2.1.

Site Description

The study was conducted during two rice-growing seasons (2020/2021 and 2021/2022) on a commercial property in the Murrumbidgee Valley near Griffith in South-Eastern Australia (34°17′18″ S, 146°03′03″ E). The soil is a typical rice growing soil in the region, classified as a self-mulching clay with a 30 cm brown A horizon over a dense red B horizon.³⁷ The 0 to 15 cm soil consists of 34% sand, 9% silt, and 56% clay, as reported in Ref. 25. The climate is characterized as temperate, with hot dry summers and low humidity. Mean annual rainfall is 385 mm for Griffith with a typical in-season rainfall of 150 mm and evapotranspiration of 1150 mm.³⁸ The monthly rainfall and mean daily ${\mathrm{ET}}_0$ (mm) for the two growing seasons when the study was conducted are presented in Fig. 1.

Fig. 1

Monthly rainfall (mm) represented by vertical bars on the left axis and daily average ${\mathrm{ET}}_0$ (mm) represented as dots on the right axis during the cropping seasons of 2020–2021 (year 1; red) and 2021–2022 (year 2; blue) at Griffith, NSW.

The field, which is shown in Fig. 2, was a 9-bay/replicate, down the grade border-check layout (34 ha) with replicates 500 to 600 m in length and 50 to 75 m wide ($\sim 3.5\mathrm{ha}$ each). Earthen banks 30 cm high were used to separate replicates. Irrigation water was supplied at the top of the replicate using lay-flat rubber stops controlled using portable automated self-powered AutowichPro’s (Padman Automation, Strathmerton, VIC, Australia). The timing of the first two-three irrigation events was based on agronomic and herbicide requirements to guarantee good rice establishment prior to the installation of soil moisture sensors for monitoring and irrigation decisions. Subsequently, various irrigation deficits were investigated during the remainder of the vegetative period until panicle initiation. Irrigation and agronomic information are detailed in Ref. 25. In brief, short-season semi-dwarf c.v. Viand was sown at a rate of 130 and $120\mathrm{kg}/\mathrm{ha}$ into wheat stubble and bare ground in years 1 and 2, respectively. Nitrogen was applied as granular urea (46% N) at a rate of $220\mathrm{kg}\mathrm{N}/\mathrm{ha}$ in a three-way split. SMT was measured using two watermark sensors (Model 200SS, Irrometer Company inc., Riverside, California, United States) to average results per replicate. Sensors were installed beside plants after establishment to ensure rootzone SMT was measured at a depth of 15 cm below ground. In year 1, sensors were connected to WiField loggers that offer low cost, low power data collection, storage, and transmission to the Google Cloud Platform in real time using an on-farm Wi-Fi network.³⁹ In the second year, solar powered LoRaWAN IoT communication stations (SensorPro’s, Padman Automation, Strathmerton, VIC, Australia) were used to collect and send soil moisture data to the Padman Automation platform for real-time monitoring. SMT data presented in this manuscript are from four replicates in year 1 and five replicates in year 2 in which treatments aimed not to surpass either $-15\mathrm{kPa}$ or $-40\mathrm{kpa}$ during the vegetative period from establishment until panicle initiation as such an irrigation regime was reported by Ref. 25 to achieve sound water productivity without detrimentally delaying crop development.

Fig. 2

Location of the commercial field site where the 2-year study was conducted with NDVI from 06/01/2021 (year 1).

2.2.

Crop Evapotranspiration Estimation

Crop evapotranspiration (ETc) was estimated using the FAO56 method,³² with water requirements calculated as the product of reference evapotranspiration (${\mathrm{ET}}_0$) and a crop coefficient as detailed previously in Eq. (1). Hourly reference evapotranspiration and rainfall were calculated from nearby weather stations using the standardized American Society of Civil Engineers equation. ⁴⁰ Hourly data were used due to the effect that the time of day of irrigation may have in regard to ${\mathrm{ET}}_0$. Crop coefficients were derived using IrriSAT³⁰ (temporal ($\sim 8$ days) and spatial ($10\times 10\mathrm{m}$) resolution), which use Eq. (2) to estimate the actual crop coefficient from satellite multispectral images as³³

Canopy cover was hypothesized to influence the period after an irrigation event at which soil transitioned from a saturated to a non-saturated state. Remote sensed images to measure canopy cover were obtained from cloud-free Sentinel-2 satellite images every 3-4 days with a pixel resolution of 10 m and top of atmosphere reflectance during both growing seasons. The GeoTIFF images of the whole field site were processed in QGIS (Version 3.10), where polygons of $900{\mathrm{m}}^2$ were created at the site of the soil moisture loggers. The zonal statistics tool available in QGIS was used to obtain the mean and standard deviation data for each band at each polygon and used to compute NDVI⁴¹ using Eq. (3), where $R$ represents reflectance at the specific wavelength

2.3.

Model Creation and Validation

The model to estimate SMT (presented later in Sec. 3.2) was created using year 1 data from four-replicates and validated using data from five independent replicates collected in year 2. To evaluate the model performance, the coefficient of determination ($R^2$) and root mean squared error (RMSE) were calculated rather than the percentage error as the percentage can be substantial for SMT near 0 kPa⁴² but of little relevance. RMSE was calculated as per Eq. (4), where $n$ is the number of data points, $y_i$ is the $i$’th measurement, and ${\hat{y}}_i$ is the corresponding predicted measurement:

3.1.

Soil Moisture Tension and Crop Evapotranspiration

SMT during the vegetative stage in relation to the number of days since the pre-emergent (first) irrigation of the individual replicates investigated in this study in years 1 and 2 is presented on the left side of Figs. 3 and 4, respectively. In all cases, SMT returned to 0 kPa after an irrigation event, indicating that sufficient irrigation water was applied to refill the soil profile at the monitored depth. The extent of the soil moisture deficit can be seen to vary between replicates and years as a result of different irrigation thresholds under investigation (detailed in Champness et al.).²⁵ It was due to the possible influence of climatic variability between seasons that data was segmented by year with various deficits in a year to understand if a single model approach could be used. The estimated crop evapotranspiration (with rainfall subtracted) in relation to the number of days since the pre-emergent irrigation of each replicate in years 1 and 2 is presented on the right side of Figs. 3 and 4, respectively. ETc was reset to 0 mm at the time of irrigation or when significant rainfall was recorded.

Fig. 3

Left: SMT (kPa) and right: ETc with rainfall subtracted (mm), in relation to number of days since pre-emergent irrigation with letters (a)–(d) used to distinguish individual replicate data from year 1. Note: IoT loggers to monitor SMT were not installed until establishment; hence data in the establishment phase is not presented, with logger issues in replicate “A” limiting the period that data was collected.

Fig. 4

Left: SMT (kPa) and right: ETc with rainfall subtracted (mm), in relation to number of days since pre-emergent irrigation with letters (a)–(e) used to distinguish individual replicate data from year 2. Note: IoT loggers to monitor SMT were not installed until establishment; hence data in the establishment phase is not presented.

The relationship between SMT and cumulative ETc since the last irrigation for each individual replicate in both years is shown in Fig. 5. When an irrigation event occurred, ETc values were reset to zero. Therefore, within each graph, each line of declining soil moisture represents an individual deficit event within that replicate. For this reason, for certain points on the $x$-axis (ETc) there are multiple points on the $y$-axis (STM). The slope of the individual lines in Fig. 5 represents the relationship between SMT and ETc.

Fig. 5

(a)–(e) Relationship between SMT (kPa) and cumulative ETc with rainfall subtracted (mm) since last irrigation for each replicate. Data from year 1 is presented on the left, and year 2 is on the right.

The point of inflexion on the $x$-axis in Fig. 5, known hereafter as “$c$,” represents the cumulative ETc when SMT began to decline ($<-1\mathrm{kPa}$). The value of $c$ was not consistent between replicates or within replicates (Fig. 5). However, $c$ generally increased within each respective replicate throughout the season. This can be seen in Fig. 6, demonstrating that the time taken for soil to transition from saturated to non-saturated conditions increased throughout the season. This is explained by relatively high soil evaporation rates experienced when little canopy cover exists and is exacerbated with readily available water at the surface immediately after an irrigation or rainfall event.³² The rate of soil drying decreases over time as the vapor pressure at the soil surface nears that of the atmospheric vapor pressure.⁴³ As canopy cover increases throughout the season, an increasing proportion of soil moisture loss is attributed to transpiration and the rate of loss from evaporation decreases, however, not proportionally.³² Therefore, increasing canopy cover reduces the initial rate of soil moisture loss when compared with relatively bare soil, thus increasing the value of $c$ as the rice continues to develop canopy closure.

Fig. 6

Example of the relationship between SMT and cumulative ETc (mm) since last irrigation with different colors used to show how the point of inflection ($c$) increases throughout the season. (blue = 55, green = 65, and red = 84 days after pre-emergent irrigation).

3.2.

Developing Model to Forecast Soil Moisture Tension

To determine the average relationship between SMT and ETc across all replicates in year 1, the $x$-axis of each deficit period for each replicate was translated $c$ units horizontally to create an inflexion at zero [Fig. 7(a)]. In general, the gradient of the individual irrigation events in different replicates is somewhat similar, with an $R^2=0.71$. This demonstrates a relatively consistent soil moisture decline in relation to ETc throughout the period under investigation. This relationship would likely differ in different soil types or under a more severe irrigation regime as the rate of soil evaporation is highest in wet soils and decreases with declining moisture availability.⁴³ The gradient of the linear relationship ($m$) between SMT and cumulative ETc of $-1.0$ implies that, for every mm increase in ETc beyond $c$, SMT declined 1.0 kPa. However, to build a predictive model, determination of the cumulative ETc when $m$ can be applied (point $c$) is required.

Fig. 7

a) Translated graph of SMT (kPa) and ETc with rainfall subtracted (mm) from point $c$ ($\mathrm{kPa}<-1$). Each line represents a deficit period that has been translated $c$ units horizontally with data from all replicates included from year 1. (b) “$c$” versus NDVI in year 1, demonstrating the relationship between canopy cover and the cumulative ETc at which soil moisture begins to decline.

Due to increasing canopy cover slowing the rate of soil evaporation and increasing the time taken for soil to transition from saturated to non-saturated (when soil evaporation is the major influence of soil water decline),^32,43 NDVI was used as a proxy measure to account for the change in canopy cover experienced throughout the season. As hypothesized, $c$ was found to be correlated to NDVI values [$R^2=0.59$, Fig. 7(b)].

Using the relationship between ETc and SMT [$m$, Fig. 7(a)], and the point at which SMT declines [$c$, Fig. 7(b)], a model to predict SMT was obtained by adding the point of inflection to the product of $m$ and cumulative ETc since last irrigation (with rainfall subtracted) according to the following equation:

3.3.

Assessment of Model to Predict Soil Moisture Tension

To assess the robustness of the model across seasons, remote sensed ETc and NDVI data collected from individual replicates in year 2 was used to predict SMT according to Eq. (5) and compared with the measured values presented in Fig. 4. The accuracy of the predicted SMT for the five individual replicates and all data combined is provided in Fig. 8.

Fig. 8

Actual versus predicted SMT in year 2 of five individual replicates (a)–(e) and all data combined from the five replicates presented in panel (f) with a 1:1 line included for perspective. The model was developed from data collected in year 1.

The model developed from data collected in year 1 to predict SMT of five replicates in year 2 achieved an $R^2$ ranging from 0.40 to 0.88 with the combined data achieving 0.70 [Fig. 8(f)]. Of greater importance is the RMSE, which ranged from $\pm 3.5$ to $\pm 7.0\mathrm{kPa}$ in the individual replicates with the combined RMSE $\pm 5.8\mathrm{kPa}$ (Table 1). The reduced accuracy of the model in replicate $C$ is potentially due to soil and therefore crop variability within the area immediately surrounding the SMT sensor. In this area only, a water reservoir previously existed prior to recent filling with coarser textured soil. The model can be seen to underestimate the extent of soil moisture deficit, with the increased frequency of soil moisture deficit events likely a result of the lower water holding capacity in this particular area. This highlights the limitation of the model to be used in other soil types. Although the area was not ideal for SMT sensor placement, this location was selected as the same IoT logger was used to monitor water progress down the replicate during an irrigation event and was required in this location to ensure full irrigation of the replicate while minimizing excessive run-off. This highlights the limitation of point-based sensing as it may not have adequately represented the $900{\mathrm{m}}^2$ area in which the spectral indices were calculated.

Table 1

Individual replicate (A–E) and all data combined (F) R2 and RMSE values 0 to −10  kPa, <−10  kPa and overall, for forecast SMT in year 2 using the model built from year 1 data.

The combined RMSE of $\pm 5.8\mathrm{kPa}$ is considered sufficient for scheduling irrigation in a commercial environment. However, it must be highlighted that after an irrigation event, SMT returned to 0 kPa and remained at or close to that value for some time. Therefore, many measured data points were 0 kPa (58%), with 76% of data points occurring under saturated conditions (0 to $-10\mathrm{kPa}$). Accuracy beyond $-10\mathrm{kPa}$ is of greater importance for irrigators as this is when irrigation decisions are required. RMSE calculated separately for 0 to $-10\mathrm{kPa}$ and beyond $-10\mathrm{kPa}$ is provided in Table 1. From 0 to $-10\mathrm{kPa}$, the RMSE was $\pm 2.8\mathrm{kPa}$ and $\pm 10.7\mathrm{kPa}$ for SMT beyond $-10\mathrm{kPa}$ when all replicate data were combined.

3.4.

Evaluation and Limitations of the Model

The model is shown in Fig. 8 to under-estimate SMT in year 2, indicating that the rate of soil moisture decline in year 2 was greater than in year 1. A comparison between year 1 and year 2 on the rate of soil moisture decline in relation to ETc is shown in Fig. 9, where the gradient, $m$, in year 1 was $-1$, but in year 2 it was found to be $-1.4$. For every mm increase in ETc beyond $c$, SMT declined 1.4 kPa in year 2, as opposed to 1.0 kPa in year 1.

Fig. 9

Translated graph of SMT (kPa) and ETc with rainfall subtracted (mm) from the time $\mathrm{kPa}<-1$. Each line represents a deficit period that has been translated $c$ units horizontally with data from all replicates included from year 1 in red and year 2 in blue.

The different rate of SMT decline is likely a result of the residual stubble from the winter cereal grown prior to rice in year 1 acting as a mulching layer to slow the exchange of water between the soil and atmosphere by buffering the soil surface temperature, reducing the heat flux, and therefore reducing soil evaporation losses in year 1.³² Rice stubble was burned prior to seeding in year 2, a practice known to increase soil evaporation.⁴⁴ This highlights the impact of management practices on soil moisture and therefore a limitation of the model for widespread adoption.

Furthermore, decreasing the frequency of irrigation (extending the SMT beyond the $-40\mathrm{kPa}$ reported here) would likely further increase the rate of decline in SMT compared with ETc due to increased resistance of moisture movement associated with declining soil water.^32,43 However, increasing soil moisture deficit beyond the values analyzed in the current study was found to have adverse yield implications and is not recommended in a temperate Australian environment, as reported in Ref. 25. Nevertheless, the model proved to estimate SMT within an overall RMSE of $\pm 5.8\mathrm{kPa}$ and is considered appropriate to schedule irrigation in this environment. Short term (7-day) ${\mathrm{ET}}_0$ forecasts are considered reasonably accurate in the region, providing opportunity for future research to incorporate forecast ${\mathrm{ET}}_0$ into the model to predict SMT multiple days in advance. This would enable forward scheduling of irrigation, which is required when irrigation water is not available on demand and water ordering requires a lead time. Such forecasting has been demonstrated for use in cotton production by Refs. 36 and 45 and offers potential to be incorporated into an automated irrigation system for water saving rice production. Although generally not an issue in semi-arid regions, cloud cover may limit the accuracy of predictions, with increased frequency of imagery preferable. This may be possible with the incorporation of other satellite-based sensing alternatives; however, the fusion of spectral indices may be required to ensure that properly calibrated NDVI values can be achieved.