Operationalising satellite and drone imagery to improve decision-making: a case study with regenerative grazing

Drought is one of the greatest existential threats facing contemporary food security. The emergence of cloud computing, big data analytics, and machine learning promise more timely land management, potentially affording landholders with hitherto unseen tactical decision-making opportunities prior to, during, and after drought. To investigate such promise, we invoked machine learning together with imagery from drones and satellites, using “regenerative agriculture” and adaptive grazing management by way of case study. High-intensity sheep grazing treatments were conducted in small fields (<1 hectare) at 6,000-8,800 dry sheep equivalents per hectare (DSE/ha) for short durations (10-12 hours) near Triabunna in Tasmania, Australia, following guidelines from the Regenerative Agriculture Network of Tasmania (RANT). Paddocks were spelled post-grazing to facilitate pasture recovery (3, 6, 9, 12, 15 months); regenerative treatments were compared with controls comprising lighter stocking rates for longer continuous periods in larger paddocks (10-54 ha at 2,000 DSE/ha for one week). Botanical compositions comprised native grasses including wallaby grass (Austrodanthonia spp.), kangaroo grass (Themeda triandra); as well as introduced species including tall fescue (Festuca arundinacea Schreb.), phalaris (Phalaris aquatica) and cocksfoot (Dactylis glomerata). Biomass was sampled destructively and longitudinally to elicit total standing dry matter (TSDM), standing green biomass, standing dry biomass and trampled biomass.

 

Using ultra-high definition drone imagery (5 mm²) in concert with Sentinel-2 satellite imagery (10 m²), we invoked machine learning to quantify temporal TSDM, standing green and dry biomass. All datasets were georeferenced using ground control points at paddock corners. Drone imagery was used to produce digital elevation models and orthophotos, with the change in pasture height pre- and post-grazing providing relationships between pasture height delta and biomass. Across plots, botanical compositions and treatments, results suggested that for each 0.1 m increment in pasture height, pasture biomass increased by 1,927 kg DM/ha.

 

Recalibration of existing algorithms based on remote sensing (Cibo Labs) using measured pasture data improved model performance relative to the default Cibo Labs model, although simulated values were generally lower than observed values, particularly when observed biomass was high (>5,000 kg DM/ha). Very high biomass quanta (>12,000 kg DM/ha) were not well simulated, although such biomass levels are atypical for the region. Despite this, pooled measurement uncertainty in the field data was similar to the root mean square error associated with biomass modelled from satellite imagery (~900 kg DM/ha). In general, prognostics developed using satellite imagery and machine learning adequately approximated observed spatiotemporal variability within and across treatments, particularly for standing green biomass.

 

Faced with cool wet La Niña conditions, high intensity short duration grazing did not significantly impact overall sward productivity. However, regenerative treatments increased material trampled, evidenced by increased litter, enhanced surface organic matter and decomposition rates thereof. This increased soil water infiltration, nutrient cycling and soil surface stability. Pasture digestibility and sward uniformity were greatest for treatments with the least spelling (three months), whereas standing senescent and trampled material were greater for treatments with 15-month spelling periods. Pre-grazing green biomass was lower in treatments rested for more than three months, with 1,135 and 4,560 kg green DM/ha available at the initiation of grazing. In contrast, treatments spelled for only three months had as much as 14,535 kg green DM/ha. Grazing utilisation was low for most grazing events, ranging between 0% and 50%, typically below 30%, and averaging 15%. Such low utilisation levels are likely to be a function of the time required for animals to adapt to the new grazing regime, and the animal genotype used in this experiment (Merinos). Following the fourth grazing event for example, trampled material ranged from 4,418 to 12,903 kg DM/ha.

 

All treatments recovered significant biomass during spelling, with the 15-month rest treatment having the most standing biomass. The proportion of standing green material in other treatments was low, increasing with grazing intervention and with proximity to the peak of spring, and reducing with rest duration. Much of the green material comprised stem, demonstrating that while standing biomass was high for grazing treatments spelled for more than three months, forage quality was relatively low. 

 

Regenerative grazing trampled significant amounts of biomass resulting in thick, dense litter layers that decomposed rapidly when conditions allowed, but in some cases persisted. The productivity of the three-month spelling treatment was likely due to grazing maintaining the dominant tall fescue in a vegetative state, together with La Niña conditions and regular import of nutrient in excreta afforded by frequent grazing. In contrast, treatments spelled for 12 months accumulated 11,096 and 11,421 kg DM/ha of total standing feed, but of this, only 2,354 and 1,135 kg DM/ha of green material was available for grazing at the 12 month point. For longer rest plots, post grazing trampled litter comprised mostly dead stems which, in some cases, persisted for the duration of the trial, slowing subsequent rates of organic matter decomposition. We contend that the three- and six-month spelling treatments were conducive to improvement in the biological functioning of soils, and enhanced conversion of macro-organic matter into organic molecules and micro-organisms. Rich organic odours of regenerative soils suggested presence of Streptomyces spp., which, if confirmed, would evidence fungal decomposition of macro-organic matter.

 

Across treatments, soil organic carbon (SOC) was greater in surface layers (0-10 cm) and lower deeper layers (30-60 cm), with SOC concentrations in surface layers ranging from 3.7% to 6.6%. SOC in deep layers was low (typically less than 2%) and varied little, implying low propensity for management interventions - such as deep rooted pastures - to impact on SOC in a substantive way. We did not observe any association between pasture productivity and SOC for introduced pasture species, suggesting that other factors may be more closely linked with pasture productivity (e.g. soil fertility, grazing management and prevailing weather). Native pasture ecotypes tended to have greater SOC in surface layers compared with introduced pastures, suggesting propensity for higher SOC in uncultivated soils, although native pastures were generally of lower productivity and nutritive value compared with introduced pasture species. We recommend subsequent long-term SOC sampling to conclusively determine impacts of regenerative grazing on SOC; a minimum period of at least five years is advised to ensure that interacting effects of the vicissitudes of weather and grazing management on SOC are sufficiently captured.

 

Numerous extension events were conducted to engender research adoption and impact. These included radio interviews, co-development of newspaper and online articles, workshops on farm with RANT, farmers, First Nations members, University of Tasmania students and the local community, engagement with the TasAg Innovation Hub and NRM South, social media, and research demonstration to the Tasmanian Minister for Primary Industries and Water, Hon. Jo Palmer MLC.

While we were not afforded the opportunity to examine implications of regenerative grazing under drought, we can conclude that high intensity regenerative grazing with short recovery periods (in this environment, three months) were more conducive to increasing surface organic matter and litter under high rainfall conditions. We speculate that high intensity grazing with longer spelling periods (relative to lighter grazing for longer durations) is likely to beneficially increment soil organic matter through increased litterfall, trampling, enhanced organic matter cycling, and perhaps also soil organic carbon, though more evidence is required to support the latter notion.

 

We highlighted a clear need for high resolution drone imagery in terms of integration with imagery derived from satellites. It is plausible that relationships between drone and satellite imagery taken on one paddock could be extrapolated to other paddocks for which only satellite imagery was available, suggesting that drone imagery would not need to be taken across whole farms. We contend that there is and will be a role for drones and satellite imagery in operational decision making and in improving the timeliness of management. From a research perspective, we argue that much remains to be done in this endeavour, with advances in the state of the art likely to benefit practitioners in their prioritisation and ability to balance environmental stewardship and commodity-based production objectives.

 

This experiment challenges hitherto principles underpinning guidelines for grazing rest duration and frequency, litter production, legume content optima, and nutrient and energy flow from pasture to livestock production. We underscore a place for adaptive grazing management with reasonable spelling periods, although it remains to be seen how such management influences landscape functioning and livestock production under drought. Operational constraints associated with regenerative grazing - such as the need to establish numerous small paddocks and watering points (or mobile watering points implemente