1. Introduction
Queensland’s rangelands occupy over 80% of the state and are extensively grazed by sheep and cattle, estimates in excess of 20 million exist [
1]. The value of this industry to the state and national economy is estimated at over
$5 billion per-annum [
2]. Native pastures account for the majority of the feed-base of this industry. The climate of these native pastures is highly variable. The initial antecedent climatic conditions, including the existing soil, water and pasture growth state, is an often-overlooked aspect of systems thinking for managing climate variability [
3]. This simple concept provides the starting point to any future projection of productivity driven by how the future seasonal climate may unfold. Even if no formulated forecast position is taken, and median conditions (climatology) or persistence of the current situation might be assumed, any future projection of pasture growth still requires a current system state. Pasture resource assessment, that is current pasture biomass (growing, senescing and dead), pasture quality and grazing land condition, is the key to setting sustainable livestock numbers. Describing the ambient antecedent condition of Queensland pasture resources is problematic due to the heterogeneous nature of grazing landscapes at micro and macro spatial scale over the year.
Queensland has one of the most variable climates in the world, frequented by episodic droughts, floods, wildfires and tropical cyclones [
4]. This highly variable climate results in a wide range of seasonal pasture production of varying quality. Therefore, the matching of livestock numbers to the available pasture resource requires continual observation and fine-tuning of herbivore density [
1,
5]. Overgrazing of the pasture resource results in loss of grass cover, development of bare patches of ground which can become highly erodible, unfavourable changes in palatable pasture species composition and phenotype and increases in woody vegetation [
6]. The effect of degradation causes a loss in productivity and potentially more overgrazing to compensate, further development of erosion gullies from accelerated runoff and loss of sediment into waterways and in some cases into the Great Barrier Reef lagoon [
7].
The estimation of stocking rates depends on the assessment of the available and future pasture biomass and its quality. There are various methods for making these assessments, including (1) manual collections of pasture biomass along defined transects using quadrat harvests or estimations [
8], (2) collecting well-calibrated rising-plate measurements also along established transects [
9], (3) optical satellite-based observations such as Landsat, Sentinel and Modis [
10], (4) through simulation modelling of pasture production, hybrid satellite-pasture simulation approaches [
11], and more recently, (5) drones or UAVs used to assess pasture growth and pasture quality [
12,
13]. Benefits and limitations exist with all techniques, but UAVs whether alone or in fusion with other techniques promise to be a major enabling technology. The scientific literature over the past few years shows emerging benefits from UAV-captured photogrammetric, multispectral and hyperspectral imaging of pasture condition with examples from the USA, Brazil, Europe, Australia and South Africa [
13,
14,
15,
16]. An optimum solution using data fusion possibly exists, combining the best aspects of all technologies. Support of the simulation can be bolstered by incorporation of satellite-based remote sensing imagery and local field measurements; however, achieving sufficient calibration observations over a large area requires intensive field data. The introduction and evolution of sophisticated drone or UAV platforms are a great promise for bridging this gap. For example, UAVs can observe at centimetre resolution with photogrammetric, multispectral and hyperspectral techniques.
This study aims to build upon recent advancements in UAV technologies, to develop and test a method to map and monitor pasture resources across the large and heterogeneous rangelands of Queensland, Australia. It aims to be a method that can be applied at a range of spatial and temporal scales as an important component of both long-term broad-scale pasture monitoring programs, as well as a stand alone or complimentary onground decision support tool. It aims to do this by way of developing similar measures of pasture yield and nutrient composition to that of traditional plot based field sampling techniques in a more efficient and spatially explicit manner. This will improve integration into satellite-derived products to scale beyond the extents of a field site as a key component to broad scale monitoring and reporting. In addition, onground decision makers will also be able to precisely assess the state of the pasture resource at a finer submetre spatial resolution, building further confidence in remotely sensed data. Improved confidence and adoption of satellite derived pasture resource estimations through access and use of finer scale UAV imagery and associated pasture state estimations, aims to provide the necessary tools for setting more sustainable long-term livestock numbers by graziers.
Pasture biomass estimation from a UAV platform is a relatively new technique. To date the majority of published research has been undertaken in other countries in either improved or cultivated pastures. Natural pastures in Australia remain largely under-studied. UAV photogrammetry Structure from Motion (SfM) is one of two methods used to determine vegetation height, the other being Light Detection and Ranging (LiDAR). LiDAR uses laser technology in an active manner to measure the distances from the device to the object, whereas photogrammetry SfM is a method of measuring height and mass from many overlapping photographs [
17]. Pasture biomass has been mostly determined from the latter technique as it is a considerably less expensive technology and pastures are more likely to be in environments that are not obstructed by other vegetation such as tall or dense tree canopies. Linear modelling of the relationship between pasture height and biomass is a well established method [
18,
19]. Modelling of this relationship with UAV SfM is becoming more popular and several studies have reported strong correlation between UAV derived imagery and ground-based estimates. [
13,
14] reported average
values as high as 0.78 and 0.81 in both the native pastures of Arizona, USA and a cultivated pasture in northern Germany. This case study over a number of dominant grasslands and woodlands of the rangelands of Queensland, attempted to develop a similar technique that, as previously stated, could be both used to potentially cross calibrate satellite imagery for broader applications across the rangelands of Queensland and as a standalone measure of pasture biomass at the field, paddock and property scales for improved pasture management.
Sustainable livestock grazing, through improved matching of stocking rates to pasture resources, depends not only on the amount of pasture biomass but importantly both its species composition and nutrient status. Adequate crude protein in the diet of livestock is essential for their maintenance, growth, lactation and reproduction [
20]. Pasture digestibility becomes limiting to livestock growth and reproduction when it reaches certain levels. Acid detergent fibre (ADF) is an inverse measure of pasture digestibility. The importance of crude protein in livestock diets has been studied for many decades [
21]. The use of remote sensing, field and laboratory spectroscopy is also well established in the literature [
22,
23,
24]. Pasture nutrient status is a field of study that has recently re-emerged since the development and affordability of UAV technologies. In comparison to past airborne/satellite and field based systems, UAV hyperspectral remote sensing provides a more cost-effective and spatially explicit means of measuring known plant chemical components. Such components are useful as indicators of pasture nutrition and digestibility for livestock grazing. Crude protein and ADF are two commonly used indicators of pasture digestion. Recent UAV studies have demonstrated the use of hyperspectral sensing in detecting both crude protein and ADF in the natural pastures of other countries [
15]. UAV hyperspectral sensors are not low-cost; however, for nutrient estimation they are essential. The aim of this study will be to demonstrate the use of UAV hyperspectral remote sensing to detect both crude protein and ADF in a range of native pastures across the rangelands of Queensland.
Study Area
The study area illustrated in
Figure 1a encompasses most of Queensland’s rangelands, an area of approximately one million square kilometres. It is a mix of both dry tropical to subtropical rangelands in the north and along the coast and arid to semiarid rangelands throughout the interior. Annual average rainfall ranges from 1800 mm on the north east coast down to 1400 mm on the south east coast and 1000 mm in the north west interior to below 200 mm in the south west interior [
25]. Field sites were established in a mix of woodland and grassland communities to sample a range of pasture biomass, heights and nutrient compositions. Sites in each community were further stratified to characterise both rainfall gradients and pasture compositions.
Table 1 details the number of sites in each community and category along with the dates of each field sampling campaign.
Figure 1a illustrates the location and spatial distribution of the field sites across the study area.
Figure 1b also presents examples of several field sites seen through high-resolution UAV imagery, illustrating the mixtures of vegetation composition, structure and pasture biomass.
4. Discussion
Survey accurate ground control was important to this study for the matching of field observations with UAV imagery, processing of the SfM photogrammetry and in the orthorectification of the hyperspectral scan imagery. Survey accuracy at each field site varied depending on the base station solution. As described in the methods section the Propeller AeroPoint
TM network solution was used where it was available, and a portable base station solution, corrected through the AUSCORS network [
29], was utilised in more remote locations. Submetre accuracy of both solutions was achieved but likely varied between one another; however, in the absence of an extensive base station network across the vast rangelands of Australia this was still considered an acceptable level of accuracy.
Field sampling of pasture height, yield and quality is not without its own sampling error. Previous studies have used other methods such as terrestrial laser scanning and hand-held spectroscopy [
22,
41], although these methods were initially trialled in this study, a simple and direct means of collecting field measurements was chosen. Limitations of each method exist, in this study these included the ability of the rising plate method to accurately measure pasture height across rough and undulating ground surfaces, often resulting in erroneous pasture heights due to the rising plate base not accurately contacting with the true ground surface. Pasture yield harvesting of TSDM also has limitations, including the ability to physically harvest low sample sizes such as in arid and sparse grasslands. Other limitations in sampling efforts included the amount of samples collected that would adequately represent the variability of each metric in both their statistical and spatial distributions. Limitations in time and resources meant a smaller overall field sample set was collected than would likely be needed to definitively model each metric across the diverse range of landscapes of the study area. However, this research was able to demonstrate the applicability of the various field sampling techniques and methods to both compare and calibrate the UAV-borne imagery and associated metrics.
The UAV pasture height model (
) was compared with the rising plate field measurements (
) (see
Figure 4). Results of the comparisons across all four pasture subcategories illustrate the potential of this method. Both grassland categories showed on average a strong correlation in the lower pasture height range that weakened as height increased. This was also present in the woodland categories, however not as distinctively. Pasture heights of most grasslands of the study area were relatively low, on average less than 70 cm tall. Taller pasture may become problematic using the rising plate field measurement method. A number of limitations of the
method were discovered. The presence of tree and shrub-cover limits the ability to accurately determine the pasture sward height both in the physical obstruction or occlusion of the pasture sward by tree and/or shrub canopy cover as well as the effects of shadowing from the canopy layer and its effects on the calculation of the elevation and pasture height surfaces. An example of such a field site is illustrated in
Figure 5 and the degree of these effects upon the accurate determination of both
and the resultant modelled TSDM t ha
−1 estimates is seen in both the robust and automated ML modelling outputs. In the case of the automated ML output, spurious estimates of
have resulted in an overfitting of the model. The robust regression modelling approach, however, is able to exclude some of these potential outliers and provide a better fit. Further to this limitation, the chosen window size used in determining the ground surface elevation has a significant impact on the overall accuracy of the
surface. An iterative process was chosen whereby the strength of the correlation of
and
was used to determine the most suitable window size. Although this showed potential, its practical application may be limited, particularly in the problematic landscapes for accurate
measurement. Automated ground layer determination is problematic in high resolution photogrammetry based digital elevation or terrain modelling, particularly in finer scale (<5 cm) imagery. Numerous tree and forestry based point cloud filtering methods exist, including [
42,
43]; however, application at a finer scale becomes computationally resource intensive. This becomes further problematic in dense pastures of sloping terrain, whereby a true measure of the ground surface elevation is difficult to determine for each pixel. Several recent studies such as [
13,
31] have physically selected and/or surveyed ground surface elevations used in the calculation of a digital terrain model to then determine the pasture or canopy height surface. This approach was considered impractical in this study and problematic in the often sloping terrains of the study area. Although a number of limitations exist in the field and UAV measurement of pasture height exist, a least-squares linear regression of both the field and UAV derived measures in this study indicated the potential of the UAV based method to replace the manual method, with enough ongoing validation across a range of pasture heights, swards and species.
The ability to develop accurate UAV based models of pasture yield have been demonstrated in this study. The estimation of pasture yield or biomass was explored with both a prior knowledge robust regression modelling approach and an automatic machine learning method. Both approaches proved successful with reasonable levels of accuracy. The robust regression approach performed slightly better, likely due to its ability to deal with some of the before mentioned erroneous field measurements through its classification and weighting of statistical outliers. Although this approach could potentially reduce valid outliers, it was deemed a useful technique. The stratification of the field measurements of pasture yield and the development of separate models for both woodland and grassland pasture groups improved each model’s performance. The grassland model performed better than the woodland, likely the result of the before mentioned limitations of modelling pasture height in mixed tree and grassland pastures. The robust regression again performed better than the auto ML model in both the grassland and woodland separate models. This was particularly evident in the woodland group whereby overfitting of the model occurred as discussed previously. The accuracy assessments of each modelling method and its associated stratification shows a similar trend in that the variability of the accuracy of each model increases for the woodland group and to some degree the auto ML method. Improvements to the auto ML modelling method’s output accuracy could include optimisation and potentially weighting models in the selection procedure to include those with advanced procedures in removing or weighting outliers. The calibration of further predictive modelling of pasture yield will likely need further field measurements. The models developed in this study provide a basis for further modelling.
High correlation between specific bands of the UAV hyperspectral imaging with that of the proportions of acid detergent fibre and crude protein, in predominantly the Mitchell grassland pastures of the study area, has been demonstrated in this study. In a study with similar pastures, [
22] reported crude protein absorption features of a moist grassland in South Africa, to be in the wavelength range between 720 to 745 nm of the red edge portion of the spectrum. This study reported a similar distinctive peak in correlation of the single-band spectral feature at 759.85 nm (see
Figure 8). However, this study reported a stronger correlation with both the simple-ratio and normalised index at the 939.92 and 947.92 nm portions of the spectrum, also referred to as the third overtone absorption mechanism [
44], whereby plant tissue oil and protein compounds are absorbed; however, it is more commonly associated with the short wave near infrared portion of the spectrum. Regression of acid detergent fibre was also different to [
22] but similar to [
15] in that the best correlation was in the visible portion of the spectrum for the single-band spectral feature, specifically visible red at 651.81 nm. The visible red (651.81 nm) and protein absorption/near-infrared portions (919.91 and 939.92 nm) illustrated the highest correlation for the simple ratio and normalised index. These results, although slightly dissimilar to other studies with similar techniques, illustrate the potential for further research utilising these techniques, into the pastures of the study area. In particular, these results show the sensitivities of this type of visible near infra-red hyperspectral imaging to accurately estimate the proportions of crude protein and acid detergent fibre in the pastures of the study area in differing growth stages and/or states of pasture decay or senescence, whereby the short-wave near-infrared portion of the spectrum has been more commonly used.
A number of difficulties in accurately mapping both measures were encountered including the successful orthorectification of UAV hyperspectral imagery. Although an area that covered the majority of each field site in a single scan was used in the analysis, subsequent adjacent scans were not, due to difficulties in accurately aligning each scan. Onboard post processing differential GPS position data that can be matched to the hyperspectral sensors IMU positioning is a potential solution to improve the geometric accuracy of each flight scan. Further sensor spectral calibration is the other consideration for ongoing mapping efforts. The UAV hyperspectral sensor used in this study was factory calibrated and will likely need recalibration to ensure spectral measurements are of the highest grade. Only 10 field samples were used to develop a robust regression model of both pasture quality measures; these samples were, however, across a vast geographic range of similar mostly native rangeland pastures and provides a basis for further research. Demonstration of the modelling across an area of Mitchell grassland captured with the same hyperspectral sensor used in this study, provides a basis for further work. This includes the development and testing of further modelling to predict changes in both crude protein and ADF under differing pasture management regimes and pasture growth stages. Ongoing research and development from this initial case study will provide innovative and improved predictions of pasture quality to further develop, inform and improve sustainable grazing decision-making systems.
5. Conclusions and Further Work
The aim of this study was to develop an accurate and spatially explicit measure of pasture biomass and nutrient composition that can be used for both scaling up to satellite-borne imagery for long term broad scale pasture resource monitoring as well as a fine scale measure for improved onground pasture resource management. A number of challenges still exist as already discussed; previous studies using similar techniques have mostly been undertaken in cropped pastures. This study has demonstrated that these techniques can be applied with rigour across the rangelands of Queensland, providing further opportunities to both integrate UAV imaging into satellite-derived products and importantly as standalone measures of real-time pasture state at the paddock and potentially property scales. Improved spatial and temporal measures of pasture state, utilising the methods developed in this study, provide new opportunities for improved pasture management. The existing GRASP pasture growth simulation modelling, developed by [
45] and implemented across Queensland’s rangelands by [
46], has traditionally utilised point based measurements of climate and pasture state. Integration of both the field scale measures developed in this study and future satellite integrated measures, could provide a more efficient and spatially explicit system, providing producers with further tools to adopt more resilient and longer-term grazing strategies in an increasingly variable climate.
The limitations found in this study lead to a number of key recommendations for further improvements including; (1) further assessment of potential spatial inaccuracies of a differential GPS solution utilising a portable base station versus the Propeller AeroPointTM correction network and the potential impacts of any differences upon down stream photogrammetry processing, (2) improvements to the orthorectification of hyperspectral image products through the incorporation of onboard UAV post processing differential GPS corrections, (3) testing the suitability and practicality of more sophisticated methods of collecting coincidental field measurements for ongoing UAV pasture model calibration and validation including terrestrial laser scanning and both AVNIR and SWIR field spectroscopy, (4) development and incorporation of a tree and shrub canopy and shadow image segmentation, (5) exploration of further automated methods to efficiently and accurately determine ground elevations in fine-scale photogrammetric digital surface modelling, (6) investigation of improvements to pasture yield modelling through further testing and optimising of automated machine learning techniques and importantly further field sampling in areas of poor model performance, including tall and sparse pastures, and lastly (7) further field sampling of pasture quality measurements to enable further sophisticated pasture quality modelling with a larger data set and in a broader range of pasture types in varying stages of growth.
In addition to specific improvements to the methods used in this study, to more accurately model and map both pasture quantity and quality at the field, paddock and potentially property scales, further development of methods to scale each UAV based measure up to satellite derived imagery is needed. This could begin with high to medium spatial resolution platforms such as the Dove Planet and the Copernicus Sentinel constellations of optical and radar satellites. Scaling from centimetre UAV imagery up to one to 10 metre resolution imagery would provide the first steps in understanding the challenges in integrating both platforms. A number of challenges will likely exist, such as differences in radiometric and geometric calibrations between platforms and the effects these differences may have upon the scaling of measurements from one platform to another. However, such efforts will provide further opportunities to apply accurate and efficient field scale measurements to satellite borne imagery for long term/broader scale monitoring. Similar earth observation studies, including [
47,
48], have begun to explore UAV to satellite imagery integration and along with this research provide a basis for further development in the native pastures of the Queensland rangelands.