3.1. Integration of Open-Access Remote Sensing and Geospatial Data for Catchment-Scale Flood Modelling
shows the model performances concerning varied Manning’s roughness coefficient “n” from 0.01 to 0.045 m1/3
, suggesting 0.04 m1/3
as the optimal Manning’s roughness coefficient, especially given that the model’s F-statistic began to decline at an “n” value of 0.045 m1/3
. The optimal Manning’s roughness used was consistent with a previous study in the same study area [29
], where the optimal channel and over-bank Manning’s roughness coefficient of 0.04 was adopted for the one-dimensional SOBEK model. Some descriptors of roughness parameters within the channel and floodplain include matured crops, scattered bushes, heavy weeds, short grass, early growth vegetation, sand dunes, and meandering channels [49
]. Additionally, it could be observed that the model performance improved when re-evaluated as sub-domains rather than when treated as a whole domain. This performance variation was consistent with data quality variation, decreasing downstream of the study domain. A similar model performance variation was observed by Skinner et al. [13
], where model performance uncertainty increased with data ambiguity.
In Table 3
and Table 4
, the model performance is evaluated against optical and combined optical and radar imagery derived using a raster mosaic at the Lokoja, Onitsha and the Niger Delta sub-domains. The model appeared to perform better when evaluated against the combined optical and radar flood extent, as opposed to when only compared to MODIS optical imagery. The combination of optical and radar imagery has been widely reported to improve flood extent delineation and is particularly useful for large-scale flood monitoring [66
]. At the Lokoja sub-domain, a minimal difference was observed between the model’s F-statistic when evaluated against optical MODIS imagery and TerraSAR-X flood extents, i.e., 0.729 and 0.808, respectively, due to the limited cloud and vegetation cover in the region. In the Niger Delta sub-domain dominated by seasonal cloud cover due to nearness to the Atlantic coast, the combination of Radarsat-2 and CosmoSkyMed images resulted in an improved model predictiveness of 0.187, from 0.095 for optical MODIS only, as well as an overall reduction in Bias. This improvement can be attributed to the SAR sensors’ ability to penetrate cloud cover to delineate underlying flood. Nonetheless, the relatively low F-statistic values, despite the improvement, suggest the presence of high model uncertainty in the region that can be attributed to input variables limitations, such as SRTM DEM, as well as the SAR images’ deficiency in the mangrove-dominated regions [67
]. The overall F-statistics for the whole model domain were found to be generally low (Figure 4
and Table 3
and Table 4
) due to data and process uncertainties that transitioned into flood model outcomes [68
]. The effect of data uncertainty was further revealed in the sub-domains, where the hydrodynamic model predictive capacity was affected by spatial data disparity, as previously disclosed. The flood extent Bias and the percentages of flood captured in Table 4
and Table 5
also corresponded to the variability of data across the overall domain and the sub-domains.
Upon establishing that the best fit Caesar-Lisflood model outcome was characterized by a static Manning’s roughness coefficient of 0.04 m1/3
, it could be observed that the modelled flood extent patterns for the three sub-domains were consistent with those observed from satellite imagery (Figure 4
A–C) and reflected the data variability effect, as defined by the performance matrices, with model outcome uncertainty (over-estimation) increasing downstream as data availability reduced.
Detailed floodplain and river terrain characterization have been identified as key inputs that influence the outcomes of hydrodynamic models [69
]. SRTM DEM combined with up-to-date (2011) river bathymetric data at Lokoja resulted in a model performance of F = 0.8, a matrix consistent with other studies where DEM and bathymetry data integration into flood modelling resulted in improved model outcomes [23
]. At Onitsha, where SRTM was combined with obsolete bathymetric data acquired in 2002, before dredging activities in 2010 [72
], F = 0.5 was achieved; thus, the bathymetric data likely over-estimated the river depth, consequently resulting in an over-estimated modelled flood extent (Figure 4
B). A reduced model accuracy of approximately F = 0.2 in the Niger Delta sub-domain was attributed to the lack of bathymetry data in the flat terrain area, despite the insertion of ICESat spot height, resulting in a simplified river geometry characterization that did not explicitly capture river network details such as anabranches and meandering. This caused flood model over-estimation due to the ease of water conveyance from shallow rivers to adjacent floodplains [73
]. Additionally, unregulated sand mining activities, water-saturated mangroves, and poor dredging and debris management practices were likely factors that contributed to the model uncertainty within the region [32
], as they could influence and trigger hydrological and hydraulic changes.
Furthermore, previous studies in the region based on global flood models, such as CaMa-Flood (Catchment-based Macro-scale Floodplain), GLOFRIS (Global Flood Risk Image Scenarios), JRC (Joint Research Centre), ECMWF (European Centre for Medium-Range Weather Forecasts), SSBN (now known as Fathom Global Ltd.), and CIMA-UNEP (Centro Internazionale in Monitoraggio Ambientale and United Nations Environment Program) [41
], have revealed similar inundation patterns but with a slightly less model-to-observation agreement at Lokoja, Onitsha, and the Niger Delta. The outcomes of global models also revealed decreasing accuracy from the deep and narrow constricted rivers at Lokoja to the low-lying floodplains of the Niger Delta [75
], similar to the finding in this study. Given that some local data such river bathymetry and other validation datasets were available this study, at Lokoja specifically (which is seldom available for global flood models [75
]), the outcomes of this study at Lokoja were considerably better than global models.
Overall, the flood pattern displayed in Figure 4
A–C was consistent with the geomorphology of the sub-domains and its influence on the hydraulics of the catchment areas. For instance, at the Lokoja sub-domain, flood spread out at the confluence in Lokoja where the Niger and Benue rivers meet and propagate towards floodplains; at Onitsha, extended flood areas could be observed, and these were attributed to back-water effect caused the constricted river channel at Asaba that deflects water to fill the dish-like relatively flat floodplain [76
]; and the widespread flooding across the Niger Delta region could be linked to the low-lying topography of the region, as well as the inability of the shallow rivers (Nun and Forcados) to contain the excess water coming from upstream rivers (Niger and Benue). These characteristics suggest that enhanced river channel and floodplain topography characterization is essential for shallow channels and low-lying floodplains [79
3.2. Flood Model Re-Validation in Vegetation-Dominant Region Using Freely Available Aerial Photos and SAR
The combination of optical and radar satellite resulted in an improved model-to-observation agreement, as seen in Table 3
and Table 4
, and Figure 4
A–C. However, SAR is known to be deficient in mangroves, swamps, and built-up areas [24
], as depicted by the observed minimal change in model performance from 0.095 to 0.187 when comparing the model to optical and radar and optical image-derived flood extents, respectively, in the Niger region dominated by mangrove vegetation.
To better assess the model’s performance in the Niger Delta sub-domain where SAR is known to be deficient, aerial photo data points acquired during the 2012 flood event were applied for the first time, and the results are presented in Figure 5
A–D and Table 5
. Figure 6
shows images some of the aerial photo data points distributed across the typical environmental/physical landscape variations in the Niger Delta region; Figure 5
A shows mixed land use (built-up area greater than vegetation); Figure 5
B shows mixed land use (vegetation greater than built-up); Figure 5
C shows bare land, sparsely built land, and vegetated lands; and Figure 5
D shows matured mangrove vegetation. These physio-environmental variations are known to influence SAR inundation delineation capacities and hydrodynamic model performance [80
], as seen in Table 5
, where a higher level of agreement was observed between aerial photo data points and the model (69%) compared to SAR (13%). The used geotagged aerial photos presented actual pictures of flooded areas (Figure 6
) and were captured at heights below cloud cover; thus, image pixels were not impaired by vegetation canopy. This outcome further buttresses SAR’s deficiency in delineating flooding in mangrove-dominated regions, as well as the potential limitation of SRTM DEM to under or over-estimate terrain elevation for hydrodynamic modelling [81
]. In conclusion, this assessment provides a novel approach to ascertain the deficiencies of hydrodynamic models and SAR images in complex terrains using aerial photos. Nevertheless, better value can be derived from such data if the spatial distribution is improved and if the data are collected to enable ortho-correction for pixel or area-based comparative analysis.
3.3. Quantifying the Magnitude and Impact of the 2012 Flood in Nigeria
The 1-in-100 year flood return-period is recommended by the Technical Guidelines on Soil Erosion, Flood, and Coastal Zone Management for flood risk management in Nigeria [83
]. Based on a methodology developed from a previous study [39
], 1-in-100 year flood discharge at Baro and Umaisha gauging stations are estimated as 13,887 and 19,589 m3
/s respectively (Supplementary Figures S1 and S2
) and applied to retrospectively quantify the impact of the 2012 flood event at the Lokoja sub-domain where the highest model performance was observed, i.e., inundated land area, built-up areas, roads and affected population. A similar impact assessment was also undertaken using the peak flood discharge in 2012 (See Figure S4, supplementary material
for 2012 hydrograph), and the results are presented alongside the observed satellite flood extent in Table 6
and Figure 7
The areas observed as flooded by satellite imagery were consistent with modelled flooded areas for a 1-in-100 year flood and the peak flood of 2012 (Figure 7
), resulting in a more than 95% spatial extent agreement. Furthermore, similarities were visible for the observed and modelled flood impact for the inundated land area, built-up area, major roads and affected population displayed in Table 6
. These indicators are relevant to understand exposure to flooding, impact to infrastructure, evacuation strategy, and damage to households and livelihoods to inform future flood risk management interventions.