All-Weather Flood Mapping Using a Synergistic Multi-Sensor Downscaling Framework: Case Study for Brisbane, Australia

Highlights

What are the main findings?

• High temporal resolution, low spatial resolution space-based remote sensors can provide valuable insights during densely clouded flood events.
• Data sources of disparate spatial resolutions can be harmonized and integrated using topography.

What are the implications of the main findings?

• By successfully harmonizing free, public sensor data in a multi-sensor framework, this method offers a viable, cost-effective alternative for rapid flood mapping in resource-constrained scenarios.

Abstract

Despite a growing number of Earth Observation satellites, a critical observational gap persists for timely, high-resolution flood mapping, primarily due to infrequent satellite revisits and persistent cloud cover. To address this issue, we propose a novel framework that synergistically fuses complementary data from three public sensor types. Our methodology harmonizes these disparate data sources by using surface water fraction as a common variable and downscaling them with flood susceptibility and topography information. This allows for the integration of sub-daily observations from the Visible Infrared Imaging Radiometer Suite and the Advanced Himawari Imager with the cloud-penetrating capabilities of the Advanced Microwave Scanning Radiometer 2. We evaluated this approach on the February 2022 flood in Brisbane, Australia using an independent ground truth dataset. The framework successfully compensates for the limitations of individual sensors, enabling the consistent generation of detailed, high-resolution flood maps. The proposed method outperformed the flood extent derived from commercial high-resolution optical imagery, scoring 77% higher than the Copernicus Emergency Management Service (CEMS) map in the Critical Success Index. Furthermore, the True Positive Rate was twice as high as the CEMS map, confirming that the proposed method successfully overcame the cloud cover issue. This approach provides valuable, actionable insights into inundation dynamics, particularly when other public data sources are unavailable.

1. Introduction

Globally, flooding is a significant natural hazard that is occurring over increasingly dangerous magnitudes and frequencies [1]. Coupled with growing populations in flood-prone areas, reliable flood monitoring is an essential component for effective disaster risk management [2]. Earth Observation (EO) is a popular option for flood monitoring and detection, as utilizing freely available public data makes it a scalable solution that provides data globally, even in poorly gauged regions. However, EO may encounter issues with adverse weather conditions, insufficient temporal frequencies, and spatial resolutions required to provide details of value for practical applications. Proper disaster management and emergency response require actionable flood information at least daily for optimal situational awareness [3], making it impractical to wait for the ideal clear-sky image or Synthetic Aperture Radar (SAR) overpass. In fact, recent research highlights a significant observational gap when relying solely on these high-resolution missions. Munasinghe et al. [4] estimated that nearly a third of flood events are missed, due to temporal misalignments or cloud coverage, using the Landsat-8/9, Sentinel-2, and Sentinel-1 satellites. These missions, while preferred for flood mapping due to their high spatial resolution and level of detail, cannot provide flood maps on a consistent basis individually. Some fusion approaches utilize these high spatial resolution sensors, such as Li et al. [5] and Liu et al. [6]; however, these approaches rely on procedures that reconstruct flood events rather than observe them directly. This reliance on post-event inference is especially difficult during the peak of a flood, when persistent cloud cover severely limits the auxiliary temporal information needed for an accurate reconstruction. Consequently, the introduced delays and uncertainty make the resulting maps insufficiently timely and reliable for operational applications and complete historical analyses.

Utilizing datasets with coarser spatial resolution but higher temporal frequency allows for more timely and meaningful flood updates. Some researchers have applied this principle with success to derive flood insights, particularly using the Moderate Resolution Imaging Spectroradiometer (MODIS), with a twice-daily repeat and numerous spectral bands for water and land segmentation. Brakenridge & Anderson [7] demonstrated the utility of MODIS, providing the basis for decades of hydrology research. As MODIS aged, Visible Infrared Imaging Radiometer Suite (VIIRS) was launched to help maintain continuity with similar temporal, spatial, and spectral characteristics. Li et al. [8] used this data to create an operational flood mapping system, providing the floodwater fraction detectable in each 375 m pixel. The VIIRS floodwater fraction mapping algorithm is robust, relying on a combination of indices such as the Normalized Difference Vegetation Index (NDVI), Normalized Difference Snow Index (NDSI), and the Normalized Difference Water Index (NDWI), undergoing cloud and cloud shadow masking, terrain shade masking, and eventual dynamic nearest neighbour searching to obtain the final water fraction. The floodwater is estimated by masking out the permanent water fraction derived from the MODIS global water mask. Li et al. [9] further adapted the VIIRS floodwater fraction algorithm with success to geostationary sensors, such as the Geostationary Operational Environmental Satellites R Series (GOES-R) and the Advanced Baseline Imager (ABI) and the Himawari Advanced Himawari Imager (AHI), deriving the floodwater fraction at a 1 km spatial resolution. As geostationary sensors such as ABI and AHI acquire imagery every 5–15 min, the maximum cloud-free flood extent composite may be created for each day to best characterize maximum flood extents using this imagery.

However, clouds and darkness still pose significant barriers to relying on optical imagery, making microwave sensors useful. Considering the current limitations of publicly available Synthetic Aperture Radar (SAR), it can be advantageous to exploit sensors that passively operate in the microwave spectrum which acquire data on a more frequent basis. Passive Microwave Radiometry (PMW) for flood detection relies on the strong emissivity differences between water and land and finds concrete applications for hydrological studies. Giddings & Choudhury [10] pioneered the usage of PMW emissivity differences between land and water to monitor hydrological processes. PMW for flood monitoring specifically has been revived recently, particularly with the emergence of Soil Moisture Active Passive (SMAP), as it remains a reliable option under clouded and dark weather conditions globally with a higher temporal frequency than SAR. Operating with the L-band frequency, SMAP has a native spatial resolution of 40 km. Du et al. [11] integrated SMAP and Landsat-8/9 data in a Machine Learning (ML) framework to attempt to predict flood extents. Colosio et al. [12] employed an enhanced-resolution PMW dataset to evaluate its ability to detect and monitor flood events in Bangladesh, successfully finding that PMW data can be used reliably for this purpose. In a similar vein, Zeng et al. [13] exploited the all-weather and frequent capabilities of PMW to orchestrate a two-step system, where SAR flood algorithms would be processed over an area of interest if the PMW data flagged an anomaly. PMW is a valuable data source for flood detection and monitoring; however, its coarse spatial resolution can be a barrier—PMW sensors, such as the Advanced Microwave Scanning Radiometer 2 (AMSR2), have a minimum spatial resolution of 10 km.

These potential data sources for timely flood information have variable spectral and spatial components, posing a significant challenge for data fusion. To overcome these challenges, the research community has largely prioritized ML and, more recently, deep learning approaches to integrate datasets and enhance the temporal consistency of flood mapping. A primary goal is spatiotemporal fusion, which aims to create dense time-series of high-resolution imagery by blending data from sensors with complementary characteristics, such as frequent but coarse MODIS, and detailed but infrequent Landsat imageries. Foundational algorithms like the Spatial and Temporal Adaptive Reflectance Fusion Model and the Flexible Spatiotemporal Data Fusion model established the core principles of this approach [14,15]. More recent work focuses on optical-to-optical fusion, using ML-based regression or spatiotemporal neighbourhood similarity to fill gaps in surface water images [16,17]. Advanced deep learning models now aim to predict or “nowcast” future high-resolution images between actual satellite overpasses, using models like Convolutional Long Short-Term Memory to simulate Sentinel-1 observations for improved temporal change detection [18,19].

Another strategy involves fusing multi-modal data from sensors that rely on different physical principles to leverage their complementary strengths. The most common pairing combines high-resolution optical and SAR data, marrying the rich spectral detail of optical sensors with the crucial all-weather, day-night capabilities of SAR. For instance, Markert et al. [20] developed a gap-filling framework combining Sentinel-1 and Landsat-8 to generate frequent surface water maps, a technique also explored by Lan & Wang [21] for large-scale flood analysis. Some models go even further, like the ensemble CNN developed by Seydi et al. [22] which fuses SAR, optical, and altimetry datasets for improved flood extent mapping. To achieve all-weather, high-frequency monitoring, a growing body of research has focused on integrating coarse-resolution PMW data. Studies have demonstrated the potential of integrating optical and PMW data for flood detection and developing data-driven methods for all-weather PMW inundation retrieval [23,24]. Likewise, ML models have been developed to fuse PMW with SAR data for frequent insights into floods and related fields like soil moisture and sea ice [25,26,27,28].

Recent advances have also focused on fusing data from different platform types to work towards more integrated satellite use in disaster management. One strategy involves combining high-frequency Geostationary (GEO) satellites with high-detail Low Earth Orbit (LEO) satellites. This GEO-LEO fusion leverages the near-continuous temporal coverage of GEO sensors like the GOES-R series and the Himawari series to fill the persistent gaps between LEO observations [29]. By integrating these complementary data streams, researchers can create a more complete and timelier flood product that captures significant flood dynamics that would otherwise be missed. For example, Li et al. [30] implemented a novel data fusion approach to downscale and integrate data from VIIRS, ABI, AHI, and Sentinel-3 Sea and Land Surface Temperature Radiometer (SLSTR). This approach cleverly harmonizes different data sources by using the floodwater fraction as a common variable, successfully bridging their spatial and spectral differences. This research seeks to expand upon that concept by integrating PMW data from the AMSR2 aboard the Global Change Observation Mission-Water (GCOM-W). The inclusion of PMW is critical for obtaining a more complete picture of flood events, as its all-weather capabilities can overcome persistent cloud cover. To achieve this multi-sensor integration, a reliable downscaling methodology may be valuable to gain higher-precision insights into floodwater dynamics and their impacts [30,31,32].

A variety of floodwater fraction downscaling approaches exist in the literature, each with distinct strengths and limitations. Galantowicz & Picton [33] created a relative floodability map based on each pixel’s elevation and its distance to the main river channel. While a reasonable and long-established method, a known limitation is that it can produce unrealistic water distributions by allowing water to cross separate catchment area divides if they are close in proximity and elevation. Similarly, Nguyen & Aires [34] used a neural network to derive a floodability index based on both hydrologic and topographic components. However, their framework relies on a smoothing process for hydrologic connectivity that risks removing valuable details. In contrast, Li et al. [35] propose a physically based model that iteratively fills a basin from its lowest point until the observed water fraction is matched. This is a physically intuitive approach, but its strict reliance on the absolute lowest elevation makes it highly sensitive to minor Digital Elevation Model (DEM) inaccuracies or unmapped features like dikes and levees. The proposed method builds directly on this concept but offers a key improvement: it can prioritize flooding in known floodplains based on a more comprehensive susceptibility model, even if those areas are not the absolute lowest points in the DEM This hydrologically realistic downscaling approach is the indispensable component of the proposed framework, enabling the synergistic integration of multiple sensors to overcome the challenge of infrequent, high-resolution, cloud-free imagery that hinders the timely mapping of floods. We propose a fusion-downscaling framework to create a more resilient and accurate flood mapping system. By synergistically combining passive microwave radiometry (AMSR2) with sub-daily multispectral data (VIIRS and AHI), our approach provides frequent flood extent estimations, even under adverse meteorological conditions. Our method addresses the limitations of previous approaches by integrating a multi-factor susceptibility model with a physically based filling mechanism, providing a more realistic distribution of water that can account for minor DEM inaccuracies and unmapped features. This is achieved through a downscaling-based integration approach that harmonizes surface water fraction data from multiple sensors, despite their varying spatial resolutions, generating flood maps at meaningful spatial and temporal scales, specifically 30 m flood maps every two days. The performance of this method was evaluated in the urban environment of Brisbane, Australia, during the February 2022 flood, highlighting the utility of the method in providing comprehensive flood maps when public sensors popular for flood mapping, such as Sentinel-1, Sentinel-2, and/or Landsat-8/9, are unable to provide information due to temporal or cloud cover constraints.

2. Materials and Methods

2.1. Data

The method for multi-sensor data integration involves two main steps (1) the floodwater fraction derivation and (2) the downscaling mechanism. A high-level overview of the process is provided in Figure 1. The essential datasets for this procedure are outlined in

Table 1. Key datasets necessary for the multi-sensor flood map. Although AMSR2 provides daily coverage, the exclusive use of the descending pass limits data availability for the specific area of interest to a two-day interval.

Dataset	Spatial Resolution	Revisit Time	Source
VIIRS floodwater fraction product	375 m	12 h	Li et al. [8]
AHI floodwater fraction product	1 km	5–15 min	Li et al. [9]
AMSR2 Level 3 Brightness Temperature (18.7, 23.8, 36.5 and 89.0 channels)	10 km	2 days	JAXA [36]
FABDEM 30 m DEM	30 m	-	Hawker et al. [37]
JRC Surface Water Occurrence	30 m	-	Pekel et al. [38]
HydroBASINS	-	-	Lehner & Grill [39]

.

2.2. Floodwater Fraction Derivation

The floodwater fractions from VIIRS and AHI are downloaded directly from the George Mason University archive (https://jpssflood.gmu.edu/, accessed on 1 July 2025). VIIRS 1 day and AHI 1 day composite datasets were utilized. The AHI 1 day composite is a composite of the maximum cloud-free flood fractions detected throughout the day, with the Himawari AHI acquiring data every 5–15 min. These datasets provide pre-calculated floodwater fractions, as calculated using the methodologies of Li et al. [8] and Li et al. [9]. The VIIRS floodwater fractions are provided at a 375 m spatial resolution while the AHI floodwater fractions are provided at a 1 km spatial resolution. This study utilizes the AHI constellation for its coverage of the study area, although other geostationary satellite systems, such as GOES-R or Meteosat Third Generation (MTG), could be integrated to achieve global coverage. To ensure temporal consistency with the two-day revisit cycle of the AMSR2 sensor, a final processing step was applied to both the VIIRS and AHI datasets. The maximum floodwater fractions observed within the daily VIIRS and AHI composites are calculated across a 48 h window. This process captures the peak flood conditions in clear-sky areas over that specific timeframe. This two-composite can then be integrated with the corresponding AMSR2 observation to fill in cloud-covered regions, resulting in a single, comprehensive snapshot of the maximum flood extent for each two-day interval.

Calculating the AMSR2 floodwater fraction integrates data from the 18.7, 23.8, 36.5 and 89.0 GHz channels of the AMSR2 sensor provided by the Japanese Aerospace Exploration Agency (JAXA) (ftp.gportal.jaxa.jp). These channels are all available at a 10 km spatial resolution as a Level 3 product. For this study, the descending pass is used to better account for the diurnal cycle and reduce potential biases, as the daily cycle of solar radiation can cause significant variations in land surface temperature and increase uncertainty in the derived floodwater fraction [40]. The first task in this process involves flagging pixels with active precipitation using the Scattering Index, calibrated by Grody [41]. This process first requires exploitation of the brightness temperatures from the 18.7 GHz and 23.8 GHz channels in Equation (1):

$$F=A+B{T}_v\left(18.7\right)+C{T}_v\left(23.8\right)+D{T}_v{\left(23.8\right)}^2$$

(1)

where $T_v\left(18.7\right)$ is the 18.7 GHz channel in the vertical polarization, $T_v\left(23.8\right)$ is the 23.8 GHz channel in the vertical polarization, and variables A, B, C, and D are pre-calibrated variables 450.2, −0.506, −1.874, and 0.00637, respectively [41,42]. Next, the resulting value is placed in Equation (2)

$$SI=F-T_v\left(89.0\right)$$

(2)

where $T_v$(89.0) is the 89.0 GHz channel in the vertical polarization. Precipitation flagging is based on the resulting SI value, flagging pixels with index values greater than 10 K as active precipitation [41]. The flagged pixels are used to mask the 36.5 GHz channel to yield only valid pixels for each day.

As both the vertical and horizontal polarizations of the 36.5 GHz channels are highly sensitive to surface water, these data are used to obtain the final floodwater fraction. The brightness temperature of each pixel in the 36.5 GHz channel is composed as a mixture of signals from land and water endmembers, thus linear mixture modelling is an ideal approach to obtain the pure water endmember (e.g., the surface water fraction). To extract this endmember, the Polarization Ratio Index (PRI) is obtained using both polarizations in Equation (3)

$$PRI={\displaystyle \frac{T_v\left(36\right)-T_h\left(36\right)}{T_v\left(36\right)+T_h\left(36\right)}}$$

(3)

The PRI minimizes the effects of the physical temperature of the land surface and helps amplify the signal related to the presence of surface water, as water polarizes microwave emissions differently than dry land and the PRI is very sensitive to this difference [43]. The relationship between the PRI and the surface water fraction can be approximated using Equation (4)

$$Q_{obs}=sqrt\left(PRI+a\right)+bPRI$$

(4)

where $Q_{obs}$ represents the surface water endmember, and variables a and b have values of 0.11 and −0.58, respectively [43]. $Q_{obs}$ is then implemented in a multi-endmember approach to calculate the final floodwater fraction with Equation (5)

$$f_{flood}={\displaystyle \frac{Q_{obs}-Q_{pow}{f}_{pow}-Q_{dry}\left(1-f_{pow}\right)}{Q_{pow}-Q_{dry}}}$$

(5)

$Q_{pow}$ the persistent open water endmember, is an empirically derived constant of 0.46 [43]. The $Q_{dry}$ value is dynamically calculated as the 25th percentile of the $Q_{obs}$ value from the current scene. $f_{pow}$,the persistent open water fraction, derived using the Joint Research Center (JRC) surface water occurrence dataset [38]. This layer is ingested as a pre-aligned input matching the 10 km grid of the AMSR2 data and averaged to obtain the permanent water present in each 10 km cell. The final calculated variable, $f_{flood}$, represents the floodwater fraction.

2.3. Downscaling Mechanism

The downscaling mechanism uses topography, historical water occurrence, and hydrologic basin divisions to preserve hydrological realism and connectivity while suppressing artifacts and adopts a hybrid strategy that merges flood susceptibility with physically based water filling for water fraction downscaling. The flood susceptibility map allows the model to produce more realistic inundation patterns than methods that rely purely on elevation, correctly identifying flood-prone areas like terraces or wetlands that may not be the absolute lowest points but are hydrologically predisposed to flooding. Furthermore, the method prioritizes physical realism by strictly enforcing hydrologic isolation, preventing water from artificially crossing watershed divides. The previously established downscaling methods often rely on unconstrained spatial smoothing and struggle to maintain distinct separation between adjacent catchments.

2.3.1. Flood Susceptibility

The method first uses the flood susceptibility mapping to prioritize pixel inundation within the coarse pixel footprint, differing from Li et al. [35], who use the absolute lowest point as the starting point and Galantowicz & Picton [33], who use the channel-relative elevation. Considering multiple hydrologic conditioning factors allows for a more realistic flood patterns in varied terrain.

The flood susceptibility map is based on a normalized set of Water Occurrence Frequency (WOF), terrain data and its derivatives. The terrain layers, including the Topographic Wetness Index (TWI), Height Above Nearest Drainage (HAND), flow accumulation, and slope, are derived from the FABDEM 30 m DEM using the pysheds Python v3.10 library [37,44]. The WOF and terrain layers are normalized to 0 to 1 and merged into a single flood susceptibility layer. The WOF and HAND layers are weighed the most, followed by the flow accumulation, TWI, and the slope. The final flood susceptibility layer is then normalized to 0 to 1, with 1 indicating areas with high flood susceptibility.

2.3.2. Physical Downscaling

The latter portion of the downscaling method is based on a physical mechanism, which involves physically filling based on modified topography and enforcing connectivity constraints. Firstly, the flood susceptibility layer is used to modify the original DEM, creating an effective DEM that is conditioned for the hydrological likelihood of floodwater presence [4]. The water fraction downscaling uses the fraction to determine the number of fine resolution pixels and then fills based on topography and susceptibility. First creating an effective DEM based on the flood susceptibility data layer, the algorithm conducts a priority ranking of all relevant pixels. This ranks the pixels from most susceptible to flooding to the least. Next, the number of fine resolution flood pixels available, using the DEM pixels for reference, is determined from the coarser floodwater fraction. Upon gaining this estimate, the target water surface level (WSL) for that specific footprint may be determined like the method of Li et al. [35]. The algorithm fills flood-susceptible pixels until the pixels are all distributed, and the appropriate WSL is reached, providing a physically conditioned estimation of the floodwater pixel distribution. By distributing the pixels within the coarse footprint while strictly enforcing hydrological basin boundaries, hydrological connectivity and correctness are maintained.

2.4. Final Flood Mapping

This process is conducted for all the provided data types (VIIRS, AHI, AMSR2) to yield floodwater maps at a shared 30 m spatial resolution. These maps are subsequently merged to yield one final merged map for the day of interest. While temporal misalignments between the sensor acquisitions cannot be avoided, the final merged map can be considered the maximum possible extent from the provided sensors for that day. This provides value for actionable purposes such as disaster response, flood risk assessment, and insurance. Furthermore, the derivation of the maximum flood extent for a flood event is beneficial for numerical model calibration and validation [45]. Furthermore, the final spatial resolution (30 m) matches that of publicly available global DEMs, helping ensure consistency in versatile solutions for both public services and private industry [46]. Spatial resolutions up to 30 m are generally accepted for risk assessment in meso- to large-scale hydrometeorological extremes in the insurance industry [47,48].

2.5. Case Study

Significant flooding occurred in Eastern Australia in February 2022, heavily impacting the Queensland state capital, Brisbane (Figure 2). A slow-moving low-pressure system dumped intense rain, nearly 1100 mm between 23 and 27 February, on already-saturated areas [49]. The intense precipitation delivered a volume of rainfall approaching Brisbane’s annual average in a short period, resulting in an estimated $7.7 billion AUD [5 billion USD] in economic and social costs in Queensland alone [50].

Following the event, the Brisbane City Council constructed a maximum flood extent map for Greater Brisbane by integrating approximately 2000 ground survey points with debris heights along with river and creek gauges into a calibrated hydrologic model re-run with the actual rainfall input [51]. The wealth of data and the calibrated model aided in determining the maximum extent of the flood and the timing of the peak flood level, in the early hours of Monday, 28 February. The flood extent generated from this data by the Brisbane City Council is used as the ground truth in this study. The flood peak is also evident in river water level readings at the Port Office gauge (Figure 3).

A Sentinel-2 satellite image was acquired on the same day as the flood peak. However, with 81% cloud cover, it was rendered unviable for flood assessment. Furthermore, Sentinel-1 did not pass during the event. Levin & Phinn [52] provided an initial assessment of the flood extent in this event with commercial imagery from PlanetScope and Capella, but the study’s scope did not extend to a thorough validation of the final map’s accuracy. This event thus presents an ideal opportunity to develop and validate a weather-independent flood mapping approach derived exclusively from public data. In addition to the flood extent from the Brisbane City Council, we used a 50 cm resolution Maxar GeoEye commercial image from 28 February 2022, processed by the Copernicus Emergency Management Service (CEMS) to further benchmark the proposed method.

2.6. Comparison and Validation Procedure

The results are validated using standard performance metrics for one-class classifications, listed in

Table 2. Performance metrics used in flood map validation. TP = True Positive, FP = False Positive, and FN = False Negative.

Metric	Equation
Overall Accuracy (OA)	${\displaystyle \frac{(TN+TP)}{(TN+FN+FP+TP)}}$
True Positive Rate (TPR)	${\displaystyle \frac{TP}{(TP+FP+FN)}}$
False Alarm Rate (FAR)	${\displaystyle \frac{TP}{(TP+FN)}}$
Critical Success Index (CSI)	${\displaystyle \frac{FP}{(TP+FP)}}$
F1 Score	${\displaystyle \frac{2TP}{(2TP+FP+FN)}}$

. The Overall Accuracy (OA) is a measure of the proportion of correctly classified flood and dry pixel instances. The True Positive (TP) metric quantifies the proportion of correctly classified flood pixels, whereas the False Positive (FP) quantifies the proportion of incorrectly classified flood pixels. The TP and FP are both important for quantifying the costs of misclassification and evaluating the model’s overall sensitivity [53]. The False Alarm Rate (FAR) indicates the proportion of dry pixels that are incorrectly classified as flooded. The Critical Success Index (CSI) provides the overlap between the predicted and actual flooded areas by dividing the TP by the sum of all pixels involved in the event [54]. The CSI minimizes biases due to the areas that are dry land in both modelled and truth data, as there tends to be a large imbalance in the quantity of dry land pixels versus flood pixels [55]. Finally, the F1 score yields the harmonic mean of precision and recall, to evaluate if there is balance between the high precision (e.g., minimal false positives) and high recall (e.g., minimal false negatives). All metrics range from 0 to 1. A score of 1 indicates a perfect match for all metrics except FAR, for which 0 indicates a perfect match. The benchmarking datasets are available at a higher spatial resolution than 30 m, thus these layers were bilinearly resampled to a 30 m spatial resolution, for consistent comparison, before evaluation.

3. Results

The multi-sensor downscaling approach successfully captured the progression of extreme flooding in Brisbane, providing a detailed view of the flood dynamics over the course of the flood event (Figure 4). The maximum extent, seen in Figure 5, was closely evaluated against the other flood extent sources for a thorough validation assessment.

The performance of the proposed approach was assessed in its overall accuracy against both the flood extent from the Brisbane City Council and the high-resolution Maxar image processed by the CEMS. The CEMS map only covered a small portion of the event, so this portion was utilized for the performance metrics (Figure 6).

The performance of the CEMS map against the Brisbane City Council map was poor, as the Maxar GeoEye image had approximately 60% cloud cover present, thus the flood extent was not fully depicted (

Table 3. Performance metrics of the CEMS, proposed method, and the Brisbane City Council flood extent dataset.

Method Tested	Benchmark	OA (%)	CSI (%)	TPR (%)	FAR (%)	F1 Score (%)
CEMS	Brisbane City Council	86.23	33.56	35.02	11.05	50.26
Proposed Method	CEMS	87.37	34.66	84.48	63.21	51.26
Proposed Method	Brisbane City Council	90.40	59.49	71.30	21.78	74.60

) [56]. Despite this, the FAR was still higher than expected. This is likely due to the false positive floodwater pixels at the eastern bank of the river, circled in Figure 6. The proposed method experienced a similar error in that location, potentially indicating that the global DEMs utilized by both the CEMS and the proposed method did not appropriately capture microtopography or other urban changes in this area.

The evaluation of the proposed method against the Brisbane City Council’s final extent yielded satisfactory results, particularly in comparison to the CEMS results. As previously highlighted, there were some pixels wrongly identified as flooded on the eastern bank of the river. The CEMS evaluation against the ground truth yielded a TPR (35.02%) that indicates the CEMS map missed a significant portion of the floodwater, likely due to the cloud cover present in the image. While the FAR (11.05%) was lower than that of the proposed method (21.78%), the proposed method detected far more floodwater and maintained a much higher F1 score and OA. The F1 score (74.60%) also confirms that the proposed method achieved satisfactory spatial agreement with the actual extent, providing a balanced performance in terms of correctness and completeness. The proposed method’s F1 score is significantly higher than CEMS (50.26%) assessed against the Brisbane City Council’s extent. This advantage is mainly due to the proposed method’s success with true flooded pixels, as shown by the higher TPR (71.30%) in comparison to CEMS. The CEMS map missed nearly two-thirds of the flooding, primarily due to the cloud cover present in the imagery. Furthermore, the proposed method’s CSI (59.49%) is higher than the evaluation of the CEMS (33.56%), indicating a more accurate representation of the flood’s true extent.

4. Discussion

4.1. Enhanced Spatiotemporal Flood Monitoring

This work aims to close the observational gap in flood monitoring by enhancing the spatiotemporal resolution of flood observations. High-resolution sensors like Landsat-8/9, Sentinel-1, and Sentinel-2, while excellent for detailed mapping, often fail to capture the peak of a flood due to their infrequent revisit times and susceptibility to cloud cover [57]. This was the case during the 2022 Brisbane event, providing an ideal scenario to test the proposed method. By leveraging the high-frequency observations of geostationary sensors and the all-weather capability of passive microwave radiometry, our framework ensures that critical flood dynamics are captured even under difficult atmospheric conditions. This multi-sensor approach builds both redundancy and resiliency into the mapping system. If one data stream is compromised, whether due to technical malfunctions or data availability issues, the framework can still generate a valuable flood map from the remaining sources.

The practical value of this resilience was highlighted when comparing the proposed method to the CEMS product for this event. The CEMS map, derived from high-resolution commercial Maxar imager, was compromised by approximately 60% cloud cover. This impact is quantified by the TPR of only 35.02% against the Brisbane City Council’s flood extent, indicating that it failed to capture nearly two-thirds of the actual inundated area. Consequently, it also had a low CSI of 33.56%. In contrast, the multi-sensor approach achieved a TPR of 71.30%, nearly double that of CEMS, and a CSI of 59.49%, representing a 77% improvement.

While the framework performed well in this complex urban setting, it is equally important to analyze the sources of its inaccuracies, particularly its higher FAR. The CEMS map’s low FAR is partly due to its cloud-covered data faps. As the CEMS map detected very little flooding, it had few opportunities to generate false positives. In contrast, the proposed method’s FAR still suggests erroneous flooding that must be explained. Some errors in the final flood map likely stem from the global DEM’s inability to resolve local microtopography. This is strongly suggested by the observation that both the proposed method and the independent CEMS map produced identical false positive errors in the same location, suggesting a shared limitation in the underlying global elevation data. Furthermore, the modelled flood extent from the Brisbane City Council did not appropriately capture certain flood dynamics, such as overland flow from small creeks [51]. As a result, it is possible that the remotely sensed datasets detected this overland flow, while the Brisbane City Council dataset did not, due to the hydrologic modelling limitations.

4.2. Utility for Diverse Applications

The approach demonstrates success in providing flood extent observations at a high spatial and temporal resolution using coarse resolution imagery. Its performance against a variety of benchmarks, including ground truth and commercial imagery illustrates its utility in delivering flood information. A particularly relevant point in this public-data focused approach is its scalability and cost effectiveness compared to high-resolution commercial imagery. While commercial satellites provide high levels of detail, the extent is limited and thus may not be acceptable for large-scale flood events, their limited spatial extent often requires tasking an expensive mosaic of hundreds of images to capture a large-scale flood, a process vulnerable to high costs and delays. This is significant, as low- and middle-income countries face the highest exposure to flood events [58]. As the commercial satellite sector continues to grow, numerous constellations composed of many satellites may overcome the aforementioned issues providing timely and cost-effective data. However, especially in the optic of promoting open science, public imagery represents the best possible option for a variety of applications and providing insights for impacted communities.

The value of such insights, however, is determined by their fitness for a specific purpose, as end-user requirements are variable. For instance, flood mapping for emergency response assistance prioritizes timeliness above all; data is needed within 12–48 h, while spatial resolution is less critical [47,59]. In contrast, the insurance industry for risk assessment often requires relatively high spatial resolutions, but timeliness may be less important [47,48]. Our proposed framework can contribute valuable information to both user groups. The operational, low-latency nature of the AHI, VIIRS, and AMSR2 sensor data makes our 30 m, two-day product a viable and reliable tool for meeting the rapid timelines of disaster response. Simultaneously, it addresses a key limitation for the private sector. While the 30 m resolution is a constraint, the framework’s ability to capture flood information when high-resolution commercial sensors fail, as demonstrated by the CEMS map, provides a valuable dataset for post-event damage assessments.

4.3. Limitations

However, it is important to consider the sources of errors and limitations of this framework. First, any inaccuracies in the terrain can propagate further errors in the downscaling estimation. A 30 m global DEM, such as the FABDEM, cannot sufficiently capture elements of the microtopography, such as small depressions, peaks, and ridges, rather it will smooth these features. This smoothing is known to overestimate hydrologic connectivity [60]. For the proposed downscaling method, this could cause a fundamental misrepresentation of how floodwater may be distributed. It would be prudent to replicate this approach using a higher quality DEM. This uncertainty in global datasets also applies to the permanent water mask; errors in the global WOF dataset may contribute to errors. For instance, Mueller et al.’s [61] WOF dataset was created specifically for Australia and may be more accurate. Furthermore, the downscaling method has been successfully evaluated in a pluvial-fluvial flood event but not yet in a purely pluvial situation. As it is designed, the downscaling method prioritizes fluvial flooding conditions. This is an important element that requires further testing to ensure true transferability, not solely across environments but also across other hydrometeorological extremes.

Additionally, the estimation of floodwater fractions may be erroneous. Sensors like VIIRS and AHI can face issues with dense vegetation and shadows from terrain or clouds, in addition to the cloud cover [8]. The AMSR2 sensor faces several ambiguities as well, including the presence of heavily saturated soils, which can have a similar signal to surface water [62]. Coastal areas expose some difficulties as the coarse resolution PMW pixels cover both land and ocean in a single pixel, with ocean contamination making these pixels less reliable and susceptible to false positives or overestimations of floodwater [34,63].

5. Conclusions

This study proposed and validated a novel multi-sensor downscaling framework to address the persistent observational gap in timely, high-resolution flood mapping. By synergistically integrating data from LEO multispectral (VIIRS), GEO multispectral (AHI), and passive microwave (AMSR2) sensors, our method overcomes the individual limitations of infrequent revisits and cloud cover. The framework harmonizes these disparate sources using surface water fraction as a common variable and downscales them with a hydrologically informed susceptibility model to generate 30 m flood maps every two days. The evaluation of the February 2022 Brisbane flood, using an independent ground truth dataset, confirmed the framework’s effectiveness. It successfully produced a comprehensive map of the flood’s peak and outperformed the cloud-contaminated CEMS map, demonstrating its utility as a resilient monitoring tool when other data sources are unavailable.

Although the evaluation in Brisbane, Australia, was successful, establishment of the framework’s global transferability requires testing in more diverse environments. This is especially true for regions where common high-resolution flood mapping sensors are already known to fail. In Southeast Asia, for instance, dense vegetation severely limits the utility of Sentinel-1 SAR, contributing to extensive exclusion masks where floodwater cannot be appropriately mapped [64]. This framework, which is independent of SAR, is theoretically well-suited to provide an alternative in these challenging regions.

Future works will include the integration of other sensors to continue reducing the temporal gaps between flood observations and enhancing the flood map quality. The framework’s reliance on a common variable, the surface water fraction, makes it highly adaptable. Any sensor capable of producing this can be integrated into the procedure. A prime candidate is Global Navigation Satellite System-Reflectometry (GNSS-R) from missions like CYGNSS, which provides all-weather observations sensitive to flood inundation, even beneath vegetation. Furthermore, data from other multispectral sensors, such as Sentinel-2 and Sentinel-3, could further increase the density of observations and the overall robustness of the final flood product. Finally, the adaptation of the framework to other geostationary sensors, such as from the GOES-R or MTG, could make the framework truly global.