**1. Introduction**

Because of its mostly hot, dry, and erratic climate, wildfires in Australia and many parts of the world are frequently occurring events during the hotter months. Between 2017 and 2019, severe drought developed across much of eastern and inland Australia including Queensland, New South Wales (NSW), and Victoria, also extending into parts of South Australia and Western Australia. As at late 2019, many regions of Australia were still in significant drought, contributing to water restrictions

and extreme fire conditions [1]. The 2019–20 Australian bushfire season, colloquially known as the "Black Summer," began with several serious uncontrolled fires in June 2019, an early start to the wildfire season as it normally starts in early October in NSW [2]. Throughout late 2019 and early 2020 these wildfires both multiplied and combined, to create mega fires that burnt predominantly throughout the southeast of the country, peaking during December–January, and having been since contained and/or extinguished [2–4].

In NSW, the three-year (2017–2019) severe drought had left forests tinder dry, facilitating rapid expansion of fires across the state. Sydney—the nation's most populated city (about 5 million residents)—had been under water restrictions since late 2019, when its dams fell below 45% capacity. The wildfires in NSW between September 2019 to January 2020 (the 2019-20 fires) were unprecedented in their extent and intensity [4]. The NSW Rural Fire Service (RFS) reported that the 2019–2020 fires burnt 5.4 million hectares (including ~11,264 bush or grass fires across 6.7% of the State) and destroyed 2439 homes [3].

Wildfires not only create risk to lives, infrastructure, and properties, but also cause land and ecosystem degradation through increased soil erosion. The subsequent sediment deposition in rivers, lakes, and reservoirs is of great concern for drinking water quality, aquatic habitat, and environmental degradation. When large rainfall events occur in a short period of time, runoff will wash a lot of ash and sediment into waterways and dams. Rain also transports other contaminants, such as building debris, dead animals, and pollutants from fire retardant. Soil sediment, ash, and other contaminants pose a hazard to human health when mobilized in drinking water catchments. It is therefore necessary to quantitatively estimate soil erosion after severe wildfires in order to assess the extent and magnitude of post-fire soil erosion risk and the effectiveness of any rehabilitation or mitigation actions [5].

The Bureau of Meteorology (BoM) recorded 51 mm rainfall in January 2020 and 337 mm in February 2020 over the Sydney drinking water catchment (SDWC) areas. The rainfall events (especially in February) were widespread and helped extinguish the wildfires. However, the rainfall also caused severe runoff and hillslope erosion. As rainfall intensity is the leading agent contributing to greater erosion rates, and the subsequent effects on water quality, it is necessary to obtain timely rainfall information for fire recovery and erosion control practices. The rapid assessment of these risks and timely mitigation actions can significantly reduce risks to public safety, infrastructure, and the environment [5].

Hillslope erosion (including sheet and rill erosion) is the major form of water erosion and the dominant source of sediment in waterways in Australia and many parts of the world [5–7]. Monitoring of hillslope erosion after wildfires may be by direct measurement, such as flumes and sediment traps, or from tipping bucket sampling techniques. However, these techniques are commonly expensive, episodic, and impractical to apply across a large catchment. Observers with different levels of expertise make it difficult to consistently measure or predict long-term soil loss or erosion risk. A common model and consistent datasets are required for reliable soil loss prediction to provide consistent and continuous erosion information for short- and long-term soil and water quality management [5–7].

Many erosion models have been developed to predict soil loss in different regions in the world. Among them, the Universal Soil Loss Equation [8] or the revised USLE (RUSLE) [9] and Water Erosion Prediction Project (WEPP) [10] are widely used to estimate long-term average soil loss rates using rainfall, soil, topography, and land cover and management as inputs. These estimates have been used in long-term planning and soil condition assessment, and have been applied in Australia continent [11,12]. However, the impact of highly variable and extreme rainfall events cannot be estimated using the long-term averages erosion. Severe erosion and sediment transport are often caused by short but strong storm events [5,6].

It is increasingly important to construct an event-based erosion model to predict the extreme erosion risks given the predicted increase in climate variability and fire intensity [13]. Ideally, areas at risk of severe damage should be identified and prioritized for assessment and remediation in the event of fire [14,15], especially for drinking water catchments. Yet, there are few studies and applications of erosion models on estimation of storm event-based rainfall erosivity and erosion. One of the major limiting factors is the lack of rainfall data at high spatial and temporal resolutions, as well as computational capacity of large quantity of spatial data [16,17]. It remains a research challenge to provide quantitative and timely assessments of hillslope erosion after wildland fires during individual storm events to support water and catchment management [5].

Weather radar data have very high temporal resolution (15 min or less) and spatial resolution (1 km or less) with the potential for estimating event-based rainfall erosivity or the 30-min rainfall erosivity index (EI30). Such data sets have recently been applied in event-based erosion modeling to compute a spatial EI30 index [18] and to monitor erosion after wildfires at Warrumbungle National Park in NSW, Australia [5,6]. Its application over a large area or catchment at near real-time of rainfall events is still a research and implementation challenge. The emerging new technologies, such as machine learning, Google Earth Engine (GEE) processing, and high-resolution (spatial, spectral and temporal) satellite remote sensing, can be employed for the efficient implementation of the event-based erosion model, especially when large catchments or regions are concerned [19]. These together can provide timely erosion risk estimation that explicitly link soil loss and sedimentation with vegetation cover and land management, especially in events of wildfires and storms. The improved capability to predict impacts on water quality as a result of wildfires and erosion helps prioritize the mitigation actions in the drinking water catchments.

The aim of this study was to develop a rapid approach to assess the post-fire erosion in near real-time during storm events. We developed an integrated approach using the RUSLE, remote sensing, and Geographical Information System (GIS) to map the potential erosion risk over space and time, with the SDWC as the pilot study area. Water managers were able to use the results to prioritize monitoring points and areas for erosion assessment and interventions following storm events.

### **2. Materials and Methods**

#### *2.1. The Study Area*

The SDWC area includes five main catchments and 204 sub-catchments (or drainage units). It extends from north of Lithgow in the upper Blue Mountains, to the source of the Shoalhaven River near Cooma in the south, and from Woronora in the east to the source of the Wollondilly River west of Crookwell. These catchments cover an area of almost 16,000 km2, about seven times larger than the Australia Capital Territory (ACT) to its southwest (Figure 1). The dam across the Warragamba River forms Lake Burragorang which provides drinking water to Sydney and surrounding regions for more than five million people or 60 per cent of the NSW population.

The annual average rainfall in SDWC is about 841 mm based on the BoM gridded rainfall data [20] over the 30-year period between 1981 and 2010 (or "climate normals") as used in climate maps and statistics in Australia and many parts of the world [21]. The rainfall in 2019 (before and during the 2019–2020 wildfire) is only about 473 mm which was far below the average. The SDWC area experiences significant seasonal variation in monthly rainfall. The months with the highest rainfalls are February (100 mm), November (85 mm), and January (79 mm). The months with least rainfalls are July (55 mm), September (57 mm), and August (59 mm) over the SDWC area.

The dominant land uses are livestock grazing (35%), nature conservation lands or national parks (30%), crown lands and reserves (16%), and others including intensive agriculture, horticulture, mining, and reservoirs (19%) based on the 2017 land use map over the SDWC area [22]. Soils types in SDWC are variable though strongly influenced by lithology and landform. Texture contrast soils with acidic subsoils (Kurosols) dominate the catchments (46%). Sandy soils with minimal soil development (Rudosols and Tenosol) are also common (30%) and often very shallow in steep-sloped terrain. The elevations range from 21 m to 1463 m (a.s.l) with an average slope of 16.7% over the SDWC area. The elevations range from 41 m to 1356 m (a.s.l) with an average slope of 28.7% at the sub-catchments near the Warragamba Dam and 70% of areas being mountainous terrain.

The 2019–2020 fires have severely and extensively burnt major forests within the drinking water catchments for Sydney, about 30% of SDWC areas have been burnt, and 16% of them are in high or extreme severity based on the fire extent and severity mapping (FESM) [23]. The fires were more severe near the Warragamba Dam, where about 81% of designated Special Areas burned (areas with little to no public access), with 30% at high to extreme fire severities (Figure 1). These drinking water catchments are managed by Water New South Wales (WaterNSW) and the New South Wales Department of Planning, Industry and Environment (DPIE).

**Figure 1.** Location of the Sydney drinking water catchment area and the Warragamba Dam (Lake Burragorang) Special Areas with fire severities, New South Wales (NSW), Australia. The background map is the hill-shaded elevation from the 5 m digital elevation model.

### *2.2. The Datasets*

The primary datasets used in this study were radar rainfall data, satellite images (Landsat-8 and Sentinel-2 with cloud cover < 5%), MODIS-derived fractional vegetation cover (FVC), fire severity, soil and water quality data. The Lansat-8 (OLI Level-2 surface reflectance) and Sentinel-2 (Level-2A BOA reflectance images) datasets for the 2019–2020 wildfire period were obtained via the Earth Engine Data Catalog [24]. The FVC datasets, derived from MODIS Nadir BRDF-Adjusted Reflectance product (MCD43A4) Collection 6, include monthly Photosynthetic Vegetation (PV), Non-Photosynthetic Vegetation (NPV), and Bare Soil (BS) at a spatial resolution of 500 m [25,26] were obtained online [27].

Obtaining clear satellite images at high resolution data, for SDWC area is difficult because of clouds and smokes during the 2019–2020 wildfire period. This problem was partially overcome by using Sentinel-2 with a more frequent (5-day) revisiting frequency. The Sentinel-2 multi-spectral instrument (MSI) sensor provides 13 spectral bands with a spatial resolution of 10–20 m in vegetation mapping. In addition, Sentinel-2 offers three new red edge spectral bands, which has the advantage in improving the accuracy for estimating vegetation indices. The blending of the 5-day Sentinel-2 and the 16-day Landsat-8 data made it possible to create time-series vegetation mapping and RUSLE C-factor estimation in high spatial and temporal resolutions.

Radar rainfall can play a significant role in representing the rainfall intensity, especially in areas without a high density of gauge networks [28,29]. Even where rain gauges or pluviograph rainfall stations exist, they are unlikely to replace radar-derived rainfall estimates, because of the high spatial and temporal resolution from radar data. In Australia, BoM produces real-time quality controlled, rainfall estimates (namely Rainfields) and forecasts using radar, rain gauges, and numerical weather prediction models [30]. It converts real-time radar observations of atmospheric reflectivity into quantitative precipitation estimates (QPEs) via several processing steps including quality controlling, cleaning, analyzing, and integrating data from radars and rain gauges in real time, offering improved spatial and temporal resolution in comparison with rain gauges in the areas covered by weather radars. Rainfields (version 3) includes more radars with QPE products and improved resolution, regional mosaic grids at spatial resolution of 1 km<sup>2</sup> and the national radar mosaic at resolution of 2 km2. In this study, we obtained and used the merged accumulation (Level 2) rainfall for NSW mosaic radar (IDR311MQ, available via FTP to registered users) with the highest quality of QPE at a spatial resolution of 1 km<sup>2</sup> and a temporal resolution of 15 min. These were blended radar and rain gauge rainfall estimates where the calibrated radar estimates were used to add spatial details between the rain gauge locations. Merged rainfall data located at grid points coincident with individual rain gauges are likely to be very similar to gauge observations available at the time of the generation of the product. Rainfields products were stored in NetCDF format and arbitrary coordinate system. We developed automated scripts (R and GIS) to process the Rainfields data including format conversion, re-projection, resampling, and calculation of precipitation. The accumulated precipitation values are calculated by multiplying a scale factor (0.05) and adding an offset (zero in this study) according to the user's guide [30]:

The FESM is a semi-automated fire severity mapping approach in NSW [23] which used a machine learning framework based on the sentinel-2 satellite imagery. The severity map has standardized classes to allow comparison of different fires across the landscape. The FESM severity classes include i) unburnt, ii) low severity (burnt understory, unburnt canopy), iii) moderate severity (partial canopy scorch), iv) high severity (complete canopy scorch, partial canopy consumption), and v) extreme (full canopy consumption). FESM is used in this study for statistical analysis and assessment on impact of fire severity on hillslope erosion.

In addition, recent land use map, soil data, and LiDAR-derived digital elevation models (DEM) [31] were also used to estimate the RUSLE factors (i.e., LS-factor) and statistics. Table 1 summarizes the primary datasets used in this study and their sources.


**Table 1.** Summary of the primary datasets used in this study.


#### **Table 1.** *Cont.*

<sup>1</sup> LiDAR = Light Detection and Ranging; <sup>2</sup> DPIE = New South Wales Department of Planning, Industry and Environment; <sup>3</sup> CSIRO = Commonwealth Scientific and Industrial Research Organisation Australia; <sup>4</sup> NASA = The National Aeronautics and Space Administration of the USA; <sup>5</sup> BoM = The Bureau of Meteorology, Australia; <sup>6</sup> ESA = European Space Agency.
