• Login
• Admin

Howells, Leah (2025) Remote sensing and machine learning for the field-scale prediction of maturity and yield in vining pea (Pisum sativum L.). PhD thesis, University of Nottingham.

PDF (Final submission- corrections made) (Thesis - as examined) - Repository staff only until 12 December 2027. Subsequently available to Anyone - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader Available under Licence Creative Commons Attribution. Download (56MB)

Abstract

Vining pea (Pisum sativum L.) is an important legume species that is cultivated in the UK as a staple vegetable for freezing and canning. Vining peas are harvested prior to physiological maturity, where quality attributes are the main determiner of optimal harvest date. A short harvest window means production is carefully managed. Since the mid-20th century, the primary method of maturity forecasting in vining peas has been the use of accumulated heat units (AHU), which are used to space sowings and estimate the order of harvests. Labour-intensive field sampling is employed in the days or weeks approaching harvest to ensure crops are harvested at the optimal time. Peas are highly sensitive to environmental conditions, and crops are prone to incurring waste as a result of bypassing due to accelerated maturation, or when higher than expected yields cannot be processed due to factory capacity constraints. The threats to global agriculture from climate change are well understood, but many crops, including vining peas, are still forecasted using techniques developed under past climatic conditions.

To understand the climatic changes that vining pea agriculture has experienced in the past decades and determine whether the AHU remains a reliable forecasting method, a long term analysis of historic Processors and Growers Research Organisation (PGRO) variety trials data was conducted. Evidence was found for a significant increase in maximum (+1.46 °C) and mean (+0.74 °C) temperatures at trial sites between 1985 and 2022. There was additional evidence for declining growth period lengths of standard early vining pea varieties as a result, with total growing days decreasing by over 3 weeks over the same period. Concurrently, the number of AHU building up between sowing and harvest of these standard varieties was found to have decreased, due to the decrease in the number of growing days, and opportunity for heat units to accumulate. The applicability of using 11.6 AHU as a standard daily estimate, based on an average daily temperature at harvest of 16 °C and a base temperature of 4.4 °C, was tested. Analysis of tenderometer reading (TR) curves of four standard varieties between 2015 and 2022 showed that using 11.6 AHU as a proxy for a day is still a valid assumption, but that basing estimates of harvest date on a fixed target AHU value at harvest was the main driver of a decline in predictive ability.

In a study to improve temperature-based harvest date forecasts, six different supervised machine learning models were tested for predictive ability, and a Cubist model was selected as the best performing algorithm. Using data from over 14000 crops, including agronomic data and meteorological variables from 2001-2022, a Cubist model was developed for estimation of harvest dates. An average mean absolute error (MAE) of 1.10 days was achieved, rising to 2.01 days across four recent individual years 2019-2022. Using a daily time step and forecasted weather data, the model was tested to simulate daily forecasting within-season, and predictions generated 5-15 days before harvest were not significantly different to those produced 2 days before harvest, which represents the current limit of accurate forecasts.

There is currently no standardised method for yield prediction beyond pod, node and plant counts. Six Sentinel-2 multispectral canopy reflectance-derived vegetation indices (VIs) were analysed for association with final yield at different dates. Correlation with yield was highest on the day prior to the date of full-flowering. Using area under the curve (AUC) analysis, it was determined that the window from 23 days prior to full-flowering to 11 days after was optimal for Sentinel-2 data acquisition to correlate most strongly with yield, whilst maximising data availability. Using agronomic and meteorological data for over 4000 crops between 2016 and 2022, and associated remotely-sensed Sentinel-2 data, a Cubist model was trained to predict yield with an average MAE of 0.61 t/ha. When validated on individual years 2021 and 2022, MAE was 1.05 t/ha and 0.81 t/ha, respectively.

The forecasting ability of the yield model was further tested in combination with the harvest date model, providing predicted harvest dates in place of observations. Results indicated only a slight drop in predictive performance, with an overall MAE of 0.63 t/ha, and MAE of 1.06 t/ha and 0.98 t/ha when tested in 2021 and 2022, respectively.

The ability to forecast harvest dates and yields of vining peas, and thereby implement a comprehensive within-season forecasting system has significant positive implications for reducing wastage and improving efficiency in the vining pea harvest process. The work in this thesis represents a novel innovation in vining pea forecasting, and has resulted in a first-in-the-world, commercially available online application available to processers and growers, which has the potential to significantly improve industry operations.

Item Type:	Thesis (University of Nottingham only) (PhD)
Supervisors:	Wilson, Paul Sparkes, Debbie Crout, Neil
Keywords:	Vining pea; Maturity forecasting; Yield prediction; Harvest date
Subjects:	S Agriculture > SB Plant culture
Faculties/Schools:	UK Campuses > Faculty of Science > School of Biosciences
Item ID:	82768
Depositing User:	Howells, Leah
Date Deposited:	12 Dec 2025 04:40
Last Modified:	12 Dec 2025 04:40
URI:	https://eprints.nottingham.ac.uk/id/eprint/82768

Actions (Archive Staff Only)

Edit View