Urban land-use land-cover extraction for catchment modelling using deep learning techniques

Floods are one of the most destructive and frequent natural disasters, causing many personnel casualties and property losses every year (Dawson et al. 2008; Jha et al. 2012; Zaman et al. 2012; Willner et al. 2018). The process of urbanisation has increased the flood risk to human society with the rapid growth of impervious areas such as buildings, roads and parking areas, which break the infiltration of stormwater to the deep soil and permanently change the natural rainfall–runoff process of the catchment (Arnold & Gibbons 1996; Jha et al. 2012; Ball et al. 2019). Unlike the nature catchment, the impervious surface degrades nature flood resilience by directly transferring rainfall to runoff, which increases the flood volume and leads to the earlier arrival of the flood peak (Holman-Dodds et al. 2003; Li et al. 2017). The weakening of catchment nature flood resilience results in the vulnerability of communities to flood when exposed to the storm burst beyond the capacity of the urban stormwater drainage system. In addition, other factors such as climate change, irresponsible and inappropriate city planning and stormwater management strategies are growing trends that aggravate the devastation of floods in urban areas.

To understand the rainfall and runoff processes that interconnect the catchment hydrological components of a catchment and to further study the occurrence and consequence of flood events, catchment models were introduced to predict the catchment response to rainfall; limitations of measurement techniques and extremely complicated inner mechanism of catchment hydrology also encouraged the development of catchment models (Beven 2012; Peel & McMahon 2020). In the computer era, distributed catchment modelling systems have experienced significant progress with the development of remote sensing technology, geographic information system (GIS) and increased computation power (Salvadore et al. 2015); all of these developments allow the modeller to reproduce the catchment rainfall–runoff process at a scale suitable for the available data. These sophisticated catchment models require a mass of catchment data and the associated data digitalisation, assimilation, interpretation, parameterisation and calibration to maintain the robustness of the model predictions. Meanwhile, data quality and quantity are more important than ever before; the adequacy and reliability of a dataset play a critical role in the model performance (Gan et al. 1997). However, identification and digitisation of the spatial characteristics of urban catchments such as land use land cover (LULC) have become a significant problem that plagues many aspects (e.g. catchment modelling and urban planning) with the expansion of urban areas and the changing of land use (Liu et al. 2015; Hu et al. 2016). The importance of land-use information for catchment modelling systems is reflected in determining the impervious ratio, infiltration rate and the associated surface roughness coefficient (e.g. Manning's n value) (Ball et al. 2019). Therefore, quickly providing accurate and reliable datasets to parameterise the catchment spatial characteristics for modelling purposes has become a topic that needs to be solved in the modelling community (Abulohom et al. 2001; Ball et al. 2019; Peel & McMahon 2020).

Breakthroughs in remote sensing techniques allow modellers to obtain numerous catchment spatial data with a visualisation form (Wulder et al. 2012). A typical representative of visualised catchment spatial data is the remote sensing imagery, which has become the indispensable dataset of catchment modelling systems (Klein & Barnett 2003; Mackay et al. 2003; Beven 2012). The interpretation of spatial information from the remote sensing imagery is required to estimate the parameters for catchment modelling purposes (Foody 2004). This interpretation includes identification, delineation and digitalisation of LULC (e.g. road, building, impervious/pervious area, vegetation and waterbody) (Boegh et al. 2004), which are time-consuming and require solid geographic knowledge (Niemi et al. 2019). In addition, the accuracy and reliability of the manual interpretation solely rely on the interpreter's ability and cannot be reproduced by other modelling groups or systems. This reduces the model's consistency when the catchment is experiencing LULC changes.

In recent years, the development of deep learning techniques has greatly facilitated the development of computer vision (CV) applications; the CV is a subbranch of machine learning with interdisciplinary knowledge involving how computers acquire information, process data and gain advanced understanding from digital images or videos (Voulodimos et al. 2018; Szeliski 2010). Examples of CV applications can be found in autonomous driving (Grigorescu et al. 2020), face recognition (Parkhi et al. 2015), image classification and segmentation (Ronneberger et al. 2015; He et al. 2017).

Meanwhile, deep learning has also achieved excellent performances in remote sensing applications. For example, Rezaee et al. (2018) proposed a deep learning neural network (DNN) to classify wetlands from remote sensing imagery. Zhao et al. (2017) introduced an object-based convolutional neural network to discriminate building types and footprints in a catchment of Beijing and achieved over 90% accuracy. Huang et al. (2018) modified the classic DNN architecture to adapt the multi-spectral data of high-resolution remote sensing imagery and achieved 91.25% precision in the land-use classification for the study area of Shenzhen and Hong Kong. Yang & Li (2015) used the linear spectral unmixing method to conquer the effect of vegetation for extracting road from an urban remote sensing image. In summary, these works focus on extracting individual spatial features or classifying images with low LULC heterogeneity (e.g. rural catchments and urban catchments with low landscape coverage). There is no universal methodology for modelling urban catchments, particularly in high heterogeneous LULC areas such as various landscape coverage and diverse building types (Salvadore et al. 2015); whether deep learning can achieve classification and segmentation of LULC for parameterising and modelling urban catchment, extensive research has not yet been undertaken.

Inspired by related works, this research develops a novel approach utilising a deep learning-based CV tool. This approach aims to produce the LULC classification and segmentation of urban catchment remote sensing imagery with heterogeneous spatial features and present a reproducible workflow for any other Australian urban catchment under the same remote sensing database. The feasibility and practicability of this approach for catchment modelling systems are discussed also. The remainder of this paper is organised as follows. Following the introduction, the study catchment is described, the proposed approach. The section ‘Result and discussion’ demonstrates the outputs and discusses their suitability for catchment modelling. Finally, the section ‘Conclusion’ presents the conclusions derived from the research.

The Alexandra Canal catchment was selected as the test area for this study. This catchment is located south of the great Sydney CBD area, Australia. The catchment area is 11.50 square km and consists of four highly urbanised regions: Sheas Creek, Rosebery, Munni Street-Erskineville and Alexandra Canal, with multiple land uses inclusive of residential (approximately 40%), industrial (approximately 25%), road (approximately 10%), parkland (approximately 22%) and water body (approximately 3%). The catchment land covers are varied and include single dwelling, terrace, dense apartment, industrial plants with large impervious areas and sizeable pervious areas of parkland and golf courses. The drainage system of the Alexandra Canal catchment consists of subsurface pipes, pits, subsurface channels, open channels, culverts and flood mitigation structures. The location of the catchment with respect to continental Australia and within the greater Sydney region is shown in Figure 1 (Esri 2021).

A significant feature of the catchment is the high heterogeneous LULC, such as the various roof colours and shapes, perennial vegetations, water bodies, railways, walking trails, roads, large pavement areas and buildings’ shadows. A few ground features are difficult to identify on the imagery due to the obscuration of tree canopies. Additionally, in some regions of the catchment, the land zoning has changed from industrial land to commercial or residential land, leading to discontinuities in land cover between the expected and actual. The identification of these spatial features is essential for the catchment model. For example, a critical inferred control parameter ‘Impervious percentage’ in the stormwater management model (SWMM) is highly related to the measured geometry of catchment LULC and associated roughness coefficient, which requires the modeller to identify and delineate all impermeable ground features (e.g. roof and road) (Choi & Ball 2002). In summary, the LULC identification and classification of the catchment is a complex and very challenging task from the perspective of both CV and catchment modelling.

The proposed LULC classification and segmentation methods of the catchment imagery consist of three algorithms, which are the clustering algorithm MeanShift (Cheng 1995), and the pixel classifiers U-net (Ronneberger et al. 2015) and DeepLabV3+ (Chen et al. 2018).

In the beginning, the catchment raw imagery was pre-processed to adapt to the deep learning environment. The MeanShift clustering algorithm was introduced to simplify the spectral information of the catchment imagery by producing a preliminary segmentation result. The segmented image was then used as the input of the pixel classifier U-net. Pure U-net and DeepLabV3+ classifiers were also implemented to classify and segment the LULC information using the catchment raw image as input for the control experiment. Then, the F1 score was adopted to evaluate the models’ prediction accuracy on the validation set, and the associated images to demonstrate the bias between the ground truth and classification outputs. Finally, the performance of trained models on the test set was quantified using the confusion matrix to assess LULC accuracy associated with different catchment modelling systems. In machine learning, both the F1 score and confusion matrix are effective measures to analyse the accuracy of the model's prediction, where the F1 score is the harmonic mean of the precision and recall, the confusion matrix is an n × n matrix by comparing the actual class of raw image and predicted class of classified image (Powers 2011; Chicco & Jurman 2020). Detailed procedures are demonstrated in Figure 2.

Image pre-processing is the critical step in remote sensing applications. Many commercial software packages like ArcGIS and MapInfo have mature techniques and associated functions to process images for different industries and issues. The remote sensing data processing includes geometric referencing, image fusion, image mosaic, image cropping, cloud removal, shadow processing and atmospheric correction (Camps-Valls et al. 2011; Yang & Li 2015).

The focus of remote sensing image pre-processing is diverse to the different application requirements of various industries. For example, atmospheric correction is an essential process for remote sensing-based agricultural management that reduces the impact of atmospheric scattering and clouds obstruction on images (Wu et al. 2005). Regarding catchment modelling, the focus is on LULC recognition and geometric accuracy, which requires high image resolution and geometric correction. Therefore, it is necessary to process the image fusion by using the panchromatic image as the reference image to register the multi-spectral image to obtain the RGB remote sensing images with geographic coordinates and high resolution (Thomas et al. 2008; Bai et al. 2015; Ghassemian 2016).

High-resolution remote sensing imagery for the catchment in 2019 was collected from MapServer (Spatial Services of New South Wales Government), with three bands (Red, Green and Blue) and a resolution of 25,344 × 20,992 (0.29 m/pixel width) (NSW Spatial Service 2021). The coordinate system of source imagery is WGS 84/UTM zone 56S. A digital surface model (DSM) was utilised as the base layer for the treatment of multi-spectral image orthorectification and thereafter to geometrically correct the image to ensure it was planimetrically projected. The GIS platform ArcGIS Pro Version 2.8.2 was applied to delineate the catchment vector boundary. Following these steps, the catchment boundary was used to clip the remote sensing imagery and unify the geocoordinate system.

In machine learning, a common approach is to use a large number of task-related data sets to:

• train the model,

• adjust the model's parameters through the continuous iteration of training error on the data sets and

• get the optimised model that best fits the model data set.

This is also referred to as supervised learning (Hastie et al. 2009). The construction of the training set requires a large number of labelled data (pictures, text files, videos, etc.) to provide a mapping channel for the model training. In the CV field, data labelling is used to identify the original images or videos and add one or more meaningful labels, which objectively establishes a mapping channel from data to results for the machine learning model to learn information from it (Cowie et al. 2011). For example, in the task of LULC recognition of remote sensing images, the label can be the binary representation of an object within an image, or it can be the semantic delineation of ground surface features at the pixel level. The starting point for data labelling requires humans to make judgments on designated unlabelled data. The trained model can be applied to make predictions on new data sets by learning hidden patterns within the human ascribed labels (Li et al. 2021). Labels are treated as objective criteria to train and evaluate the supervised machine learning model, often referred to as a ‘standard answer’. The model training and accuracy of the prediction depends on the correctness of the ‘standard answer’, so it is essential to allocate sufficient time and resources for labelling to ensure the high accuracy of labels.

The ratio of the training and validation was set to 9:1. The trained model will be applied to make predictions on the entire catchment imagery (Figure 3) for testing the model's performance on the rest area of the imagery without labels. In catchment modelling systems, the determination of representative class schema depends on the comprehensive land-cover conditions and the catchment spatial features that affect the catchment rainfall–runoff process. Therefore, I defined the catchment LULC class schema for this study as roof, road, railway, water, impervious, pervious and tree by referring to the zoning map from the City of Sydney Council and observable features of the catchment remote sensing imagery. The manual labelling approach was used to maintain the accuracy and consistency of label masks. It is worth mentioning that some ground surface features are occluded by the tree canopies, which leads to identification difficulties and incorrect labelling. To address this issue, I ignored the occluded ground features but delineated and labelled the tree canopies above them to ensure the stability of the mapping channel. The tree can be seen as an uncertainty class, where its prediction should be analysed with other conditions to obtain the correct parameter for catchment modelling purposes. Approximately 25% of the image area was labelled. An image clipping tool was applied to grid the labelled area into 5,636 small images with a size of 256 × 256 and then rasterise them to generate the label masks by snapping to the pixel alignment. Figure 3 shows the labelling area and class schema of the study.

The MeanShift image segmentation algorithm extracts general information from the complex features through coarsening and then segmenting the image (Comaniciu & Meer 2002). The concept of MeanShift was first proposed by Fukunaga & Hostetler (1975). In unsupervised learning, MeanShift refers to a density-based non-parametric clustering algorithm that assumes the data of different clusters conform to the corresponding probability density distributions and finds the density vector of any sample points in which the highest vector means the direction of iterations. These sample points will eventually converge to the area with the maximum local density, and the points that converge to the same local maximum are considered members of the same cluster (Comaniciu & Meer 2002; Szeliski 2010). In image segmentation, MeanShift can be used to segment image objects by clustering the neighbouring pixels with similar spectral and texture information.

Pixels of remote sensing imagery usually contain two types of information: coordinate information and colour information, which constitute the spatial features of the image (Foley et al. 1994). In the urbanised Alexandria Canal catchment, the regular geometric of artificial structures and the similar colours of vegetation lay the foundation for the use of MeanShift for preliminary image segmentation. In addition, the spatial features of the imagery are very complex, and the associated data amount is enormous due to the high heterogeneity of the catchment LULC. Therefore, it is necessary to take preliminary segmentation measures to simplify the spatial features of the image. In this study, the MeanShift segmentation result was used as the input of the DNN classifier to train the model and test set to make predictions of LULC. Figure 4 shows the examples of segmentation by MeanShift.

Two deep neural network-based algorithms, such as U-net and DeepLabV3+, were selected to conduct the study catchment's LULC classification and semantic segmentation. In the distributed catchment modelling systems, any catchment's internal spatial characteristics and distribution that affect the rainfall–runoff process should be considered and parameterised, where the spatial patterns are expressed by pixel in remote sensing imagery. Therefore, the pixel-based semantic segmentation methods were chosen instead of the object focused instance segmentation method.

U-Net is one of the earliest algorithms for semantic segmentation using fully convolutional networks (Ronneberger et al. 2015). The proposed symmetrical U-shaped structure with innovative encoding and decoding paths has affected the development of many segmentation networks such as U-net++ (Zhou et al. 2018) and SegNet (Badrinarayanan et al. 2017).

In this study, the pre-trained model Resnet-34 was applied as the backbone for enhancing the downsampling process. A schematic diagram of U-net with the pre-trained backbone neural network ResNet-34 (He et al. 2016) is demonstrated in Figure 5.

The left side of the network is a series of downsampling operations composed of four convolution and max-pooling layers, and the right side of the network is called the expansive path with four upsampling layers to recover the feature map to its original resolution. The upsampling outputs are connected to the corresponding cropped feature maps from the downsampling path, where the features learned during downsampling are applied to the upsampling to enhance the accuracy of output.

DeeplabV3+ is the fourth generation of the DeepLab series of semantic segmentation models developed by Google, with encoder–decoder architecture, which is different from the progressive upsampling of the classic symmetric neural networks (e.g. U-net). DeepLab uses atrous convolution to reduce the downsampling rate without losing the receptive field, that the final feature map contains rich and elaborate semantics information. The original resolution can be restored directly through interpolation instead of the downsampling feature maps’ horizontal connection (Chen et al. 2018).

The architecture of DeepLabV3+ is shown in Figure 6. The main body encoder part is a deep convolutional neural network (DCNN) with atrous convolution, the commonly used DCNN such as ResNet can be used, followed by the Atrous Spatial Pyramid Pooling module (ASPP) for introducing multi-scale information. DeepLabv3+ introduced the decoder part to improve the segmentation accuracy of class boundary by integrating the features of different levels.

In this study, the stratified random method was applied to create 500 assessment points that are randomly distributed in each class. The number of points in each class is proportional to its class area. Then, the κ coefficient was calculated by confusion matrix to evaluate the prediction consistency.

In this study, three methods were used to predict LULC from remote sensing images of the study catchment, namely U-net, DeepLabV3+ and MeanShift+U-net, with the same test set, training set amounts and masks. The pre-trained backbone neural network and iteration parameter was set to Resnet-34 and 50 for all three methods. The difference is that the MeanShift+U-net method uses the preliminary segmented imagery as the base map instead of the raw images adopted by the other two methods. Figure 7 shows the training and validation loss curves. The training process of the three models was smooth, and the fast convergence of the loss was achieved, which indicates that the gradient is continuously decreasing and the selected learning rate is appropriate. U-net has the best training performance, followed by MeanShift+U-net, and the last one is DeepLabV3+. A vibration occurred on the verification curve of DeepLabV3+, and the fitting degree between the verification and training is not as good as the former two. Figure 8 shows the performance of the three methods on the verification set.

The advantage of the F1 score is avoiding the unstable recall and precision due to the excessive number of positive samples. Figure 9 shows the F1 score of three models in each class. The three methods have minor differences in the F1 scores of each LULC class. U-net performs better in major classes (e.g. roof and impervious), and DeeplabV3+ performs better in minor classes (e.g. tree and road), which may be powered by its excellent boundary segmentation capability. The overall performance of MeanShift+U-net is lower than the other two methods, but the prediction is more balanced, which means better consistency between input and output.

In the performance evaluation strategy, the confusion matrix and the κ coefficient were applied to assess the performance of the three trained models in predicting the LULC of the study catchment.

Five hundred random points were created on the catchment imagery as assessment points, where the number of points in each class is proportional to its corresponding area percentage.

The κ coefficient for all three trained models exceeded 0.75; the κ coefficient values are 0.7929, 0.7593 and 0.7955 for U-net, DeepLabV3+ and MeanShift+U-net, respectively. These κ coefficients indicate that the classification results of the three trained models are strongly consistent, and there is no bias towards the major classes with large areas. The U_Accuracy and P_Accuracy represent the User Accuracy and Prediction Accuracy for reflecting the model's accuracy from different observation channels. For example, in the tree class of Table 1, 16 points lost the correctness among the total 72 assessing points, and 56 points were counted to the correct class within 65 classified samples so that the P_Accuracy and U_Accuracy are 0.7778 and 0.8615. Tables 1–3 show the confusion matrixes and κ coefficients of the three trained models.

This section will discuss the suitability of the test results of the above three methods in two types of catchment modelling systems (fully-distributed and semi-distributed modelling systems). Both fully-distributed and semi-distributed models involve considering catchment spatial pattern distribution when determining the runoff generation and routing, where the remote sensing datasets such as imagery and DEM are widely applied in obtaining spatial features for modelling purposes (Bell et al. 2009; Ball et al. 2019).

The remote sensing-based catchment modelling systems have rigorous requirements for the geometric accuracy of spatial features in estimating the model's parameters. Therefore, the selected index of suitability assessment for catchment modelling should be the P_Accuracy (Prediction Accuracy) values of the CV model on the test set, which can reflect the classification and segmentation correctness of the CV model on each LULC class. Tables 1–3 illustrate that the three methods’ overall prediction accuracy values have all exceeded 0.8, yielding 0.8360, 0.8103 and 0.8379 for U-net, DeepLabV3+ and MeanShift+U-net, respectively. In addition, the κ coefficient of MeanShift+U-net is the highest among the three methods, which means MeanShift+U-net has the best consistency in predicting the LULC of the catchment. The advantage of consistency will be gradually expanded with the increase of the prediction area, which is beneficial to maintaining the stability and robustness of the parameter estimation model under the changing catchment conditions. Therefore, the MeanShift+U-net method has the best performance in extracting LULC information at the pixel scale among the three with the overall P_accuracy of 0.8379. Figure 10 shows the MeanShift+U-net LULC prediction of the entire catchment.

MeanShift+U-net prediction is also tested at the subcatchment scale under the schema of semi-distributed modelling systems, which have a pre-determined flow direction and the requirement of fewer parameters. The spatial distribution is ignored within the subcatchment scale of the semi-distributed model (Sitterson et al. 2018). For example, Choi & Ball (2002) categorised the control parameters of the SWMM into measured and inferred parameters, where the inferred parameters such as impervious ratio, subcatchment length/width, depression storage, infiltration and Manning's roughness coefficients are highly related to the classification result and segmentation geometry of catchment LULC expected to be obtained from the MeanShift+U-net prediction. Therefore, the original four sub-classes (railway, road, impervious and roof) treated as the impervious feature were merged into one class (Impervious) in the confusion matrix. The new class schema is ‘Tree’, ‘Pervious’, ‘Waterbody’ and ‘Impervious’. The purpose of this simplified class schema is to assess the suitability of the semi-distributed modelling systems by ignoring the classification errors among the sub-classes with impervious rainfall responses. Tables 4 and 5 show the confusion matrixes of the DeepLabV3+, U-net and MeanShift+U-net after merging classes. One table (Table 5) is used to show the simplified confusion matrix for both U-net and MeanShift+U-net due to their same accuracy distribution among the 500 assessing points. The P_Accuracy value of ‘Impervious’ was significantly increased in the three methods under the simplified class schema, in which the ‘Impervious’ P_Accuracy occurred from U-net, MeanShift+U-net and DeepLabV3+ are 0.9847, 0.9847 and 0.9634, respectively. The κ coefficient of DeepLabV3+ has been improved but is still lower than the other two methods.

Considering the highest κ coefficient achieved by MeanShift+U-net in the original class schema, MeanShift+U-net is recommended for estimating the initial parameters of semi-distributed modelling systems. Reasons are as follows: First of all, the semi-distributed model does not require high geometric accuracy of the spatial features within the subcatchment scale, which reduce the probability of models’ failure by downgrading the resolution of prediction error. Secondly, the estimated parameters for the semi-distributed model are more accurate as the impervious area prediction accuracy (0.9847) of MeanShift+U-net is very close to the accuracy of manual delineation. In addition, the parameter uncertainty caused by the CV model error is objectively reduced since the number of parameters of the semi-distributed model is much less than that of the fully distributed model. Finally, the higher κ coefficient of the MeanShift+U-net method illustrates that its better classification consistency could maintain a stable prediction accuracy on the sizeable urban catchment with high heterogeneity of LULC. However, the number of subcatchments in a metropolis could be very large due to the complex LULC and drainage system-controlled land divisions. In the specific implementation of MeanShift+U-net on catchment parameterisation, it is necessary to simplify the large number of subcatchments to a certain extent to reduce the impact of the error caused by the high resolution of spatial features.

Nevertheless, this method still has some significant drawbacks as follows:

• There is a significant gap in achieving accurate geometric parameters of the catchment spatial features at the pixel scale by referring to the overall P_Accuracy. This error will be gradually enlarged, with the expansion of the study catchment area, and causes the collapse of the entire model eventually.

• The occlusion of tree canopy and buildings’ shadows could lead to incorrect determinations of the underlying ground features. Further study about mitigation approaches is necessary to reduce their negative effect.

• The spectral similarity between the bare soil and the impervious area, resulting in a large portion of bare soil with pervious rainfall response was classified as an impervious feature, which will further lead to the uncertainty of the parameters and the instability of the catchment model.

• Errors and misprediction could occur when using these approaches on other remote sensing databases such as Google and Quickbird due to the pixel spectral variation caused by different image processing methods.

It is recommended to train the deep learning models on the associated remote sensing dataset before applying the approach to a practical catchment model. The LULC class schema used in this study is not fixed. Modellers should set up suitable LULC classes to represent catchment spatial features and correlate to the catchment rainfall–runoff process.

This paper proposes three methods for LULC classification on Alexandria Canal catchment remote sensing imagery. The study imagery was pre-processed, and the MeanShift algorithm was applied to achieve preliminary segmentation of the catchment LULC objects with similar spatial features. The raw catchment image samples were trained using U-net and DeepLabV3+, and the MeanShift+U-net was trained by the pre-segmented catchment image of MeanShift. Then, the U-net, DeepLabV3+ and MeanShift+U-net worked as the classifier for the study catchment imagery. Finally, the confusion matrix was applied to assessing the suitability of the prediction for catchment modelling systems.

In contrast with the pure U-net and DeepLabV3+, MeanShift+U-net has achieved the best prediction accuracy and excellent consistency in the study catchment. Both original and simplified confusion matrices indicate that the proposed methods are more applicable to estimating spatial parameters for the semi-distributed catchment modelling systems, where the MeanShift+U-net is recommended due to its high accuracy (0.9847) in classifying impervious features. (In addition, it is necessary to explore the performance of MeanShift+U-net on multi-sources remote sensing dataset and evaluate its consistency on different LULC classes further, as well as the influence of image noise such as tree canopy and buildings’ shadow on the classification accuracy, so as to improve the method's practicality in the catchment modelling field.)

The land zoning map, drainage network and cadastral map of the Alexandria Canal catchment are provided by the City of Sydney Council, Sydney, Australia. Sincere thanks are given for the support of the City of Sydney Council and their flood risk management team.

Data cannot be made publicly available; readers should contact the corresponding author for details.