SBP is working with PGRO on a Knowledge Transfer Project (KTP) to help develop a model for determining accurte estimates of harvest dates for pea crops across the country. The aim is to provide an automated tool for vining pea growers.
Ultimately, it is expected that using scientific modelling to provide accurate forecasts to maximise pea quality and tenderness ratings will improve efficiency within SBP and Eyemouth Freezers, enabling the factory to work at higher capacity and reducing the need to bypass land.
Alongside other pea groups in the UK, SBP’s involvement will be in the provision of data to help to establish accurate modelling. We will do this by gathering pre-harvest crop and field data throughout the season and a student will be dedicated to this task from June to September.
The project has been designed and developed by Leah Howells who is KTP Associate with the University of Nottingham. She has produced a report which appears in the British Peas & Beans Winter 2022 Newsletter and it is included below:
PGRO are currently undertaking a three-year KTP project in collaboration with the University of Nottingham and Innovate UK. Its objective is the development of two prediction models capable of advanced estimations of vining pea maturity and yield. Now in its latter stages, the research aims to ease harvest planning capabilities for more efficient vining pea processing across the UK using remote sensing and machine learning.
A particular benefit will be knowing with greater certainty the order in which fields should be harvested, with maturity estimations giving a narrower window for tenderometer or maturometer testing in the days leading up to harvest. This will ultimately minimise the need for time-consuming sampling. Advanced predictions of yield will primarily help factories plan for above- or below-average yields, ensuring a constant supply of peas can be handled at factories, allowing them to work consistently at capacity and minimising the bypassing of fields due to unexpectedly high yields.
Remote sensing: a hands-off approach
As a notoriously unpredictable crop, vining pea yield and rate of maturation are highly affected by environmental conditions, with temperature, rainfall, daylength and soil type among the many factors which must be accounted for. Add to this the many different available vining pea cultivars, each with varying yield potential, maturity indices and canopy architectures, and you have a truly complex task.
This project aims to specifically utilise remote sensing to acquire the necessary data to make predictions, without the need for in-person sampling. Remotely sensed data can be automatically collected at regular intervals throughout the vining season for a comprehensive picture of the development of each crop.
Growers’ own on-farm weather stations of course provide the most accurate and location specific climatic measurements. However, Met data has been found to be sufficient for the areas when grower data is not available, and is more than capable of picking up anomalous weather events like heavy rainfall or prolonged periods of high temperature. Aggregated and interpolated Met data is readily available from a number of different online services which will be utilised in this project.
There is also a wealth of satellite data available, with the clear benefit that images require no field visits or manpower to acquire. The Sentinel-2 satellite for example has a revisit time of around 3 days in the UK, with a 10-metre resolution sufficient for field-scale analysis and predictions. Whilst true-colour images like those captured by a camera can give some indication abut crop health and growth stage, multispectral canopy reflectance indices are more insightful. Vegetation indices like NDRE and NDVI can be used to measure the ‘greenness’ of a crop and give valuable information into its growth stage and yield potential. The ability to ‘see’ a crop and make visual assessments is the most effective way of assessing its development, and using satellite data will allow a more detailed look, albeit from space.
In model training and testing, estimated harvest dates were generated with a mean absolute error of 0.98 days, whilst average error of yield predictions was 0.77 t/ha. These results are promising, and with estimations able to be generated over a week in advance of harvest, indicate the models have many potential benefits. The models were also tested in real time across Humberside and Yorkshire with Birds Eye during the 2021 vining season, and showed a good ability to predict the higher-than-average yields that were seen across the UK last year.
Long term trends of course cannot be ignored. Against a backdrop of rising global temperatures and increasingly frequent extreme and adverse weather events, it is essential that the prediction models are flexible enough to account for gradual changes and adjust estimations as a result. The aim is for the models to be dynamic; incorporating new data year by year, ultimately increasing their accuracy as time goes on.
The resultant models will eventually be available through the PGRO website, giving grower groups equal access and equal benefit to greater planning power.
Leah Howells – KTP Associate with the University of Nottingham
In addition, Leah has written an article which appears on p.12 of The Vegetable Magazine (Winter 2021/22).