Speaker
Description
- Introduction
Crop growth models can provide real-time forecasts of upcoming drought or nutrient deficiencies and can thus in principle be used to support decisions about irrigation and fertiliser application. To be useful for supporting in-season crop management, forecasts need to be (1) frequently updated and (2) sufficiently accurate. This second point is problematic in practice, because crop models tend to deviate from reality, due to insufficient calibration, and because not all relevant processes are included. A digital twin (van Evert et al. 2021, Knibbe et al. 2022) combines crop growth modelling with updating model state variables based on in-season observations. A digital twin provides daily updated estimates of growth forecasts and also shows how these forecasts are corrected with observations. Ideally, such a digital twin results in greater forecasting accuracy, for the end-user more confidence in model predictions and ultimately more efficient resource management. We developed a fully automated operational digital twin for a strip cropping experiment in the Netherlands.
- Materials and methods
The digital twin strip cropping (DTS) has the following properties:
1. Field experiment in centre of Netherlands, with in total 7 crops and 11 strips per crop. Not all crops were modelled
2. Strip specific management data retrieved through Farmmaps platform (Been et al 2023)
3. Location specific soil and weather data retrieved through Farmmaps
4. Bi-weekly drone flights. Stitching and geo-referencing by a private company then imported into relational database as soon as available
5. Two crop growth models used: Tipstar for simulating potato (early and late cultivar), and WOFOST for fababean, onion and winter wheat
6. Ensemble Kalman Filter (EnKF) adjusts model states based on LAI observations derived from drone images, taking into account both the uncertainty in the model (estimated by perturbating model parameters) and the uncertainty in the observation.
7. NMODCOM simulation framework for integrating different models and enabling EnKF
8. Forecasting by stitching weather till present date, a 14 day forecast, and thereafter from a year in the past (e.g. 1991, 1992, etc). Simulations are repeated 30x with 30 past years to get a plume visualising uncertainty due to future weather
9. Automatic running the digital twin overnight for all 5 crops x 11 strips x 30 weather combinations
10. R-scripts generated daily figures. A new website was developed to present the most recent forecasts every day.
- Results
Figure 1 presents an illustration of forecasts presented to the farm manager. More daily updated figures are available from https://farmofthefuture.nl/data-precisietechnologie/gewasgroeimodellen/ (in Dutch).
Figure 1. Mid July 2024 forecast of Leaf Area Index (LAI) and tuber yield. Figure showing an ensemble of simulations (grey dots) and median of simulations (black line). Green dots show LAI estimates based on drone images. Four green dots in April-May were pre-emergence and ignored. Green dots mid-June and July cause a downwards correction of simulated LAI. Simulations continue with the downward adjusted LAI.
- Discussion
The work presented represents one of the first operational digital twins in the agricultural domain. Farm managers showed greater appreciation for potential of models to support their decision making because we provided daily updated forecasts and visualized how these were updated with data from drone images (e.g. Fig 1).
Scientifically we are still in a pioneering stage. A major advantage of using a filter to update the state variables that it allows to correct the predictions for processes that are not currently in the model, e.g. canopy diseases. We are currently working on an advanced version of the digital twin where in addition to leaf area index also soil moisture is updated using satellite images.
- References
Been, T.H., et al., 2023. Akkerweb and farmmaps: Development of Open Service Platforms for Precision Agriculture. In: D. Cammarano et al. (Eds.), Precision Agriculture: Modelling. Springer International Publishing, Cham. p. 269-293.
Knibbe, et al., 2022. Digital twins in the green life sciences. NJAS: Impact in Agricultural and Life Sciences 94, 249-279.
van Evert, F.K., et al., 2021. A digital twin for arable and dairy farming. In: J. Stafford (Ed.). Precision Agriculture '21. Wageningen Academic Publishers, Wageningen. p. 919-925.
| Keywords | digital twin ; potato; yield forecast; nitrogen; irrigation |
|---|
