Perspectives on error and uncertainty in farm-scale digital platforms for decision support: A case study using the North Wyke Farm Platform
Year: 2025
Type: conference-paper
Abstract: Ensuring and maintaining food security, in the context of climate change, is a global challenge of the 21st century. It is necessary to maintain or enhance productivity from agriculture, while at the same time balancing environmental and social priorities – not only for the near future but the distant future also. To strike this balance at the farm-scale, timely decisions on agronomic management, system conversion and/or the adoption of a particular farming philosophy are warranted. Agroecosystem process-based models (PBMs) can be used to inform such climate smart decisions. For sustainable farming, PBMs can simulate processes for nutrient cycling and pollutants to water and air, which can be balanced with simulations for crop or livestock performance. As mathematical models, PBMs have relatively small data requirements, and thus are ideally suited to facilitate decisions for most farms. However, the simulated processes of a PBM deployment can be enriched when they are coupled with on-farm in-situ and off-farm remotely sensed data, together with routine data measurements and records. The data streams themselves can be characterised via data-driven models (DDMs). In doing so, this PBM-DDM coupling creates a digital shadow (DS) of a farm’s agroecosystem, rather than a digital model (DM) that involves a PBM only. In turn, if the architecture of the DS provides actionable feedback to the farm by positively influencing a farmer’s decision-making while refining the farm’s data capture programme, then the DS becomes a digital twin (DT). A farm decision support tool (DST - one of DM, DS or DT based) should facilitate practical, feasible and right-time decision making - both short- and long-term. Critically, DST forecasts need to be given with accurate estimates of forecast uncertainty to quantify decision risk. Set within an interactive visual environment, a farm DST should enable the farmer (or advisor) insight and discovery of the farms’ past, current and future status. For a PBM-based farm DS/DT, analytical choices for refining the chosen PBMs, with data streams typically take three major forms: (a) data assimilation on the PBM simulations, say using an ensemble Kalman filter; (b) data assimilation where the PBM, or a component of it, is replaced by its DDM emulator, say through machine learning; and (c) a coupling of PBM simulations with those from a DDM to form a hybrid. Choosing the PBM refining form depends on the situation, the questions being asked of the DS/DT, data availability (quality, quantity, coverage) and the flexibility of the PBM for its adaptation. All refining forms inherently improve the capture or estimation of uncertainty over that found with a PBM alone. Typically, choice (a) suffices for forecasting broad trends and patterns; choice (b) reduces computational overheads found with choice (a) thus ensuring timely decision making; while choice (c) better captures process detail, such as trends in variance for indications of declining system resilience. Any farm DS/DT is limited by the inherent complexities (in space, time, and scale) of its agroecosystems where for most of its component processes, the associated measuring or sensing technologies are either limited or non-existent. This means a fully DDM-driven farm DS/DT is not currently viable (i.e., it is not possible to dispense with PBMs altogether). In summary, a farm DS/DT consists of a series of coupled predictive models (PBMs/DDMs), including ensemble forms, where their performance can be evaluated with historical data (in-sample) to provide objective value to (out-of-sample) forecasts. Out-of-sample performance can be routinely re-evaluated as data is captured via dynamic in-sample assessments. The models of a farm DS/DT need to be both accurate in their predictions and accurate in their estimates of prediction uncertainty, where there are numerous sources of error and uncertainty that need to be captured to facilitate robust decision making. The quantification of uncertainty needs to account for data sources, model (choice) sources and computational sources. A farm DS/DT can be designed to facilitate sustainable, resilient, and/or net zero farming. Each endeavour requires a specific data capture design for specific farm processes together with specific analytical toolkits. Although such DS/DTs would share common components, say for data management, data preprocessing, a DS/DT servicing all three endeavours would unlikely be feasible or even desirable. This paper illustrates uncertainty quantification for farm-scale DSTs (via DMs, DSs, DTs) using the open datasets of four instrumented research farms at Rothamsted Research’s North Wyke Farm Platform (NWFP) long-term experiment in south west UK (https://www.rothamsted.ac.uk/national-capability/north-wyke-farm-platform). The NWFP was established in 2010 to facilitate system-scale research (Orr et al. 2016), where to date over 400 in-situ sensors have been deployed and over 100 million measurements have been captured. The component processes (for soils, crops, livestock, biodiversity, disease, pests, water, and gaseous emissions) of the farms are multi-scale with complex biological, chemical, and physical connections, both above and below ground, that in turn change over space and over time – all of which provide significant challenges for the capture and quantification of uncertainty. However, given an accurate capture of uncertainty, the risk of poor farm management decisions using DM/DS/DTs can be quantified, facilitating planning for climate smart farming.
Source: ARPHA Conference Abstracts
Authors: Paul Harris, Phil Le Grice, Bruce Griffith
Institution: Rothamsted Research
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDGs: Climate action, Zero hunger
Open Access status: diamond
OpenAlex ID: W4410826295