N/S–60°N/S and in the high-latitude between 60°N/S and 90°N/S to 12 km and 8 km, respectively. The cloud bottom height is assumed to be 0 km. Therefore, only the inner and outer border (with a two pixel margin) of the cloud is projected on the ground as shadow. This significantly reduces the computational effort of the algorithm as the extent of the cloud shadow can be estimated by linearly interpolating between the border of the cloud and the border of the projected cloud-shadow.

Due to these simplifying assumptions, the simulated cloud shadow is usually larger in extent than the real cloud shadow. After the cloud shadow computation all pixels classified as flood outside the cloud shadow mask are assigned to the flood class with a high probability. Pixels covered by the cloud shadow mask are characterized by a lower probability to belong the class “Flood”. These image elements are assigned to the temporary class “Flood in Cloud Shadow” (FCS) (Figure 3). Further criteria, based on temporal and spectral information, are used to identify flooded regions within the simulated cloud shadows. If a pixel of the class “FCS” is labeled as flooded in the preceding MODIS classification result, at T-1 with a time distance of <36 h to the acquisition time of the most current MODIS scene at T0, it is assumed to belong to the class “Flood”. If the pixel is labeled as “Cloud” at T-1, further classification results of T-x of the last 36 h are iteratively integrated into the test. Only these pixels of the class “FCS” are assigned to flood elements, which fulfill the following spectral criterion [31]: NDVI (ratio between the difference and sum of spectral bands ρ_NIR and ρ_RED ≤ −0.15 and a ratio of ρ_NIR/ρ_RED ≤ 0.75).

In a final step, the separation between flooded regions and standing water areas is performed using the Global Raster Water Mask (MOD44W) at 250 m spatial resolution [22].

The fully automated TerraSAR-X flood processor is designed to derive individual scene dependent threshold values for global data acquired with different sensor configurations (i.e., polarization, beam mode, and incidence angle). For this purpose, a parametric tile-based thresholding, proposed in [2], is modified for more robustness and adapted for SAR data, radiometrically calibrated to σ₀ (dB) [16]. The method identifies n image tiles, which are likely to represent a bimodal distribution of the classes to be separated (i.e., water and land surface). The Kittler and Illingworth thresholding approach [39] is used to derive local threshold values τ_n using a cost function, which is based on modeling the sub-histograms of each tile as bi-modal Gaussian mixture distributions. A global threshold, τ_g, is obtained computing the arithmetic mean of the local thresholds. If the derived global threshold exceeds a certain value (e.g., −10 dB) it is assumed that either no water areas exist in the covered region; or the water extent is very small; or water bodies do not appear as dark backscatter regions. This could be due to wind-induced roughening of the water surface or protruding vegetation leading to volume or double bounce scattering of the radar signal. In this case, the threshold τ_g is approximated using an empirically derived linear function depending on the center incidence angle of the SAR scene [16].

The initial labeling derived by applying the threshold τ_g to the image is further enhanced using a fuzzy logic-based post-classification algorithm. The fuzzy set incorporates SAR backscatter, digital elevation, and slope information, as well as the size of water bodies. The shape of the four membership functions (standard Z and S membership functions) is either determined according to statistical computations or is defined based on empirical studies. The fuzzy thresholds for each function are individually computed for each image element. The corresponding fuzzy elements are combined into one composite fuzzy set by computing the average of the membership degrees of each pixel. The membership degree of the composite fuzzy set is assigned by a membership degree of zero in the event that a single fuzzy element has a membership degree of zero. The flood mask is created in a threshold defuzzification step, which assigns each image element with a membership degree > 0.5 to the class “Flood”.

A subsequent region growing step is performed in order to integrate the transient shallow water zone between open flood water surfaces and non-flooded areas. The water bodies extracted by using the defuzzified classification result are used as seeds for dilating the water regions. The water areas are progressively enlarged until a fuzzy logic based tolerance criterion is reached, where only image elements located in the neighbourhood of the flood areas are iteratively scanned.

Finally, the GIM is integrated into the classification process to remove areas incorrectly classified as “flood” due to radar layover and shadowing effects. To differentiate between flooded, standing, and receding water bodies, the classification result is compared to a global reference water mask.