Merging weather radar and rain gauges for dryland agriculture

2.1. Case study area

The Wimmera region lies within western Victoria, Australia (Fig. 1), covering ~30,000 km². The region is one of Australia’s most productive grain growing regions due to the soils and climate (Agriculture Victoria 2023; Wimmera Catchment Management Authority 2023). Parts of the region comprise Grey Vertosol soils, which are the principal cropping soils of western Victoria. The Red and Brown Vertosol soils are also used for cropping but are not as high yielding as the Grey Vertosol soils. Most of the landscape is generally either flat or undulating. The Wimmera has a semi-arid to sub-humid Mediterranean type climate with rainfall varying from 475 to 615 mm year^–1 (Wimmera Catchment Management Authority 2024).

Fig. 1.

Location of the Australian Bureau of Meteorology Rainbow weather radar and the contributing rain gauge network (Table 1) within the coverage of the Rainbow weather radar in the state of Victoria, Australia.

2.2. Rainfall data

Within the case study region, rainfall data are generated by three basic methodologies: rain gauges, radar and satellite.

2.2.2. Gauge-based gridded rainfall data

Gridded analysis of rainfall data is used by major meteorological organisations across the world with continental-scale daily climate surfaces derived by interpolating the observed point data. Australian modellers have historically used two publicly available archived products that both cover Australia on an ~5- × 5-km grid: Queensland Department of Science, Information Technology and Innovation SILO (Scientific Information for Land Owners) and the Australian Bureau of Meteorology Australian Gridded Climate Data (AGCD).

The SILO dataset is interpolated rainfall observed at Bureau gauging stations across Australia, based on smoothing splining and Kriging techniques described in Jeffrey et al. (2001). AGCD is a product developed by The Bureau and CSIRO as part of the Australian Water Availability Project. Similar to SILO, the AGCD dataset is interpolated rainfall observed at Bureau gauging stations across Australia, with the interpolation method using a two-step process (Beesley et al. 2009), interpolating anomalies of daily rainfall using Barnes’ successive correction method (Jones et al. 2009) that applies a weighted averaging process to the station data. On the maps, each grid point represents an approximately square area with sides of ~5 km (0.05°), as illustrated in Fig. 2. This grid point analysis technique provides an objective average rainfall for each grid square. However, in data-rich areas such as south-east Australia or in regions with strong gradients, this applied analysis results in grid point values that may differ slightly from the exact rainfall amount measured at the contributing stations, due to data smoothing effects.

Fig. 2.

Longerenong DATA Farm rain gauge network (Map data Google Satellite, WMS, see https://mt0.google.com/vt/lyrs=s&x=x}&y=y}&z=z}, accessed 17 February 2024).

Fu et al. (2022) compared the SILO and AGCD, showing that both grids could produce significant differences in rainfall at the local scale. This is because the products use different interpolation algorithms and different gauging stations for compilation and interpolation. AGCD is limited to data obtained from ‘qualified’ stations in the Australian Data Archive for Meteorology (ADAM), whereas SILO also uses data from organisations such as local governments, mining companies, national parks and private properties.

2.3. Radar data

The radar data used in this study are sourced from the long-term archive of The Bureau operational radar network. The Rainfields 3 dataset (Seed et al. 2022) is used operationally in Australia for Quantitative Precipitation Estimate (QPE) applications where accurate estimates of rain rate are required from radar. The Rainfields 3 legacy datasets are accessible 48 h behind real time and are covered by open Creative Commons licence CC4-BY-NC.

Merged radar rainfall data are calculated using the radar reflectivity data from rain drops, which is an indirect instantaneous measurement of rainfall at a known height. The main steps to estimate rainfall from radar in real time follow: (1) measurement of radar reflectivity and removal of known sources of errors such as beam blockage, reflections and ground scattering of the radar beam; (2) Z–R conversion: the algorithm used to convert reflectivity (Z) to rainfall (R) rate, which can vary from location to location and between different precipitation types; and (3) merging with rain gauge network rainfall data to remove any field bias. Radar rainfall estimation is made for a 500- × 500-m pixel that is 1000 m above the radar, whereas the rain gauge observes the rainfall accumulation at a point on the ground. Surface rainfall maps are constructed by using a Kriging interpolation technique (Seed and Pegram 2001; Wesson and Pegram 2006) to interpolate the radar observations. The radar data have been blended with selected rainfall observations from The Bureau rain gauge network to account for any systematic biases, adding spatial detail between the rain gauges based on Kriging interpolation using a set of algorithms to estimate the rain rate at ground level from the radar reflectivity observations in the atmosphere.

The QPE merged radar rainfall is downloaded in the form of 15-min rain accumulation grids, from network Common Data Form (netCDF) libraries, with a horizontal grid spacing of 500 × 500 m. The metadata of these data files contain a list of The Bureau station identifiers of the contributing gauges (Table 1), used in the Kriging algorithms for that 15-min data collection. The mapping projection is Albers Conical Equal Area with all times and dates in Universal Coordinated Time (UTC). The 15-min data are appropriately summed to provide a reading of the total amount of rainfall measured in the 24-h period preceding 09:00 hours on the day of reading, converting to local time, and subsequently reprojecting the gridded data to EPSG:4283 – GDA94, a common geodetic coordinate system used in Australia. The data are then in a format that facilitates further analysis by comparison with other daily rainfall data collected other devices such as GPS located rain gauges.

Most of the Wimmera region is covered by The Bureau’s weather radar located on private farmland near Rainbow, Victoria (latitude 35.9975°S, longitude 142.0133°E, altitude 148.225 m, height 27 m). The Rainbow weather radar is a Meteor 735CDP (high resolution C-Band, Doppler and Dual Polarisation enabled), and generates data (typically) 24 h day^–1 centred in a 256- × 256-km grid with 500- × 500-m spatial resolution (Fig. 1).

Beam blockage is mostly due to topographic features, and the Rainbow radar is situated in a landscape of flat to gently undulating plains. The radar beam has good coverage extending up to 200 km with minimal impact of any obstructions and has a full 360° view (Fig. 3).

Fig. 3.

Beam blockage map for the Rainbow weather radar installation (8 December 2021, Australian Unified Radar Archive, see https://open-radar-beamblock-maps.s3-ap-southeast-2.amazonaws.com/95_0_CBB_map.png). ASL, above sea level.

2.4. Statistical methods

Regression analysis, a statistical method used to explain the strength of the relationship between a dependent variable and one or more independent variables, was used in this case study.

The independent variables are the rain gauge network daily observations (G_obs) and the dependent variables are the Rainfields 3 QPE merged radar rainfall daily observations (R_obs) as they depend on the algorithmic processing of the selected contributing Bureau rain gauges and radar data.

To statistically analyse the comparison between the rain gauge network daily observations and merged radar rainfall daily observations, five standard metrics were used considering time series of G_obs and R_obs, each with the same number n of time steps. The linear regression line describes the relationship between G_obs and R_obs. The resulting linear equation is used to calculate predicted daily rainfall values (R_pre) for corresponding daily G_obs values.

The Pearson correlation coefficient (Corr ∈ [−1, 1]) estimates the strength and direction of the linear relationship between time series:

The Pearson correlation measures the strength of the linear relationship between two variables (a measure of linear correlation between two sets of daily rainfall data) and has values between −1 to 1: a value of −1 meaning a total negative linear correlation, 0 being no correlation and +1 meaning a total positive correlation. If the coefficient value lies between ±0.50 and ±1, then it is said to be a strong correlation. If the value lies between ±0.30 and ±0.49, then it is said to be a medium correlation. When the value lies below ±0.29, then it is said to be a weak correlation.

The coefficient of determination (R²) measures how well the least squares regression line fits (with range 0–1), 1 is best fit for a radar value predicted perfectly by the rain gauge observation.

Sum of squares of residuals, RSS:

Total sum of squares (TSS) is the sum of all squared differences between the mean of an observation and the individual values in that observation.

The R² is the square of the correlation coefficient.

The root mean squared error (RMSE) measures the average difference between predicted and observed values:

The lower the RMSE value, the better the linear regression is. A zero value indicates that the predicted and observed values match precisely. Higher values indicate a poorer fit to the least squares regression line.

The Nash–Sutcliffe Efficiency (NSE ∈ (−∞, 1]) measures how well the outputs of a model reproduce observations against a model that uses only the average of the observed data:

The NSE ranges from −∞ to 1 (1 = perfect score, optimal value), with NSE = 0 representing a radar daily rainfall accuracy corresponding to the mean of the gauge daily rainfall. Values of 0.6–1.0 are generally viewed as acceptable levels of performance, whereas values <0.0 indicate that the mean observed value is a better predictor than the simulated value, which indicates unacceptable performance.

Lin’s concordance correlation coefficient (CCC) is the concordance between a new test or measurement (Y) and a gold standard test or measurement (X). This statistic quantifies the agreement between these two measures of the same variable, namely daily rainfall, with the rain gauge measurement (G_obs) represented by X, and the radar measurement estimation of rainfall (R_obs) represented by Y. Values near +1 indicate strong concordance between G_obs and R_obs, values near −1 indicate strong discordance and values near zero indicate no concordance (Lawrence and Lin 1989).

3.1. Daily merged radar rainfall maps

The daily merged radar rainfall maps and the locations of the chosen contributing Bureau rain gauges, detailed in the metadata of the 15-min Rainfields 3 QPE data files as listed in Table 1, are presented in Fig. 5.

Fig. 5.

Accumulated daily rainfall from the Rainbow QPE merged radar rainfall dataset for the five selected rainfall events.

The halo effects displayed in the maps around some of these rain gauge sites chosen as Kriging points, are likely due to erroneous artifacts introduced by the Kriging interpolation algorithms implemented in the Rainfields 3 system, also referred to as the ‘bullseyes’ artefact (Chua et al. 2022), illustrated in Fig. 5 and 6.

Fig. 6.

(a) Rainbow QPE merged radar rainfall 15-min data file, ‘95_20230327_034500.prcp-m15.nc’, one of the 96 15-min data files used to determine the quantities and the spatial distribution of the daily rainfall and (b) the accumulated rainfall for the 24 h preceding 09:00 hours on 27 March 2023. ‘Bullseye’ anomalies (yellow circles), located around the contributing Bureau gauges detailed in the metadata (Table 1), dot symbols on the map. Red ellipses in map (b) indicate the wind farm anomalies as displayed in Fig. 7.

3.2. Issues with radar data

One shortcoming evident in this case study of the Rainbow weather radar installation is the anomalies introduced by the presence of wind farms (Fig. 7). On many occasions, the wind turbine artefacts resulting from the interference of the reflected radar signal are not identified by the Rainfields 3 system and subsequently are not removed from the displayed output.

Fig. 7.

The red ellipses highlight wind farm anomalies interfering with Rainbow radar reflectivity on a rain-free day 6 February 2023. The wind farm locations are: (1) Diapur, two turbines; (2) Kiata, nine turbines; and (3) Murra Wurra, 99 turbines. This is a derivative of http://www.bom.gov.au/products/IDR953.loop.shtml, available under the Creative Commons Attribution Licence 3.0.

Norin (2015, 2017) discusses the effects of wind farms on the determination of rainfall by weather radar, suggesting that their impact is identifiable as a distinct and highly repeatable signature which can be used to identify and filter out the anomaly caused by the wind turbines. There is a need to be able to classify the anomalous propagation that is caused by wind farms and correctly map the true rainfall pixels quickly and easily.

3.3. Comparison of daily rainfall – Bureau rain gauge network with radar merged rainfall maps

Rainfall data pairs from each of the 26 contributing Bureau rain gauges and the QPE weather radar merged rainfall for the five rainfall events were analysed, with seven outliers removed, leaving a total of 123 data pairs. The results of this analysis are described both in Fig. 8 and Table 2. As expected, there is an overall relationship between G_obs and R_obs. As a result of the algorithms employed in the Rainfields 3 processing, each of The Bureau official contributing gauges (as specified in the metadata of the dataset files) is the selected Kriging point for the Kriging interpolation. This is particularly evident in the results for 26 February and 7 April 2023 with Lin’s CCC values of 0.9309 and 0.9454 respectively, indicating substantial concordance between G_obs and R_obs. Factors other than the Kriging interpolation that are part of the Rainfields 3 system, such seasonal factors, and non-compensated radar reflectivity interference, are possible contributors to the variability of the Lin’s CCC, range 0.7960–0.9454, for the five rainfall events. The NSE values are a measure of how well R_obs represents a radar rain rate accuracy corresponding to the mean of G_obs, using only the average of the observed data. For the five events, values within 0.7092–0.9121, indicate ‘passable’ to ‘very good’ level of performance. The Pearson coefficient, range 0.8472–0.9551, also indicates a strong correlation.

Fig. 8.

Comparison of daily rainfall measured by The Bureau rain gauge network with determinations of rainfall from the QPE weather radar merged rainfall maps for the five chosen rainfall events.

Jackson et al. (2018) refer to an unpublished study conducted by HydroNET in 2015 to validate the accuracy of various rainfall data products in Australia, including calibrated radar rainfall data. There are significant differences in precipitation types, with a distinction being made between stratiform and convective precipitation which consequently affects the reflectivity of the radar signal and therefore how the Rainfields 3 QPE algorithm adjusts the mean field bias using the contributing gauges from The Bureau rain gauge network (Jackson et al. 2018).

The R² values (Fig. 9) are an indicator of how well the QPE merged radar rainfall observations fit the observed rain gauge data. The study reported R² values for weather radar merged rainfall data: 0.83 for all events, 0.82 for convective events and 0.9 for stratiform events. The results in Table 2 present a similar range (0.72–0.91) of R² values.

Fig. 9.

Regression analyses graphs for daily rainfall comparisons of QPE merged radar rainfall data v. Bureau rain gauge data for all of the rainfall events.

3.4. Comparison of daily rainfall – Longerenong DATA Farm rain gauge network with radar merged rainfall maps

Rainfall data pairs from each of the 33 rain gauges installed at the Longerenong DATA Farm and the QPE weather radar merged rainfall were analysed for each of the five rainfall events, with 14 outliers removed, leaving a total of 151 data pairs. The purpose of this analysis was to illustrate the spatial variability of a dense network of single-point daily rainfall rain gauge measurements with the QPE merged radar rainfall product at the same location. The analysis of this data must consider that the Longerenong DATA Farm rain gauge network only covers a relatively small area of ~5 × 5 km (2500 ha). This privately managed network consists of low-cost tipping bucket rain gauges that do not meet WMO standards for the installation and maintenance of rain gauges as required for The Bureau site rain gauges. As a minimum, the rain gauge manufacturer recommends annual maintenance and checking to ensure that gauges remain accurate. Consequently, the expected levels of accuracy may be somewhat less than Bureau site rain gauges.

The results of the analysis described both in Fig. 10 and Table 3 show no overall statistical relationship between G_obs and R_obs. This is particularly evident in the results for Lin’s CCC, with values ranging from −0.0851 to 0.1486, indicating strong discordance between G_obs and R_obs for the five rainfall events. The NSE values range of 0.0361–0.5002 indicates unsatisfactory levels of performance. The Pearson coefficients, ranging from −0.5174 to 0.271 for the first four events indicate very small correlations, whereas a value of 0.7072 for the 8 June event indicates a strong correlation. However, Fig. 10 and 11 show a large spread of the residuals, which does not support this conclusion.

Fig. 10.

Comparison of daily rainfall measured by the Longerenong DATA Farm rain gauge network with determinations of daily rainfall from the QPE merged radar rainfall maps for the five rainfall events.

Fig. 11.

Comparison of daily rainfall measured by the Longerenong DATA Farm rain gauge network with determinations of daily rainfall from the QPE merged radar rainfall maps for the five rainfall events using linear regression.

The other four rainfall events experienced more evenly spread rainfall as implied by the QPE merged radar rainfall observations (Fig. 11). The R² values (0.0361–0.5002) also indicate that the radar observations are a very poor fit for the observed rain gauge data from the Longerenong DATA Farm rain gauge network.

The data in Table 4 compare single-point values at the location of The Bureau Weather Station 79028, for each of the rainfall events chosen for the study. The differences between SILO and AGCD are explained by the source data used in their modelling as they use different datasets and calculation methods (Beesley et al. 2009). Fu et al. (2022) state that the mean annual rainfall difference over the Australian continent (1960–2016) is ~2.6%, and that the differences between SILO and AGCD (e.g. 7 April 2023 in Table 4) can be so large for certain rainfall events in certain locations that one must wrong, and it is likely both are inaccurate.

4.1. An alternative to conventional rain gauge installations

Owing to the sparsity of rain gauge sites and the high spatial variability of rainfall at the field level, there is a need by modellers, agronomists and farmers for much improved rainfall data both in density and quality. Grid-based models such as SILO and AGCD are unsatisfactory due to their limited resolution, which can lead to inaccuracies in model outputs. For areas of Australia, the Rainfields 3 QPE merged radar rainfall product provides areal coverage at high resolution and at a significantly reduced cost to the alternative of expanding existing rain gauge networks. Other countries also have similar weather radar installations such as the Operational Program on the Exchange of Weather Radar Information (OPERA) in Europe (Saltikoff et al. 2019) and the Next-Generation Weather Radar (NEXRAD) in the USA (Seo et al. 2011). Ansh Srivastava and Mascaro (2023) studied the spatial frequency analysis of extreme precipitation events with 19 years of merged radar rainfall data generated from the NEXRAD network tested against a network of 204 high-resolution rain gauges in central Arizona, USA, resulting in spatially seamless merged radar rainfall data that addressed the shortcomings from sparse rain gauge measurements.

Even though rain gauge data are point measurements, this is the only source of independent truth available to validate a rainfall event, assuming the gauge meets the WMO standards for the installation and maintenance of rain gauges.

Comparison of sampled areas of coverage for different methods of rainfall estimation is shown in Table 5.

In the case study of the Rainbow radar installation, rainfall data from 26 contributing rain gauges (Table 1) were used to merge with the Rainbow weather radar data, resulting in an average spacing of ~50 km between gauges.

4.2. Localised intense rainfall events

Within the region of the case study, changing patterns of rainfall have become evident with an observed increase in localised summer thunderstorm events of intense rainfall over relatively small areas and of short duration (Pepler et al. 2021).

The rainfall event of 7 April 2023 consisted of localised heavy rainfall (most likely convective), with high spatial variability within the Longerenong DATA Farm rain gauge network (Fig. 13, 14). The rain event occurred in the 9 h after midnight with a peak intensity of ~15 mm h^–1 in the vicinity of the Longerenong Bureau weather station (79028). Large differences of ~8–13 mm of rainfall were evident in the SILO and AGCD modelled rainfall for this event (Beesley et al. 2009).

Fig. 13.

Daily rainfall map of QPE merged radar rainfall for the 24 h preceding 09:00 hours on 7 April 2023 with a resolution of 500- × 500-m pixels. To show the comparison highlighted in Table 4, the coincident SILO tile (and the collocated AGCD tile), including The Bureau Longerenong weather station and Longerenong DATA Farm rain gauge network are superimposed on the map.

Fig. 14.

Map of daily QPE merged radar rainfall from the Rainbow weather radar in 24 h to 09:00 hours on 7 April 2023, with daily rainfall amounts displayed within each 500- × 500-m pixel. The SILO grid is superimposed on the map, displaying the SILO and AGCD modelled rainfall for each 5- × 5-km grid cell.

An example of this localised thunderstorm activity on the evening of 7 January 2022, with the edge of the storm reaching the eastern boundary of the Longerenong DATA Farm, is illustrated in Fig. 15. The official Longerenong Bureau Weather Station, ID 079028, located within the Longerenong Agricultural College (Fig. 15), recorded 0 mm of rainfall on 8 January 2022 (Table 6). The nearest official Bureau rain gauges are Horsham Aerodrome (11.2 km east), which also recorded 0 mm, and Rupanyup (29.9 km west) with 2.8 mm recorded accumulated over 3 days from 8 January 2022.

Fig. 15.

Map of daily QPE merged radar rainfall from the Rainbow weather radar in 24 h to 09:00 hours on 8 January 2022, with daily rainfall amounts displayed within each 500- × 500-m pixel. The SILO grid is superimposed on the map, displaying the SILO and AGCD modelled rainfall for each 5- × 5-km grid cell.

This is an example of the lack of spatial density of rain gauges missing a significant rainfall event with the majority of rainfall greater than 10 mm as indicated by the QPE merged radar rainfall data. The shortcomings of both gridded rainfall models that rely on input from The Bureau rain gauge network are demonstrated with values of 0.0 and 0.1 mm respectively for the SILO and AGCD grid cells that have The Bureau Longerenong rain gauge located within the grid cell (Table 6). The localised thunderstorm was completely missed by both models; however, the weather radar data were able to indicate the spatial distribution of this rainfall event. As mentioned in the Methods section, the Longerenong DATA Farm has a private network of rain gauges, with five located on the southern end of the eastern boundary. Although these rain gauges do not meet WMO standards, they did indicate rainfall readings between 6 and 16 mm, confirming coverage of the rainfall event.

The cited example above is not an isolated one, with the rural media reporting several localised thunderstorm events of agricultural significance that occurred on 2 and 15 March 2022 in an area west of the town of Warracknabeal, Victoria, covered by the Rainbow radar. A Wimmera farmer who farms west of Warracknabeal reported he had a full moisture profile on parts of his farm after the 2 March event, stating that, ‘We got up to 100 mm on some of the farm overnight, but it varied a lot – very hit and miss’ (Verley 2022). The daily rainfall reading for the nearby official Bureau site Warracknabeal Airport, ID 078018, was 0.2 mm for the preceding 24 h to 09:00 hours on 2 March 2022. The SILO grid value for daily rainfall for that field was 17.4 mm. This illustrates how erroneous point-measured rain gauge and grid-based models are, the QPE merged radar rainfall at the very least gives an indication of the spatial variability at the field scale.

4.3. Measurement of accumulated rainfall using QPE merged radar rainfall data

An illustration of how the QPE merged radar rainfall data can be used to aid farm management decisions, especially in dryland (rain-fed) agriculture, is to generate high-resolution maps of the accumulated rainfall for a field between any two selected dates from the archive. An example is shown in Fig. 16, which depicts the accumulated rainfall and the spatial variability in the 4 months pre-sowing, from January to April 2023.

Fig. 16.

(a) Map of the accumulated daily rainfall for the Longerenong DATA Farm computed from the QPE merged radar rainfall product archive for the period 1 January to 30 April 2023. The hashed area is a single field, ‘Paddock 8’, showing the resolution of areal spatial distribution of the accumulated rainfall within the field for the time period. (Map data Google Satellite, WMS, see https://mt0.google.com/vt/lyrs=s&x={x}&y={y}&z={z}, accessed 19 February 2024). The blue square indicates the QPE merged radar rainfall pixel displayed in (b) the accumulated rainfall graph.

Comparative values of the accumulated rainfall for the period 1 January to 30 April 2023 at the case study site are:

As noted previously, Fu et al. (2022) compared the SILO and AGCD in their study, showing that both grids could produce significant differences in rainfall amounts at the local scale.

4.4. Improved output from water balance modelling

In the dryland (rain-fed) scenario, the simple water balance equation for the change in soil moisture S stored in the soil profile over time follows:

where R, T, E and D are rainfall, transpiration, evaporation and drainage respectively, all with equivalent units (mm day^–1).

Water balance studies, especially in dryland farming locations, have mainly relied on gridded rainfall data such as SILO for their rainfall variable input. The provision of the higher resolution merged radar rainfall data accounting for the spatial variability of rainfall at field level, results in higher precision output from the models, e.g. the SoilWat module (Probert et al. 1998) in the Agricultural Production Systems sIMulator (APSIM) model (Holzworth et al. 2014; Holzworth et al. 2018) (Fig. 17). This may lead to better management decisions being made from the improved modelling output by using QPE merged radar rainfall data to better represent the spatial variability of rainfall as opposed to using point-based rain gauges or grid-based data.

Fig. 17.

APSIM modelling of the summed available water content for a soil profile in ‘Paddock 8’ using the SoilWat module with different sources of daily rainfall from QPE, SILO grid and single-point Bureau weather station 70928 rain gauge, to estimate daily rainfall for the time series 1 January to 30 April 2023 inclusive.