Improving flood monitoring in rural areas using remote sensing

Australia's regions (non-metropolitan areas) play an important role in the economic growth of the nation (Australian Government 2018), producing 65% of the country's export income and employing over a third of the workforce. However, rural and regional Australia is vulnerable to the impacts of climate change and associated extreme weather events, particularly flooding (Australian Government 2018).

Flooding is generally the most devastating type of natural disaster in Australia, affecting people, the environment, and regional economies severely (Gentle et al. 2001). Levees and dams are the main structural flood control measures in Australia, and most levees were constructed many years ago in rural and agricultural areas to protect farms. Today, almost 98% of levees in Victoria are rural and in varying states of disrepair (Edwards 2013). However, due to population growth and regional development, they now also protect urban areas. This creates a new hazard to rural towns, because flood control relies on levees not necessarily designed to cope with the volume and frequency of flooding they now face in a changing climate. This will increase the potential consequences of failure, even though the likelihood of failure does not change (Edwards 2013).

Australian rural levees are not as effective as their overseas counterparts (Smith et al. 2014; SGSC 2018). Traditionally, rural levees were constructed simply by accumulating soil to act as a barrier to rising water and most were not designed to specific engineering standards. They generally exhibit a propensity to be overtopped (Smith et al. 2014). However, the aging Australian levees generally provide a lower level of protection (less than 10% Annual Exceedance Probability or AEP) than their US and European counterparts (Smith et al. 2014; SGSC 2018). Improving their protection capacity to the 1% AEP level will be difficult both economically and practically (Smith et al. 2014; SGSC 2018).

Climate change brings with it associated increases in the frequency of extreme weather and greater likelihood of flooding. Evaluation of the possible extent of flooding, and damage to individual's lives and assets is essential, to provide local authorities with information to support policies that reduce flood vulnerability and improve recovery responses.

Understanding the effects of flooding on the natural and built environments requires spatial and temporal data relating to flood occurrence. Traditionally, stream gauges mounted along catchments provide water level and discharge information for flood analysis. However, this is both time-consuming and labour-intensive, and thus not cost-effective for large areas (Northcott et al. 2000). In addition, because of topographic variation, the recorded data could be imprecise. Stream gauge distribution could also affect the possibility of accurate recording of the magnitude and impacts of flooding (Sanyal & Lu 2004; Dao et al. 2015). In such situations, the true magnitude and/or extent of flooding may be either under- or over-estimated.

Scientific improvements in satellite image analysis potentially provide an effective and relatively inexpensive solution. Satellite imagery allows the capture of timely and detailed information about inundation, and at different scales (Singh et al. 2015). Satellite-based remote sensing has been used to visualize the extent of flood inundation since the early 1970s (Deutsch et al. 1973; Rango & Salomonson 1974). However, it could also be used as a monitoring tool to assess finer-scale flooding (such as the failure of individual levees) or to detect the early stages of a flood event. Remote sensing has enormous potential in flood management due to its ability to acquire accurate and consistent data, at low cost and over large areas (Khan 2005; Serpico et al. 2012). It is one of the fastest and most cost-effective methods for observing and providing flood event information. Compared with older-style surface data – e.g., weather forecasting records and/or gauging station data – remote sensing data from different sensors can satisfy the wide global requirements of flood monitoring (Bates 2004; Sanyal & Lu 2004).

In this study, Landsat optical imagery were processed to try to detect the extension of inundated parts of the Tarwin catchment, Victoria, Australia, during winter 2016. Landsat images, provided free-of-charge by NASA, are good candidates for remote mapping studies for large-scale floods. By integrating the assessment of levee condition in the Tarwin catchment, the research outcomes identify the flood-prone areas of the Tarwin and report on the inefficiency of flood control measures in this regional part of Australia.

The study area is part of the Tarwin River in South Gippsland. The Tarwin is the main river in South Gippsland Shire and its catchment, covering about 1,500 km², is mainly rural with small residential pockets. Figure 1 shows the location of the river, as well as the locations of the two nearest hydrologic stations (Koonwarra and Meeniyan) used for the flood study in the Tarwin catchment.

The region and study area are frequently impacted by heavy rainfall events in the wetter season. Figure 2 illustrates the monthly total rainfall for the period of 2010–2015 for the Tarwin catchment.

During wet periods, runoff inundates the low-lying areas and leads to flooding. The extent of inundated areas has been determined manually in the past using stream gauge measurements (rainfall, streamflow, and tide and sea level data) and flood hydrograph modelling (Water Technology 2007 ). In this study, digital classification was used for spatial mapping of water features.

Since the recognition of water features during the heavy rainfall events needs good spectral resolution, Landsat images are a reliable option. Two Landsat-8 OLI images (Level-2 product) were captured on 25 April 2016, before a heavy rainfall event, and 28 June 2016, after it (Figures 3(b) and 3(c)), and processed to predict the extent of runoff. Several heavy rainfall events occurred during winter 2016 (24/25 June, 6 and 23 July, and 2 August, with 16 mm/d average rainfall). Both the cloud cover and date of available images were important factors in choosing the year, as the State of Victoria is mostly cloudy during winter (June to August), and 2016 was the best amongst many.

In order to distinguish remotely-sensed surface water as inundation due to levee overtopping rather than arising from rainfall accumulation, two time-periods were assessed. Landsat images of the study site from June 2015 (Figure 3(a)) and June 2016 were compared. Stream gauge data indicated that levees had not been overtopped in June 2015. By contrast, in June 2016 the same area experienced significant floods and stream gauge data indicated that levees had been overtopped. As such, it was assumed that any surface water shown in the 2015 images was accumulated rainfall, so they were used as a ‘negative control’ to test the approach. Likewise it was assumed that surface water in the 2016 images arose from levee overtopping.

Landsat-8 OLI images comprise eight multispectral bands with a spatial resolution of 30 m and one 15-m panchromatic band. The radiometric resolution is 16 bits, upgraded from previous Landsat sensors. The grayscale over-saturation in extremely dark regions can be avoided efficiently and the subtle features of water with very low reflectivity distinguished by enhancing the images' radiometric resolution (Gao et al. 2016).

The images were georeferenced to UTM projection, zone 55. The study area is on one Landsat image, so there is no need to mosaic images together, which is usually accompanied by normalizing to a master image (Ghofrani et al. 2014). As the Tarwin catchment has smooth topographic elevations, topological correction is not needed.

The Tarwin deltas, nearby, include a broad system of levees along the lower part of the catchment with a crest height of approximately 3 m Australian Height Datum (AHD). The levees were generally constructed by private landowners, as a result of a ‘Drainage Area’ declaration, between the early 1940s and the late 1960s, primarily as a means of controlling flooding (SGSC 2018).

Although the levees are intended to protect farmland and people, most are inadequate for the purpose. The Tarwin Lower levee systems can protect the region from small floods (nuisance flooding) and storm surges, but are generally over-topped by large floods (1% AEP events) and large storm surges, and cannot prevent inundation. As the levees were not planned to hold back the volume of water faced, it is not surprising that they are ineffective in coping with flooding impacts. The existing structures are old (built between 1940 and 1980), and there is little evidence that they have been maintained well (Water Technology 2007; Edwards 2013).

South Gippsland Shire Council commissioned Water Technology (a water infrastructure consultancy firm) to undertake a condition assessment study of its coastal levees. This included a field survey of points of weakness (POW) via a desktop GIS analysis to identify potential POW using LiDAR survey and aerial photos. POWs are locations along a levee that do not offer full protection. They are visible features affecting the levee structure or shape that may reduce its performance. They include erosion and proximity to the river; tracks and pipes traversing levees; the presence of vegetation such as saplings and tree regeneration; and animal burrows.

The survey revealed that most of the levees are in poor to fair condition and seem to have been poorly maintained since construction. Almost half of all POWs identified appear to be caused by wombat burrows (SGSC 2018). Six of the assessed levees are within the Tarwin Lower study region (identified by ‘Levee ID’ numbers and reported in Table 1) as shown in Figure 4. The condition of each levee was rated between good and very poor on the basis of the field-based evaluation – Figure 4 and Table 1 (SGSC 2018).

In their study, Water Technology evaluated the level of protection offered by levees in the Tarwin catchment (SGSC 2018). The levee crests were compared, in their existing condition, to various storm-tide events to determine the circumstances under which the levee might be overtopped.

The levee protection level results for the Tarwin Lower are presented in Figure 5 for both the 10% and 1% AEP storm tide events, and for both current and predicted 2100 (+0.8 m) sea level rise (SLR). The figure shows the levee lengths (km) above (protected) and below (unprotected) the relevant storm tide levels.

The maximum daily river levels from two hydrologic stations (Koonwarra and Meeniyan) around Tarwin River were obtained from the Bureau of Meteorology (Bureau of Meteorology 2018) for June 2015 (orange) and 2016 (blue) – see Figure 6. These show that river level in the Tarwin exceeded 3 m (levee height) in June 2016 but not June 2015.

To evaluate the method's accuracy in this study, GeoEye high-resolution images were used for 16 August 2016 (the closest available high-resolution image to the heavy rainfall events). The images were already mosaicked, taken from 680 km above ground level, with a spatial resolution of 49 cm.

Mapping of water areas is achieved using digital classification with the normalized difference water index (NDWI) and modified NDWI (MNDWI), accompanied by online visual interpretation. Both NDWI and MNDWI highlight water features (Xu 2006, 2007).

Figure 7 illustrates the spectral reflectance profiles of three land cover categories (water, crops and built-up areas), from a case study test area. All bands in Figure 7 have been rescaled to [0–1].

The study area is rural and covered mainly with croplands, water bodies and small residential zones. It is clear from Figure 7 that the reflectance of built-up land in the green (OLI 3) and NIR (OLI 5) bands differs from that of water in rural parts of the Tarwin catchment. Water features reflect near infrared light less than green light, the reverse of what is seen in built-up areas in rural regions. Consequently, calculation of NDWI and MNDWI generates positive values for water, and negative values for crops and built-up areas.

The classification of Landsat images using NDWI and MNDWI indices is illustrated in Figure 8. The left hand set/column – Figures 8(a), 8(c), 8(e), 8(g), 8(i) and 8(k) – shows alternating NDWI and MNDWI images for April 2016 (8a & c), June 2016 (8e & g), and June 2015 (8i & k). Water features appear as white tones as a result of the NDWI and MNDWI analyses.

Subsequently, a threshold value of zero was applied to extract water features from both the NDWI and MNDWI images. Water and non-water regions were identified in the NDWI and MNDWI images using unsupervised classification, based on index values and classified image tones. Figures 8(b), 8(d), 8(f), 8(h), 8(j) and 8(l) show water bodies extracted, alternately, from the NDWI and MNDWI images, respectively, for April 2016, June 2016, and June 2015.

Comparison of the extracted water features in the April 2016 (autumn) (Figures 8(b) and 8(d)) and June 2016 (winter) images (Figures 8(f) and 8(h)) shows an increase in surface water coverage during the winter. Similar comparison of the June 2015 (Figures 8(j) and 8(l)) and June 2016 (Figures 8(f) and 8(h)) images shows the difference between surface water as inundation arising from rainfall accumulation (2015) and/or levee overtopping (2016). As noted previously, the water level data for 2015 and 2016 (Figure 6) confirmed that the river stage exceeded 3 m (levee height) in June 2016 with extensive inundation due to levee overtopping, but that there was no overtopping in 2015.

Cloud cover in the June 2015 Landsat image was not minimal but this did not affect the analysis because clouds were mostly prevalent over residential areas (lower left corners of Figures 3(a) and 8(j)) and cloud cover along the river was negligible (Figures 8(j) and 8(l)).

The classification result of Landsat June 2016 image was evaluated by comparing it with that from a high-resolution GeoEye image acquired in August 2016 over cloud-free sites. The confusion matrix was computed qualitatively to assess the accuracy of the result generated by the approach. The result is presented in Table 2.

Both NDWI and MNDWI performed well, but NDWI identified water features more efficiently in the catchment's rural areas with 96.04% accuracy (Table 2). This could be because of NDWI's larger standard deviation, which allows it to contain more information that can be used to establish high-contrast areas for detection (Table 3). The contrast between crops and water is also higher in the NIR than the MIR band (Figure 7), so, the greater difference between the contrasts of water and crops in NDWI, makes it more efficient in rural areas where crops dominate land-use.

Analysis of the Landsat images provided good delineation of the extent of inundation in the Tarwin catchment. Water extraction from Landsat images of the study area for winter 2016 shows that levees 15, 16, 21, 22 and parts of 17 and 30 were overtopped during rainfall (Figures 4, 8(f) and 8(h)). This confirms Water Technology's results concerning the level of protection provided by the Tarwin Lower levees (SGSC 2018). Based on these results (Figure 5) and current condition, about 16 kilometres of the levee system can protect the river's lower region from a 10% AEP event, while about 7 kilometres cannot. Equally, only 7 of 23 kilometres can protect the region from a 1% AEP event. Projections suggest that, in 2100, only 2 kilometres of the levee system will be able to protect the region from a 10% AEP event and only 0.8 kilometres from a 1% event.

The potential consequences of failure were determined by considering the possible residential regions, main roads, and farmlands under threat of inundation. In winter 2016, the inundated farms became unusable, and flooding due to over-topping of levee 17 posed a substantial threat to residential areas, and public roads and assets, while over-topping of levee 16 posed a threat to a major public road. Otherwise, it was primarily private farmland that was at risk from more frequent and potentially permanent inundation.

The results of the Landsat image analysis and assessment of the condition of structural measures in the Tarwin catchment support the notion that no seawall or levee system can provide enough protection from flooding under all conditions.

To mitigate the impacts of ineffective levees, either repairing or upgrading the levee systems can be considered. Repairing the failure points and maintaining current levee height, could provide short-term benefit in some places, but the large number of wombat holes and other potential points of weakness found along most levees suggests that a significant amount of repair work would be involved. Furthermore, at current crest heights, most levees are likely to be overtopped frequently by 2100, anyway.

Upgrading levees involves significant cost because they are fairly long. Moreover, the levees are often constructed to different heights due to the different protection priorities for residential and agricultural lands, and they are often maintained in segments. Ad hoc height changes for these segments may exacerbate flood conditions, as alterations result in complicated feedback loops in the depth and velocity of stormwater that are difficult to predict (Sanyal 2017).

The extensive implications without significant benefits in applying either solution, make them short-term. In order to build safer and more resilient rural communities, a new flood mitigation approach is needed for the Tarwin catchment, which takes into account the potential impacts of climate change.

Transformational rather than incremental change would seem a better solution when adapting to climate change, etc, with respect to the Tarwin catchment levees. Moving away from the traditional levee building and upgrading approach, towards one that provides multiple functions and allows more room for the rivers by re-engaging floodplains, enlarging flood channels, and allowing flooding of some low-lying areas, could increase the level of flood protection tenfold and have a planning horizon of more than two centuries (de Groot & de Groot 2009). This so-called Dutch approach has already proved to be one of the most effective solutions to the potential effects of climate change, moving from traditional structural reinforcements toward ‘room for river’ measures to enhance flood protection and to meet other societal goals (de Groot & de Groot 2009; Ghofrani et al. 2016). It provides sustainable underpinning for development that protects against floods while supporting biodiversity, agriculture and the recreational use of water resources (Ghofrani et al. 2016) at a low cost compared to conventional systems (Joksimovic & Alam 2014).

This study confirms that remote sensing can be used as a useful tool to assess flooding in rural areas, and also supports the need for a new solution for flood control and climate change impact mitigation in rural and regional Victoria, Australia. Using Landsat satellite images acquired in winter 2016, levee overtopping has been detected with image processing and remote sensing technology. The levee condition assessment of the Tarwin Lower region has been supplemented and this has confirmed that the levee system is not working properly. Straightforward and sustainable solutions must be considered and applied promptly, to prevent the irreversible damage that could occur because of climate change.

The authors thank Mr Geoffrey Davis (Assets Planning Engineer at South Gippsland Shire Council) who provided information from a peer reviewed draft of the ‘Assessment of Coastal Levees Study’ by Water Technology, Bureau of Meteorology, and to West Gippsland Catchment Management Authority for providing the necessary data.