The sensors were deployed at the Commonwealth Scientific and Industrial Research Organisation's (CSIRO) Lansdown Research Station, located 50 km south of Townsville, Queensland, Australia (19∘39′42′′ S and 146∘51′12′′ E, elevation 63 m). Paddocks used in this study contained pastures dominated by Urochloa spp., Chloris spp., and Stylosanthes spp. Data were collected over 545 days between 23 September 2011 and 21 March 2013.

Based on daily precipitation and temperature data collected by the Bureau of Meteorology (BoM) from the Woolshed station (approximately 45 km NW of the study site), the tropical climate in the study region is characterized by a wet season from November to April where monsoonal storms bring intermittent periods of heavy rainfall, and a winter dry season with little or no rainfall. The average annual rainfall of 1139 mm falls mainly during the wet season, and the average monthly temperature range is 20.8 to 28.5 ∘C in January and 10.4 to 21.8 ∘C in July.

Each of the two sensor nodes (Fig. 1) were mounted with the same array of equipment (i.e. multispectral sensor, digital camera, soil moisture sensor, wireless networking infrastructure) and provided spatially coincident data with both high temporal and spatial resolution. The nadir-pointing sensors were located at a height of 2.5 m above the ground. At this height the downward-pointing multispectral sensor had a 25∘ field of view (FOV) sensing approximately 0.97 m2 of the area at ground level, although this area changes across the season with vegetation height. The digital camera's FOV was approximately 2.8 m × 2.0 m at ground level and would have been able to capture the 1 × 1 m area with a vegetation height up to approximately 1.5 m. See Balzarolo et al. (2011) for a discussion of optical sensor configurations.

The nodes were approximately 200 m apart in areas of the paddock visually assessed to be similar at the time of installation. One node was unfenced, permitting access to the area under the node by cattle grazing in the paddock. The second node was enclosed by a 30 m × 30 m fence, which excluded cattle from grazing within the enclosure, but allowed access by kangaroos and other small herbivores. The decision to place only one of the nodes within a grazing exclosure was made to improve the likelihood that the vegetation observed in each node would be at different heights. Although the paddocks were grazed by beef cattle for short periods during the sensor deployment, due to the lack of feed in the paddocks at those times and the low grazing pressures there ultimately was no discernible difference in vegetation height before and after the grazing.

Each node included a solar-powered sensor hub which relayed captured sensor data to a wireless sensor network (WSN) installed on the research farm via an internet connection to a centralized enterprise database. All equipment was temporarily removed for a week during a controlled property burn in mid-December 2011.

A Decagon 5TM soil moisture sensor (Decagon Devices, USA) was installed at each node to monitor the volumetric water content (VWC) of the soil. The VWC is the volume of water per unit of total volume, determined by measuring the dielectric constant of the soil, as well as soil temperature from a thermistor. The 5TM sensors were buried at a depth of 15 cm under the soil surface below the multispectral sensors. This depth was used to capture soil moisture near the surface, yet reduce the possibility of damage from trampling by cattle. The 5TM sensors recorded soil moisture and soil temperature readings at 1 min intervals. We extracted an average of VMC for the period between 12:00 and 13:00 (all times are local time) for each day, resulting in a time series of daily VWC (i.e. SoilMoisture) and soil temperature data during the study period.

The nearest BoM weather stations were at Woolshed, Charters Towers Airport (both inland), and Townsville Airport (coastal), approximately 45 km NW, 70 km SW, and 45 km N of the study site respectively. Daily maximum ambient temperature averaged for the two inland stations had a strong relationship with temperature data from 12:00 from the 5TM soil moisture sensors, so these data sets were used interchangeably. The 5TM soil moisture sensors were additionally used as the main source of soil moisture data.

At the time of this study a new meteorological station at the Lansdown Research Station had recently been installed, but the data were not available for the study period. The national interpolated climate surfaces from BoM were thought to be too coarse for our small study site as precipitation events are typically spatial heterogeneous. Instead, a comparison of data from nearby BoM stations with the in situ soil moisture sensors at our nodes showed a strong correlation with the average of the precipitation recorded at Charters Towers Airport and Townsville Airport stations (Pearson product–moment correlation coefficient of 0.61 during the wet season period of data collection).This average precipitation was therefore used as the best option, as the only alternative was to use an interpolated data set.

The start and end of the wet season were determined using a method designed for the northern Australian climate (Lo et al., 2007) in which the start of the wet season is defined as the date after 1 September when 50 mm of precipitation has accumulated. Bureau of Meteorology precipitation data from the Townsville Airport station were used to define the start and end of the wet and dry seasons, as this station had the most complete time series of the nearby stations. Using this method, the 2011/2012 wet season at our study site started on the 5 December 2011 and the 2012/2013 wet season started on 1 January 2013.

Digital cameras were deployed at the study site to provide an automated assessment of ground cover (see Zerger et al., 2012) to serve as a visual cross-check of the multispectral data and assist in identifying surface water. At each of the two nodes we deployed a Pentax Optio WG-1 digital camera in a downward-pointing position, centred on the area sensed by the Skye sensors so that the images overlapped the FOV of the multispectral sensors.

This camera model was selected as it was inexpensive, weatherproof, and had an inbuilt intervalometer to enable automatic shooting at fixed intervals. At 2.5 m the 13.8 megapixel digital cameras recorded images with an approximate resolution of 0.6 mm at the ground. The cameras were configured with flash off, sensitivity at ISO 200, autofocus, and automatic white balance enabled. The decision to use an automatic white balance was based on similar studies (e.g. Macfarlane and Ogden, 2012), although other studies have used a manual/fixed white balance in order to minimize changes in illumination (e.g. Toomey et al., 2015; Sonnentag et al., 2012). Digital images (approximately 1 to 4 MB each) were captured every 30 mins and were manually downloaded at approximately 2-week intervals.

The images from the cameras contained uncalibrated red, green, and blue (RGB) spectral bands. There has been extensive work on automated and semi-automated classification of such time series of digital photographs for the purposes of vegetation monitoring (e.g. Ewing and Horton, 1999; Karcher and Richardson, 2005; Bennett et al., 2000). As the focus of the current study was on the calibration of the multispectral sensor data, we chose to use a semi-automated method, VegMeasure (Johnson et al., 2003), to extract a green cover fraction from the time series of digital camera images at each node. VegMeasure has been utilized and validated in a number of studies (e.g. Booth et al., 2005; Louhaichi et al., 2001) and provides a rapid method for classifying a series of images into green and non-green using the green leaf algorithm (GLA). The GLA also acts as an alternative sensor measurement of green fraction to that derived from the multispectral data set.

The GLA protocol requires deriving a single threshold value from a single image, which is then applied across the whole time series of camera images. The GLA applies the following spectral band ratio (Louhaichi et al., 2001): (G-R)+(G-B)(G+R+G+B), where G is the digital number of the green band, R is the digital number of the red band, and B is the digital number of the blue band. The proportion of the pixels in each image, in which the band ratio exceeds a user-defined threshold, is reported as the GLA.

For each day in the study period, the camera image taken nearest in time to 12:00 was selected to minimize shadows and to ensure as consistent an illumination as possible, and the time series was quality controlled for days when there was site maintenance work under the node. One photo with a mix of PV and NPV vegetation was manually selected as a calibration image (14 May 2012, 12:13:55, on the unfenced node). To derive a threshold value for the GLA, one hundred random points were identified using the “calibrate threshold” function in the VegMeasure software, and assigned to two classes: “white”, which is green vegetation, and “black”, which is non-green vegetation and background material including litter and soil. The resulting GLA threshold of 0.095 was verified using a random selection of images and was then applied across the whole time series of camera images to extract the green proportion. The single threshold value used in deriving the GLA is a necessary feature of using the GLA, as well as having been applied in other vegetation studies (as cited).

We used a paired sensor set-up (Fig. 1) with the downward-pointing sensor having a conical field of FOV of 25∘ as indicated by the manufacturer, allowing it to sense reflected light only from the ground directly beneath the sensor. The upward-pointing sensor was fitted with a cosine diffusing filter to alter its FOV to a full hemispherical view, permitting the albedo of the surface to be assessed relative to the incident solar radiation. Sensors were checked and cleaned fortnightly and the sensor station was coated with insecticide to deter crawling and flying insects.

The multispectral sensors mounted on each of the two nodes were paired Skye SKR-1850 four-band weatherproof sensors (Skye-Instruments, 2012b), which were calibrated individually by Skye, with band choices based on our specifications. Each sensor was configured with bands in the green (0.545 to 0.547 µm), red (0.644 to 0.646 µm), near infrared (NIR) (0.834 to 0.837 µm) and the lower SWIR (1.028 to 1.029 µm) spectral range (wavelengths in brackets indicate band widths). These bands were chosen as the NIR region of the electromagnetic spectrum is widely used in monitoring vegetation “greenness” from multispectral sensors (Tucker, 1979), and the SWIR region is sensitive to plant moisture content (Tucker, 1980). Both the SWIR and upper NIR spectral data can be used to help differentiate PV from both NPV and soil (Asner, 1998), and broad-band SWIR indices have been used to capture seasonally varying NPV proportions resulting from repeat grazing of pastures by livestock (Handcock et al., 2008). We were not able to choose the fourth sensor to be in the 1.55–1.75 µm range recommended by (Tucker, 1980), but were limited to using the longest wavelength possible for this sensor configuration to try and capture senescing vegetation. The band choice was verified before sensor creation by comparing the band to reflectance for green and dry pastures from the Advanced Spaceborne Thermal Emission and Reflection (ASTER) spectral library (Baldridge et al., 2009). This comparison confirmed that, while the discrimination between green and dry pastures is not as distinct at 1.029 µm compared to that at 1.55–1.75 µm, there was still enough potential for discrimination to confirm this wavelength choice for the fourth band.

The NIR region is sensitive to vegetation “vigour” or “greenness”, and vegetation indices, such as the widely used normalized difference vegetation index (NDVI) (Tucker, 1979), utilize the NIR spectral range. A variety of vegetation indices are possible from combinations of the four broad spectral bands of our Skye sensors. Due to the algebraic complexity of calculating indices from this particular Skye sensor model (see the description in the paragraph below), our index choice was limited to simple ratios and normalized difference band ratios (Jackson and Huete, 1991), which we derived to highlight seasonal aspects of the green and dry mix of the tropical pastures (Table 1).

The Skye sensors returned a calibrated numeric output for each spectral band every minute, and data volumes were small enough to be transmitted in near real time via the WSN. After calibrating raw sensor data using individual Skye sensor calibration coefficients, vegetation indices were calculated. The Skye SKR-1850 sensor does not permit the calculation of reflectance directly from the raw current. Instead, Skye provides formulae which use the measured sensitivities of the individual sensors to calculate ratio-style indices such as NDVI (Skye-Instruments, 2012a). These indices are mathematically equivalent to those calculated from reflectance. Using the NDVI example from Skye, we developed formulae for the vegetation indices shown in Table 1.

We illustrate the types of processing required for high-frequency multispectral time series with an example of a typical diurnal time series of multispectral data with a reading every minute (Fig. 2). Both raw sensor current and the calculated NDVI values are typically low during the night-time hours. The period of rapidly increasing sensor values at dawn is extremely noisy due to variable early morning illumination and the scattering of sunlight through a thicker atmosphere at low elevations. At dusk this pattern of sensor values is reversed (data not shown), which is also seen in Weber et al. (2008, Fig. 3a). Apart from the spike in high NDVI when a green leaf was held in front of the sensor (approximately 13:00), the middle part of the day is the period of relatively stable values of NDVI, with only random variations that occur due to the noise in the raw current or from ephemeral variations in illumination such as from sun glint.

For the entire time series of multispectral sensor data taken every minute, a time series of daily values was determined by selecting the vegetation index values from the middle part of the day (12:00 to 13:00) and calculating the median value to reduce noise due to small fluctuations in illumination. Data from a particular day were discarded if they met any of the four categories of filtering criteria listed in Table 2. Data were not discarded under conditions where changes in the spectral values were considered to be a signal rather than noise. For example, rapid increases in NDVI values over time corresponded to rapid growth at the start of the wet season, so were not filtered. Questionable multispectral data were also visually verified against the digital camera images. In developing these filtering rules, the vegetation indices stood as proxy for their individual constituent bands since, as discussed, it was not possible to use spectral reflectance from the Skye SKR-1850 sensors directly. Table 2 is divided into four different filtering categories as follows.

The first category of filtering criteria (Table 2a) were developed to screen the daily multispectral data series for large fluctuations such as data outliers, spikes, high noise levels, data out of range, clipping, and calibration issues, which can commonly result from anomalies at the sensor or during data transmission (Collins et al., 2006; Ni et al., 2009). For example, the night-time raw current reading should remain relatively constant, excluding minor night-time light reflections or electronic noise. Large deviations from night-time baseline current values indicate a technical issue. Such issues were identified from the night-time (00:00 to 01:00) median value of raw current by flagging where one or more of the multispectral sensor bands in the paired node had a night-time reading greater than 10 000 mV or where these values were greater than 3 standard deviations from the band mean value. The daytime (12:00 to 13:00) median value of the multispectral indices was also used to identify data quality issues, for example where NDVI was not between 0 and 0.1. This threshold value of NDVI was chosen based on typical values for this environment (Holben, 1986; Jackson and Huete, 1991) and would have to be adjusted if the sensors were deployed elsewhere, for example to monitor snow and ice, which may have negative NDVI values. Data were also masked when the daytime RatioNS34 dropped to zero but within 1 day had returned to its previous value. All instances where the RatioNS34 remained at zero for more than 1 day were visually cross-checked with the deployment records to see if this indicated sensor failure or some other issue such as an insect infestation.

The second category of filtering criteria (Table 2b) is for logistical and physical issues. For example, the data for a day was screened if there was a maintenance ladder underneath the sensor. When a sensor was swapped for new equipment, it was required that a new baseline current value be used in calculations that use raw current. A flag was also set to indicate days for which there was no data during the midday period from one or more of the sensors, which would restrict the calculation of a full suite of indices.

The third category of filtering criteria (see Table 2c) covers filtering rules based on the expected spectral response of tropical pastures, for example, if NDVI was less than zero. This flag is a companion test to the range tested in Table 2a, as it flags NDVI ranges that may indicate catastrophic failure of the sensor resulting in values extremely out of range. All of these cases were visually examined through the photographs and by inspecting the sensor infrastructure during site visits. Other indices were also used for testing data out of range. For example, if RatioNS34 values were greater than 2, it indicated a technical error as pastures should not have values in this range. This filtering rule would need to be adjusted if the sensors were deployed to a different environment. When values of gNDVI were less than 0 or values of NVI-GR were greater than -0.10, and the date and weather data indicated that the readings were made in the dry season, this again indicated values that were out of range rather than due to wet season surface water.

The fourth category of filtering criteria (Table 2d) covered filtering rules where valid spectral signals were excluded, not because they were errors, but because they covered physical conditions which were not applicable to our goal of monitoring pastures. For example, surface water under the vegetation due to heavy rainfall was identified by visual inspection of the camera images combined with the soil moisture data and filtered because it was not a valid measurement of the pasture status, even though it was a valid sensor signal.

In designing the field sampling for this project, it was necessary to balance the project goals with staff resources and logistics of travelling to the remote site every 2–3 weeks for the multiple years of the sensor deployment. All field observation methods were designed to be quickly completed by field technicians during these visits, while also maintaining the technical infrastructure of the sensor deployment. This trade-off between time and resources (Catchpole and Wheeler, 1992) resulted in field observations successfully being obtained over the multiple years of the study, but also resulted in a large degree of uncertainty in the field observations.

During the study period there were 32 visits to the study site to make field observations. All the measurements were made by the same two field technicians, with the majority (71 %) by one technician. Where possible, measurements were repeated by both of the main technicians or other staff (6 days). For the 45 % of days where more than one technician made measurements, the data from that day was averaged. Visual examination of the raw field data noted no systematic differences between the data collected by the different field technicians, so measurements were not further controlled for operator differences. All observations were made within the sensors FOV in a 1 m × 1 m area under the sensors identified by small pegs hidden by the vegetation.

The goal of this part of the project was to assess whether the sensors were able to deliver a reliable source of data that can be calibrated to biophysical values. Our goal was not to develop definitive relationships for prediction purposes, as the quality and volume of the field data is not sufficient for that purpose. We instead assess only the strength of the relationship between the sensor and field data, and do this separately for data from the wet and dry seasons and across the whole year. We use these results to recommend when and how data should be collected in a full sensor deployment for on-farm monitoring.

Data from the two nodes were combined as there were no discernible differences between the fenced and unfenced samples due to grazing of the pastures by cattle. Of the original 33 days of field measurements from across the whole project, Table 3a shows the number of days on which the field sampled data matched the filtered sensor data at each node. Data subsets were also created for the wet season period from January to April (days 1 to 130 of the year), and the dry season (May to December) (Table 3b). The remainder of the field samples were made during periods in which the sensor data were filtered using the rules in Table 2 and so could not be used for further analysis.

The final group of independent variables included vegetation indices derived from the filtered daily data set from the multispectral sensors (i.e. NDVI, gNDVI, NVI-GR, NVI-SR, and RatioNS34) and the digital cameras (i.e. GLA). The dependent variables were the visual biophysical measurements and other observations of the pasture status made at the field sites (TotalBiomass, %BareGround, %Litter2D, %TotalVegetation2D, %Green2D, %Green3D, and VegetationHeight).

A common problem in calibrating and validating models between remote sensing and field data is the small number of field samples and the inherent variability in biophysical data, resulting in models that are not robust (Richter et al., 2012; Harrell Jr. et al., 1996). Richter et al. (2012) provide a good overview of statistical techniques useful for such data sets, including the use of cross validation and bootstrapping methods for model development and validation. Bootstrapping is a non-parametric method that does not assume normality of the data set, making it suitable for developing robust estimates of the population from limited sample data such as in the present study. The estimated model coefficients are assumed to be the best estimates of the population values (Harrell Jr. et al., 1996), of which our field observations are just one sample of the entire population. The advantage of the bootstrapping method is that the entire data set can be used to assess the model performance in the one process, rather than having to split it to create a validation subsample (Harrell Jr. et al., 1996). The distribution of model parameters resulting from the bootstrapping allows the confidence intervals and standard errors of the model parameters to be estimated (Peters and Freedman, 1984).

In the bootstrapping method, a sample is drawn from the original data set with replacement, meaning that each individual datum is selected from the whole data set and so could be drawn multiple times. For each sample, the desired model is fitted between the dependent and independent variables, and their model coefficients are determined. The sampling and modelling process is repeated many times, with 200 being the minimum recommended by Steyerberg et al. (2001). The result is a distribution of the selected model parameters from which the robust estimates of the model parameters and confidence intervals can be made.

The bootstrapping approach is particularly suited to our pilot study because we are interested in the strength of the relationships between the sensor data rather than their form. The approach also addresses the main issue with the visual assessment of pasture status, which is the high degree of uncertainty in that data. The bootstrap method replicates all uncertainty in the analysis, including operator error, uncertainty in the field observations, and that from the flexibility of the statistical model, allowing the confidence intervals around the model parameters to be assessed (Davison and Hinkley, 1997). The method is robust in cases where one variable has missing data, such as when the filtering of our spectral data resulted in field data which did not have matching sensor data.

We therefore applied a bootstrapping method to assess the strength of the relationship between the sensor and field data and the uncertainty around the model parameters. All analysis was made using the R statistical package (R-Core-Team, 2013). We used the mgcv library in R (Wood, 2011) to fit generalized additive models (GAM) (Hastie and Tibshirani, 1990) with a maximum possible dimension of four. GAMs do not assume a linear relationship, but instead use a non-parametric method to fit a model with the highest dimension possible given constraints of small data sets and missing data. The bootstrap was implemented using the boots library in R (Davison and Hinkley, 1997) with 2000 model runs and a “pivotal” method. This bootstrapping method was applied to all combinations of observations of pasture status and a single independent sensor variable.

As the multispectral measurements were made every minute, the data collection from the two nodes represents a possible 1 569 600 sets of the 8 raw current values. As a result of the rigorous data cleaning using the criteria in Table 2, for the 545 days of data collected at each node, 48 % of days of data from the unfenced node and 63 % of days of data from the fenced node were discarded. This large number of filtered days of data reflects the experimental nature of the pilot deployment of the sensors, which resulted in technical and environmental issues with the sensor deployment. However, the rigorous data cleaning we applied was necessary to ensure quality data for model development.

Figure 3 illustrates this data cleaning by showing the time series of NDVI values from the unfenced node, before (raw) and after filtering. In comparison to the digital cameras, the design of the housing for the Skye SKR-1850 sensors led to significant problems with insects such as mud wasps nesting in the sensor tubes (Fig. 4a–b), spiders building webs across the sensor openings, and water ingress below the cosine correction filters, which were fitted to the upward-pointing sensors.

The field observations made at each of the two nodes (Fig. 5) illustrate the rapid vegetation growth at the start of the wet season followed by senescence during the dry season. During the 2011–2012 wet season the TotalBiomass observed at the two nodes had similar values (Fig. 5a), despite the recognized uncertainty in these measurements. Having initially similar pasture biomass was not unexpected as the nodes were sited in an area of the paddock with similar vegetation. Although we had fenced one node with the intention of increasing the range of pasture height being monitored, due to the limited feed availability in the paddocks and the low grazing pressure, these grazing events had negligible impact on the pastures and were not considered further in the analysis. At the end of the 2011–2012 wet season the TotalBiomass observed at each node became markedly dissimilar, with differences of almost 2000 kg DM ha-1 between the nodes, and as expected the difference continues during the rest of dry season as there is no rain to promote vegetation growth. This difference in the pasture biomass between the nodes illustrates the heterogeneous nature of these pastures, where a small change in the type, size, shape, and density of the vegetation growing under a node resulted in large biomass differences. It also highlights why pasture measurements made in the area surrounding the node may not be representative of what the sensor FOV observes.

The time series of VegetationHeight (Fig. 5b) shows a similar pattern to TotalBiomass, but the differences between the nodes are less distinct. VegetationHeight also exhibits more variability between measurements despite being a quantitative measurement made with a ruler rather than a visual estimate. In contrast, the observations of %Green2D and %Green3D (Fig. 5c and d) are comparatively similar between the two nodes, except for the period of June to July 2012. As shown in the images in Fig. 6, the vegetation is tall, mixed, senesced, and increasingly lodged (i.e. no longer erect), resulting in increased variation in the observed values between the nodes.

Over the 545-day study period, the digital cameras captured 22 642 images at the unfenced node and 23 210 from the fenced node. Data capture from the cameras was more reliable than for the multispectral sensors with the loss of only 13 days of data from the unfenced node (3 %) and 10 days of data from the fenced node (2 %), both due to data card failure. A month of digital camera images was also lost in a post-capture storage malfunction, so is not counted as being a deployment-related data loss.

Figure 6 shows a time series of images from the digital camera at the fenced node, with each week represented by one image taken at approximately 12:00. The seasonal progression of vegetation is clearly illustrated by these images, from the new green growth of the vegetation at the start of the wet season, followed by senescence during the move into the dry season, and the sudden removal of all vegetation following the 2011 controlled burn. The camera images again illustrate how, as the wet season progresses, the tall grasses dominate the canopy followed by the gradual drying of the canopy in the transition into the dry season.

Figure 7 shows the daily time series of GLA calculated from digital camera images at each node. These results show that the digital cameras and GLA can successfully capture the seasonal changes in green vegetation, corresponding with the rapid growth of green vegetation at the start of the wet season followed by a decrease to zero during the dry season.

Table 4 and Fig. 8 show the bias-adjusted bootstrap point estimates and the lower and upper bound of the 95 % pivotal bootstrap confidence intervals for the distributions of R2. These distributions are from bootstrapping the GAMs for all combinations of sensor-derived indices and field observations, which were made of all data, as well as for the data subsets from the wet or dry seasons. As the bias-adjusted bootstrap point estimates of R2 are a more conservative estimate than the mean R2 of the modelled distribution, there are times when its value is negative or less than the lower bound of the 95 % pivotal bootstrap confidence interval. This occurred most frequently for the dry season data for which the model fits are generally poor (Table 4). The graphs in Fig. 8 clearly show how the various uncertainties in the study, and in particular the high uncertainty in the field observations, has resulted in wide confidence intervals for many of the models explored using the bootstrapping methodology.

The relationships between sensor and field observations for the whole year and dry season period generally performed poorly compared to those from the wet season. These results are not unexpected as the vegetation between the wet and dry season in this environment is distinctly different. The exceptions were for %Green3D (Fig. 8e) and %Green2D (Fig. 8f), which for all sensor-derived indices except RatioNS34 had strong relationships to data from the whole year and dry season. The bootstrapping analysis for %Green2D was not able to determine model parameters due to the boundary conditions inherent in those subsets of data values.

Across all time periods, the strongest relationships between the multispectral sensor and pasture observations were for the wet season data for %Green3D (Fig. 8e) and %Green2D (Fig. 8f). For all variables, %Litter2D (Fig. 8c) showed the weakest relationships with the sensor variables, and %TotalVegetation2D (Fig. 8d) showed only weak relationships. For the other pasture observations there were good relationships with at least one sensor variable. For example, the bias-adjusted bootstrap point estimates of R2 for the wet season data between TotalBiomass and NVI-SR were 0.72 (95 % CI of 0.28 to 0.98) (Fig. 8a), %BareGround and gNDVI were 0.65 (95 % CI of 0.09 to 0.92) (Fig. 8b), %Green3D and RatioNS34 were 0.81 (95 % CI of 0.53 to 1.00) (Fig. 8e), and VegetationHeight and NVI-SR were 0.66 (95 % CI of 0.19 to 0.95) (Fig. 8g). Excluding the relationships for %Litter2D, for four of the other pasture observations, the NVI-SR index had the strongest relationships to four different pasture characteristics, with RatioNS34 for one variable (%Green3D, Fig. 8e) and gNDVI for one variable (%BareGround, Fig. 8b).

Across almost all time periods, the relationship between the image-derived GLA were weaker than those from the multispectral sensor data. The one example in which the GLA outperformed the multispectral sensors was also the strongest relationship in all data and periods, being for data from the whole year, and between %Green3D (Fig. 8e) and %Green2D (Fig. 8f). These results show that the GLA method to extract green fractions from the digital camera images was very successful in this environment.

In this study, the time series of images from the digital cameras and multispectral sensors at each node clearly captured the changes in the tropical pastures from the period of green-up at the start of the wet season, the period of green vegetation growth during the wet season, and the gradual senescence and drying off of the vegetation. Even given the obvious limitations with the observations of pasture status in this study, it is clear that there are stronger relationships during the wet season than during the dry season or for the whole year. The generally poor relationships between the sensor and field observations outside of the wet season are not surprising since NPV is difficult to discern in the NIR spectral region. The SWIR band of our multispectral sensors was also in the lower part of the SWIR range (1.029 µm), which is not as responsive to dry vegetation as the longer SWIR region of the visible to near-infrared (i.e. 1.55–1.75 µm), which Tucker (1980) recommends for the remote sensing of plant canopy water status. Even if the issues with the field data quality are overcome in a future deployment, it is unlikely that the relationships between field and sensor data will improve for the dry season unless the choice of spectral bands in a future deployment was made to improve sensitively to NPV.

The results of using the bootstrapping method to explore the relationship between the pasture observations shows that the various measures of fractional cover could be successfully predicted from various indices calculated from either the multispectral sensors or the digital camera data. These results are encouraging for additional studies exploring these relationships further.

These results also showed the GLA derived from the digital images to be a useful parameter, with strong relationships to the field observations of %Green3D and %Green2D. They also support the utility of including a SWIR band in the multispectral sensors, with data from our multispectral band in the lower SWIR giving encouraging results.

The vegetation indices from the multispectral sensors were a better predictor of %BareGround than the GLA from the digital cameras. These results indicate that while both sensor types are suitable for monitoring aspects of fractional cover in this tropical pasture system, alternative indices extracted from the digital cameras would need to be explored to improve how well %BareGround can be monitored. Both sensors view the canopy in two dimensions, with the GLA focussed on the green proportion of the canopy while the band choice for multispectral indices can be made to capture both PV and NPV.

Fractional cover has the potential to be a valuable part of a multiple data source approach to providing on-farm data to farmers for sustainable pasture management. Although fractional cover is widely used in landscape degradation studies, particularly in regional monitoring (Richardson et al., 2007; Myneni and Williams, 1994; Guerschman et al., 2009), it is a more recent measurement compared to pasture biomass, which has long been used in livestock production systems. Fractional cover is therefore a less familiar measurement than biomass to interpret and use. However, as fractional cover measurements become more widely available (e.g. Guerschman et al., 2009) and examples of its use in operational farm management increase, it is likely that this will change, as occurred when NDVI started to be used in agriculture. Sensor nodes that monitored fractional cover could be strategically placed in sensitive areas to monitor areas that are becoming overgrazed, for example to signal an alert to move livestock.

Although the period at the end of the wet season is critical for on-farm decision-making, we recommend that to improve understanding of the rate of change of the pasture conditions, monitoring also be made throughout the wet season that precedes it and into the start of the dry season. One of the benefits of a data flow from proximal sensors is to understand the rate of seasonal changes and identify any periods in which the pasture conditions change rapidly or suddenly in response to weather or environmental events.

From this pilot project it is still unclear whether pasture biomass could be predicted with sufficient accuracy in this environment to allow the measurements to be used operationally in on-farm decision-making. However, the results of the present study are encouraging enough to show that further work is warranted. Assuming that the issues with the field data quality can be addressed in future work, it is expected that the relationships between field and sensor data will improve.

This study was run for less than 2 years, and covers only the limited range of pasture conditions resulting from interannual variability in climate and differing grazing and pasture management. If further studies do not show consistent relationships between sites and years, one option for calibration would be to have the farmer performing a controlled set of calibration measurements once or twice during the growing season to calibrate a particular sensor deployment. Having to make pasture measurements would require additional time from labour-poor beef producers. However, by gathering this data at the geographical location of the deployed sensors, these measurements would alleviate the cost of a much larger project. This larger project would require gathering the volume of calibration data required to develop models that would be robust for different geographical locations and different weather conditions between years, and address any recalibration requirements of the physical sensor over time. Alternatively, the time series of vegetation index data from the sensors could be used without calibration to a quantitative value, which would still provide data to indicate sudden changes in vegetation growth.

It is clear that the accuracy of field observations of pasture status could be improved for future sensor deployments aimed at developing qualitative relationships between sensor and field data. In the context of the present study, the uncertainty in our field observations does not change the main outcomes of the project, which are to illustrate practical issues around the sensor deployment, and the methods necessary for the quality control of the sensor data, necessary for designing future deployments.

We recommend that future deployments use non-destructive sampling methods such as BOTANAL, which includes a protocol for assessing and maintaining the accuracy of visual measurements of pasture biomass and composition (Tothill et al., 1992; Orchard et al., 2000). Alternatively, visual assessments could be calibrated by developing a site-specific set of reference photographs at different times in the growing season. The reference photos would be calibrated using pasture cuts (if possible for the vegetation type) and used for repeat training of field staff. This method has the advantage of controlling the data range and the biomass interval between photo standards. Pasture assessments of this type are time intensive, which could be mitigated by targeting data collections at key times during the year. It would also be useful to make additional measurements in the vicinity of the node FOV to assess the spatial variability of pastures in the surrounding area.

In the extensive database cleaning illustrated in Fig. 3 and Table 2, we focussed on post collection filtering methods, as the experimental nature of our deployment meant that data could not be screened in real time. In an operational system additional rules and approaches could be implemented on the node, such as for sensor data cleaning and outlier detection (e.g. Basu and Meckesheimer, 2007; Huemmrich et al., 1999; Liu et al., 2004), and including implementing data quality control algorithms within the WSN (e.g. Collins et al., 2006; Jeffery et al., 2006; Zhang et al., 2010). In addition to the data-cleaning rules we developed, and as the field deployment progressed, we modified the sensor maintenance protocols and infrastructure. This knowledge can also be used in future deployments.

Due to our stringent data-cleaning protocols, a large amount of data from the multispectral sensors was excluded by a combination of automatic and manual methods. In future deployments additional automatic data filtering could be implemented, for example using spectral information to filter data when surface water is present. Developing automatic filtering rules for surface water was not considered necessary in our study as visual examination of the digital camera images identified only 9 days of surface water at the fenced node and 20 days at the unfenced node. The data were excluded manually, particularly as this surface water occurred when there was water incursion into the sensor housing and the whole data period was suspect. For sensor deployments in conditions with more surface water, such as in areas of flood irrigation, having an automatic rule for surface water detection would be useful.

We found the digital cameras to be more robust than the multispectral sensors in terms of data flow, with up to 63 % of days of data from our Skye sensors being discarded during data quality control. Although the stringent filter criteria (Table 2) may have resulted in some “clean” data being excluded, this was balanced against the greater impact of having untrustworthy data for modelling. The long periods of erroneous multispectral data showed that this Skye SKR-1850 sensor model was unreliable in the environment. In comparison to the digital camera, the design of the Skye sensors led to significant problems, including insect infestations in the sensor tubes and water ingress below the cosine correction filters which were fitted to the upward-pointing sensors.

While we were able to mitigate the effects of these issues by regular maintenance of the sensors and post-acquisition data cleaning, we found that the Skye SKR-1850 sensor model was not stable enough in our tropical environment for an operational deployment on a farm. For example, we had a complete failure of one sensor which had water incursion into the sensor enclosure at the point where the wiring attached to the sensor, despite sealant being applied to the connection and the connections being regularly monitored. Given that we had a spare sensor that could be used as a replacement, the decision was made to swap the sensors out to ensure continuity of data collection, while the sensor was returned to the manufacturer for examination.

The new and improved designs for the Skye sensor housing are likely to address many of these issues by having a covered sensor face and also being able to calculate reflectance directly (e.g. the SKR 1860D 4 channel sensor design. Skye-Instruments (2013). Repeating this study with the newer sensor design would allow the focus of future studies to be on gathering multispectral measurements, not on checking and managing the technical aspects of the field deployment or on post-collection data filtering. In situations where only the earlier model Skye sensors are available for use, it may be possible to use a method employed by Harris et al. (2014), who were able to overcome similar limitations of earlier models of a SKR-1800 sensor by using a cross-calibration method between the upward- and downward-pointing sensors to retrieve reflectance. While not recommended by the manufacturer, such a method would be useful for deployments in which the calibration certificates had expired or where reflectance is a requirement.

Cross calibration of sensors could also be useful in situations in which there is a mix of sensor types deployed to capture spatial variability in the landscape. The growing availability of lower-cost sensors provides an alternative to expensive but highly calibrated sensors such as the Skye SKR-1850, with arrays of lower-cost sensors relying on multiple sensor redundancy rather than absolute sensor accuracy. Multispectral sensors have the potential to be deployed relatively inexpensively if these technical issues can be resolved.

In our pilot study the digital camera images were downloaded manually but, as described by Gobbett et al. (2013) in an operational system, the cameras could be solar powered and deliver data across a network that had sufficient bandwidth, particularly if daily image capture rather than capture every 30 min was found to be adequate. Testing the technology around sending image data across the network in this way was not the focus of this pilot deployment, but we illustrate the utility of such an approach by our transmission of the multispectral and soil moisture sensor data via a WSN.

We showed that a single image selected in the middle of the day was sufficient for seasonal monitoring, but that camera images from other times of the day were also useful for investigating unexpected data from the other sensors. The selection of camera images from the middle of the day was made to minimize illumination changes between images and used an automated white balance setting on the camera following that used in, e.g. Macfarlane and Ogden (2012). Other studies have used a manual/fixed white balance in order to minimize changes in illumination (Toomey et al., 2015; Sonnentag et al., 2012) and its use is recommended by the Phenocam network (http://phenocam.sr.unh.edu/webcam/). This aspect could be investigated further in future deployments, as it may enable even stronger correlations to be derived from the digital imagery.

There were benefits to having both multispectral sensors and digital cameras as they complement each other in data interpretation. In an operational setting with cost constraints, a single digital camera could be used to give visual feedback on pasture status to the producer while using a wide deployment of spectral sensors as the main data source. In our study, the separate soil moisture sensors at each node were used to aid in data interpretation. Additional precipitation information could also be provided by the addition of a low-cost rainfall sensor to alleviate the necessity of using rainfall data from non-local meteorological stations.

We have been explicit in this study about not expecting to capture the heterogeneity of tropical pastures with just the two sensors used in the pilot deployment, as assessing the spatial heterogeneity of the pastures was not the project's goal. The two nodes were intentionally placed in an area of the paddock which was as similar as possible at deployment, and the fencing of one node was aimed only at providing a range of pasture heights. An important question about the use of proximal sensors mounted on static nodes is whether the spatial heterogeneity of the pastures is adequately captured by the small area on the ground that the sensors observe, assuming an appropriate number of sensors are deployed. The small FOV of an individual sensor is in contrast to the spatially extensive data obtained from satellite and airborne sensing platforms, and more recently from mobile platforms such as ground vehicles (e.g. King et al., 2010), helicopters, unmanned aerial vehicles (UAV) (e.g. von Bueren et al., 2015), and robotic set-ups to move sensors (Hamilton et al., 2007). In an operational deployment of sensors it may not be necessary to spatially sample the landscape exhaustively, as occurs from an imaging platform such as a satellite; the landscape only needs to be sampled with the number of nodes and their spatial arrangement should be suitable for capturing the spatial pattern in the particular landscape. This includes considerations such as whether the spatial pattern in the pastures is relatively stable, as is more common in temperate pastures, or is more clumped and heterogeneous, as is common in tropical pastures. Spatially heterogeneous pastures can also result from pasture management such as reseeding. The assessment of landscape spatial pattern at multiple scales is a broad topic; a good overview can be found in McCoy (2005) and a more detailed example in Chen et al. (2012).

Options for addressing these spatial sampling concerns of point-based proximal sensors in an operational system include placing multiple sensors strategically in key paddock zones such that the sensors capture the range of paddock variability. Remote sensing images, even if captured only once or twice per year, could be used to aid in the delineation of suitable zones in conjunction with local farmer knowledge. Data from this set-up could then be aggregated up to the scale of a farm management unit to create a robust time series of observations. Alternatively, the sensors could be mounted on a mobile platform that monitors the pastures along a series of waypoints at set times of the day. Unlike the set revisit times of satellite-based remotely sensed images, helicopters, and UAVs have the potential to capture data under a more flexible acquisition schedule. However, data from these non-satellite platforms have more complex processing requirements due to the stability of the imaging platform and the capture of strips of image data in separate flight lines. Increasingly, these processing limitations of mobile platforms are being mitigated by advances in automating image processing (Colomina and Molina, 2014), but they still have the limitation of providing intermittent rather than continuous monitoring. More importantly, while capturing raw data from these systems is relatively easy, creating an operational system to convert the data to something the producer can use for decisions making is complex.

While there are limitations to using point-based sensors for monitoring heterogeneous tropical pastures, this is balanced by the benefits of having a near-real-time continuous data stream for monitoring. For example, an ideal pasture monitoring system would combine data from multiple sources; proximal sensing data for repeated and continuous monitoring of the pastures, and remote sensing images collected at a limited number of times when a spatial assessment of pastures is required. An automatic sensor system could also be set up to trigger a notification to a smart phone or tablet when a critical threshold in feed availability or bare ground has been reached. These data sources could also be combined with other precision farm management technologies, such as walk over weighing (González et al., 2014), and emerging low-power sensor network systems (e.g. http://www.taggle.com.au). For these combined sensor technologies to be used on-farm outside of the current research pilot deployment would require future technical development to streamline their installation and operational use.