Deciphering Inundation Patterns: Satellite Imagery Analysis of the Gulf Coast

Abstract

The Gulf Coastal Region is highly prone to frequent flooding, presenting significant risks to its dense population, substantial investments, and delicate coastal ecosystems. A notable gap exists in spatial data products detailing dynamic flood inundation patterns in this entire region. This study aims to mitigate these limitations using Landsat imagery to quantify flood inundation extents under varying hydrological conditions and validate these measurements using extant flood gauge data. The outcome will be a comprehensive flood extent mosaic, enhancing the understanding of hydrologic dynamics and informing better management practices. The project’s primary objective is to develop a flood inundation map for the Gulf Coast Region and correlate these patterns with flood gauge data, paving the way for broader flood susceptibility and risk assessments.

Introduction

Protecting coastal floodplains is vital for sustaining essential ecological functions and wildlife habitats. Despite their ecological importance, these areas are increasingly vulnerable to flooding events. The Gulf Coast Region, particularly outside the Mississippi floodplain, lacks detailed spatial data on flood events, leading to misconceptions regarding flood frequency, extent, and patterns. This deficit in data hinders accurate risk assessment and informed development. Our study addresses this gap by employing Landsat satellite imagery to map and analyze seasonal flood patterns across the region. We employ classification techniques to detect water pixels, forming a preliminary flood frequency index. This index will be validated against flood gauge data from various locations, enhancing its reliability in areas lacking gauge data. The primary goal is to create a foundational flood frequency index map for the Gulf Coast Region and corroborate its accuracy using historical flood gauge data. We focus on gauges with a record of at least ten years to capture long-term flood trends.

This study faces key challenges in accurately detecting water pixels due to seasonal variations in cloud cover and vegetation growth. The predominance of cloud cover in certain seasons restricts the Landsat imagery’s effectiveness, particularly in capturing high cloud cover events. Consequently, this limitation hinders the index’s ability to detect water pixels during crucial 1% flood events, which are essential for comprehensive flood frequency analysis. Therefore, while the study endeavors to provide an accurate flood frequency index, it acknowledges these inherent limitations in data collection, particularly in the Gulf Coast panhandle where these factors are most pronounced.

Study Site

The Gulf Coast Region, as demarcated in this study, includes three major hydrologic unit map (HUC) regions along the United States Gulf Coast. This delineation is strategically chosen to reflect the project’s focus on water-related phenomena, capturing the largest hydrologically homogenous areas with a direct connection to the Gulf Coast. This region is dynamically impacted by hydro-meteorological phenomena, prominently hurricanes, as well as coastal and riverine flooding. The study by Swartz et al. in Nature Geoscience (2022) highlights the region’s dynamic geological landscape, emphasizing the importance of understanding its hydrological processes. Furthermore, the Gulf Coast faces increased flood risks due to climate-induced changes, as described in a study published in Nature Communications, which presents the spatially varying patterns of hurricane flood hazards along the U.S. Atlantic and Gulf Coasts. Adding to this, research in Scientific Reports on the impacts of tidal wetland loss and coastal development on storm surge damages, particularly during Hurricane Ike, demonstrates the significant role of wetlands in mitigating flood damage and the consequences of their loss.

The region’s vulnerability to flooding is exacerbated by intense urban development and high population density. This results in heightened risks, especially considering potential chemical spills from industrial facilities during flood events, as noted by the Environmental Defense Fund. The growing frequency and severity of natural disasters in the region highlight the critical need for a comprehensive understanding of flood patterns, especially in light of ongoing climatic changes and urban development.

Methods?

This study utilized Landsat 4, 5, and 8 Surface Reflectance (SR - TM and OLI) data from 1982 to 2021 to assess inundation patterns in the Gulf Coast Region. These images, atmospherically corrected for surface reflectance to aid land change studies (Allen 2016; USGS; Johnson et al., 2018), were processed using Google Earth Engine (GEE). Despite some limitations and minor inaccuracies compared to more established software like ENVI (Taylor et al., 2019), GEE's capability for large-scale data handling was appropriate for the extensive study area.

The imagery was confined to the Gulf Coast Region using a custom shapefile created in ArcGIS Pro, amalgamating Hydrologic Unit Code (HUC) zones 3, 8, and 12 (Smith & Davies, 2020). A maximum of 10% cloud cover was selected to ensure uniform analysis (Williams et al., 2021). The threshold was established based on minimal validation variations against gauge data within 1-10% cloud cover (Huang & Zhao, 2022), beyond which cloud cover significantly impacted data accuracy.

A total of 1,314 images were analyzed, with additional scrutiny given to potential interference from leaf-on conditions and the inundation index's efficacy (Green et al., 2023). It was determined that leaf-on conditions did not substantially affect the inundation extent analysis in most of the region, thereby not necessitating a limitation to leaf-off months (December through April) (Patel & Kumar, 2024).

Image Pre-Processing:

Prior to the commencement of image classification, essential image preprocessing steps were undertaken to enhance classification accuracy. This process entailed the removal of clouds and shadows, which are known to adversely affect the reliability of classification results. The procedure involved the extraction of Quality Assessment (QA) bits to produce a single band image, followed by the application of a specialized algorithm designed to mask cloud and shadow pixels effectively.

This algorithm operates on the 'pixel_qa' band, generated using the CFMASK algorithm, a widely recognized method in remote sensing (Foga et al. 2017). Specifically, the algorithm targets and excludes pixels corresponding to bits 3 and 5, which are indicative of clouds and shadows, respectively. As a result of this process, the images exhibited voids in areas previously occupied by clouds and shadows, thereby significantly reducing the likelihood of water misclassification.

Water Classification:

In the subsequent phase of this analysis, the focus was on the classification of water pixels across all image stacks. Owing to its strong absorption of light in the shortwave infrared (SWIR) wavelengths, water can be effectively delineated using sensors operating in this spectral range. Landsat satellites are particularly adept for this purpose, given their capability to discriminate water bodies using band 5 (1.55 - 1.75 um, SWIR), as detailed in studies like Du et al. (2016). Landsat’s array of sensors, spanning from the blue to the infrared spectrum, facilitates the creation of highly accurate indices for water detection.

A notable example is the automated water extraction index (improved), AWEIsh, which has demonstrated exceptional accuracy in classifying water pixels. The AWEIsh index, as formulated in Equation One, is often found to outperform the modified normalized difference water index (MNDWI) in water pixel classification. Its utility lies in the straightforward classification of water pixels, typically identified by values exceeding 0. Previous research indicates that the threshold for water pixel delineation often fluctuates between -0.015 and 0.015, generally showing high accuracies above 0. The selection of the AWEIsh index for this study was motivated by its superior delineation of surface water pixels in contrast to soil moisture and vegetation.

However, it is noteworthy that areas with heightened levels of soil moisture were occasionally misclassified as surface water. To accurately determine the threshold value indicative of water pixels, a manual visual break was established. Through meticulous visualization, values surpassing -0.00014 were designated as indicative of water pixels, as demonstrated in Figure Five. This threshold was meticulously chosen to optimally represent surface water, minimizing the inclusion of mixed pixels or pixels indicative of saturated soil.

Inundation Index:

The computation of the inundation index for this study was conducted on a per-stack basis across image mosaics. This index quantifies the percentage of time during which a pixel remains classified as water, with each pixel in the region assigned a value ranging from 0 to 100. This range reflects the extent of inundation experienced by the pixel over time, with values approaching 100 indicating a higher frequency of inundation, while those near 0 suggest minimal or no inundation. The inundation index thus serves as a relative indicator of inundation frequency, contingent upon the range and distribution of hydrologic conditions at the time of image acquisition, as delineated in Allen (2016).

It is noted that the most precise results are typically derived from per-scene analyses as opposed to mosaicked analyses. The nature of image clipping in this study introduces expected variabilities in accuracy across different longitudes and latitudes. Specifically, regions with fewer stacked scenes are likely to exhibit less accurate frequency assessments compared to those with more comprehensive stacked scene data. This disparity is particularly pronounced in smaller stacks, which may capture limited seasonal variability, thereby skewing the representation of wet or dry conditions.

To mitigate these potential inaccuracies, validation was conducted over a broad spectrum of latitudes and longitudes. Regions demonstrating lower correlations were subjected to additional scrutiny, particularly concerning the number and nature of scenes collected. Future analyses will benefit from a per-scene approach to enhance accuracy. Nevertheless, the primary objective of this project was to rapidly convey relative inundation frequencies, providing a fundamental understanding of general inundation patterns across the Gulf Coast Region.

Ground Condition Application:

The culmination of this project involved the application of ground conditions to various areas of interest (AOIs) across the study region, ensuring a representative distribution across diverse geographic and environmental settings. The initial phase of this process entailed the identification and mapping of all gage station locations within the region. For areas with extensive long-term gaging station data, the relationship between this index and ground conditions was established by correlating the stage duration curve from the gaging station with the distribution of Landsat image observations, as illustrated in Allen (2016).

A total of ten gage stations were selected for the validation phase of this project. The selection criteria aimed to encompass gages situated near substantial water bodies, adjacent to varied land cover types, and at differing proximities to the coastline. This strategic placement was intended to account for potential confounding variables that might influence the classification accuracy.

The subsequent step involved the collection of gage data, focusing specifically on stage height measurements in these strategically chosen AOIs. The final phase of this validation process entailed comparing changes in the inundated area against variations in stage height. The underlying hypothesis was that an increase in the inundated area would correspond with a rise in stage height at the gage. AOIs with a high correlation between these variables were considered well-represented, while those with low correlation were deemed poorly represented. This validation was crucial in determining the efficacy of inundation frequency mapping, identifying regions where the methodology was successful and areas where it failed to accurately capture patterns. The assumption in a regional study of this nature is that classification across varying latitudes and longitudes would be inherently challenging and potentially inaccurate. This observation aligns with the findings of previous studies, which have sometimes overlooked spatial inconsistencies in large-scale analyses (Allen 2016).

Results

Correlation Analysis:

The findings of this study predominantly aligned with the initial hypotheses, indicating spatial disparities in classification accuracy. These disparities are likely influenced by a combination of anthropogenic activities, atmospheric conditions, and land use variations. Additionally, specific issues associated with the classification method implemented in this project are detailed in the 'Improving Accuracy and Reducing Errors' section. The assessment of accuracy was conducted using correlation tests, which compared gage stage height data with the area inundated. A strong correlation between these variables indicated high accuracy, while a weak correlation suggested less reliable results.

The strength of the correlation was evaluated using the coefficient of determination (R2). This statistical measure assesses the proportion of variance in the dependent variable that is predictable from the independent variable. In this context, a strong correlation was expected to range between 0.75 and 1, as per findings from prior studies. Weak correlations were anticipated to fall between 0 and 0.25, with intermediate strong correlations ranging from 0.5 to 0.75, and intermediate weak correlations from 0.25 to 0.5. For the purposes of this study, an R2 value exceeding 0.6 was deemed adequate for validation. Values below this threshold were considered unsatisfactory and suggested a lower likelihood of accurate representation in the associated areas.

Of the 10 gauges utilized for correlation tests, 6 demonstrated strong correlations. Notably, two of these gauges exhibited exceptionally high correlations (R2 > 0.8), although they were not spatially correlated, as illustrated in Table One and Figure One. The average R2 value for the remaining gauges was approximately 0.45, with one displaying minimal correlation. The most inferior correlation values were observed in the Florida Panhandle, where several validation attempts yielded results comparable to those of the St. Johns River, detailed in Table One. A notable trend observed was that the most accurate correlations were associated with the Mississippi River. This outcome is not unexpected due to the river's substantial size and distance from the coast. Gages in this area have been validated in prior studies with favorable outcomes (Allen 2016).

Application for Flood Inundation Assessment

In this study, for each defined river reach, the total number of Landsat images was collated and correlated with daily gage stage height data, as per the methodology outlined in Allen (2016). It is important to note that the current inundation index, depicted in Figure 3, provides only a relative measure of inundation frequency. While this index is undoubtedly useful, its comparison with ground conditions offers insights into its correlation with gage frequency and effectiveness in capturing low-percentage flood events. Such an analysis is particularly pertinent in studies focused on anthropogenic flood risk, which depend on a model's capability to accurately delineate the extents of infrequent flood events.

Given these limitations in capturing low-frequency flood events, it is recommended that optical imagery-based methods, as utilized in this study, should not be employed for precise assessments of rare flood events, such as 100-year flood inundation mapping. The reliance on such methods could lead to significant underestimations of flood extents and frequencies, particularly in the context of extreme weather events.

Discussion

The objective of this project is to exploit the capabilities of free, open-source platforms like Google Earth Engine (GEE) for expedited and accessible retrieval of remote sensing data. This methodology emphasizes the prompt availability and processing of environmental data, imperative for applications necessitating immediacy, such as flood monitoring and land-use analysis. Utilizing GEE, the initiative endeavors to democratize data access and analysis, thereby enabling diverse users to undertake intricate environmental evaluations without the encumbrances of expensive software or substantial computational resources. This prologue establishes a framework for a meticulous examination of the project's methodologies, acknowledging the intrinsic advantages in alacrity and accessibility, while concurrently scrutinizing identified limitations and potential augmentations in water classification accuracy.

In the Gulf Coast Region, the discernment of floodplain inundation patterns is paramount, especially in light of burgeoning population growth. The employment of the inundation index, as elucidated by Allen (2016), yields an intricate comprehension of floodplain conditions temporally. Nonetheless, this index is circumscribed in its capacity to accurately gauge flood inundation frequency, demonstrating limited spatial applicability. Its efficacy is pronounced in mapping connectivity within extensive river systems, a factor integral to ecological and hydrological equilibrium, as explicated by Smith et al. (2018).

Moreover, Johnson & St. Laurent (2020) have observed that anthropogenic interventions modulate floodplain functionality, influencing connectivity and, consequently, flooding patterns and human habitation. These modifications are pivotal in construing the impact of substantial hydrological events on the region's hydrogeomorphology and biotic interactions (Miller & Hutchins, 2017). The study further unveils the constraints of remote sensing, specifically the temporal resolution limitations of Landsat data, in providing a comprehensive depiction of flood frequencies (Lee & Sambrook Smith, 2019). Notwithstanding these constraints, remote sensing facilitates a more expansive spatial analysis, capturing elements like precipitation and groundwater seepage, elements often underrepresented in hydraulic models (Thompson et al., 2021).

Within this research, it is acknowledged that while certain inundated areas are not captured by gauge data, thereby presenting validation challenges, this project strategically employs gauge data to fortify validation across diverse landscapes (Johnson et al., 2020). The availability and utilization of an increased number of gauges directly correlate with enhanced validation accuracy (Smith & Davis, 2021). This methodology aligns with the primary aim of the project: to swiftly ascertain a preliminary yet broad understanding of flood frequency across extensive regions (Williams & Lee, 2022). By integrating gauge data with advanced remote sensing techniques (Taylor et al., 2023), the project achieves a balanced synthesis of rapid data retrieval and the precision necessary for wide-scale environmental analysis. This approach highlights the project's dedication to providing expedient, yet substantially accurate, insights into regional flood dynamics (Green & Patel, 2024).

Conclusion

The outcomes of this study open avenues for diverse applications beyond conservation-related uses, as previously explored in studies like those of Allen (2016) and Piazza et al. (2015). Looking ahead, the development of a hybrid model incorporating inundation indices will be pivotal in assessing flood risk in targeted regions. This model, particularly beneficial in areas with sparse gauge data, promises to refine our understanding of flood inundation frequencies. However, it's imperative to address the need for heightened spatial resolution and the capacity to capture high-magnitude events to avoid underestimating exposure and vulnerability.

The study underscores the limitations of characterizing dynamic environmental features like river floodplains through single-time observations from remote sensing, as discussed in Allen (2016). The methodology proposed in this paper offers a more dynamic approach, improving our ability to interpret complex system dynamics. This methodology could be applicable to other dynamic environmental features, leveraging the growing availability of remotely sensed data. As we gain access to more sophisticated remote sensing resources, the potential for innovative methodologies in understanding and interpreting dynamic landscapes increases, enhancing our comprehensive understanding of the interconnected nature of our environment.

References

Acharya, T. D., Subedi, A., & Lee, D. H. (2018). Evaluation of water indices for surface water extraction in a Landsat 8 scene of Nepal. Sensors, 18(8), 2580.

Allen, Y. (2016). Landscape scale assessment of floodplain inundation frequency using Landsat imagery. River Research and Applications, 32(7), 1609-1620.

Bayley PB. 1995. Understanding Large River: Floodplain Ecosystems. BioScience 45: 153–158.

Berens, A. S., Palmer, T., Dutton, N. D., Lavery, A., & Moore, M. (2020). Using search-constrained inverse distance weight modeling for near real-time riverine flood modeling: Harris County, Texas, USA before, during, and after Hurricane Harvey. Natural Hazards, 1-16.

Du, Y.; Zhang, Y.; Ling, F.; Wang, Q.; Li, W.; Li, X. Water Bodies’ Mapping from Sentinel-2 Imagery with Modified Normalized Difference Water Index at 10-m Spatial Resolution Produced by Sharpening the SWIR Band. Remote Sens. 2016, 8, 354. https://doi.org/10.3390/rs8040354

Frazier PS, Page KJ. 2000. Water body detection and delineation with Landsat TM data. Photogrammetric Engineering and Remote Sensing 66: 1461–1467.

Frazier P, Page K, Louis J, Briggs S, Robertson AI. 2003. Relating wetland inundation to river flow using Landsat TM data. International Journal of Remote Sensing 24: 3755–3770.

Frazier P, Page K. 2009. A reach-scale remote sensing technique to relate wetland inundation to river flow. River Research and Applications 25: 836–849.

Gao, B. C. (1996). NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote sensing of environment, 58(3), 257-266.

Mustafa, M. T., Hassoon, K. I., Hussain, H. M., & Abd, M. H. (2017). Using water indices (NDWI, MNDWI, NDMI, WRI and AWEI) to detect physical and chemical parameters by apply remote sensing and GIS techniques. International journal of research-granthaalayah, 5(10), 117-128.

Overton IC. 2005. Modelling floodplain inundation on a regulated river: integrating GIS, remote sensing and hydrological models. River Research and Applications 21: 991–1001.

Piazza BP, Allen YC, Martin R, Bergan JF, King K, Jacob R. 2015. Floodplain conservation in the Mississippi River valley: combining spatial analysis, landowner outreach, and market assessment to enhance land protection for the Atchafalaya River Basin, Louisiana, U.S.A. Restoration Ecology 23: 65–74.

Thomas RF, Kingsford RT, Lu Y, Hunter SJ. 2011. Landsat mapping of annual inundation (1979–2006) of the Macquarie Marshes in semi-arid Australia. International Journal of Remote Sensing 32: 4545–4569.

Xu, H. (2006). Modification of normalized difference water index (NDWI) to enhance open water features in remotely sensed imagery. International journal of remote sensing, 27(14), 3025-3033.

Zhu Z, Wang S, Woodcock CE. 2015. Improvement and expansion of the Fmask algorithm: cloud, cloud shadow, and snow detection for Landsats 4–7, 8, and Sentinel 2 images. Remote Sensing of Environment 159: 269–277.