A framework for predicting soft-fruit yields and phenology using embedded, networked microsensors, coupled weather models and machine-learning techniques - Royal Botanic Gardens, Kew research repository
Skip to main content
Shared Research Repository
Journal article

A framework for predicting soft-fruit yields and phenology using embedded, networked microsensors, coupled weather models and machine-learning techniques

29 November 2019

Abstract

Improving the accuracy of harvest timing predictions offers an opportunity to sustainably improve soft fruit farming. Fruits are perishable, high-value and seasonal, and prices are typically time-sensitive. Harvesting is labour-intensive and increasingly expensive making accurate phenological predictions valuable for growers. We have developed and tested a novel framework for linking mesoscale weather forecasts to local crop microclimates using embedded autonomous sensors to produce bespoke phenological predictions, using strawberries as the model crop. Seedlings were planted in polytunnels, and environmental and yield data were collected throughout the growing season. Over 1.2 million datapoints were collected by networked microsensors which measured spatial and temporal variability in air temperature, relative humidity (RH), soil moisture and photosynthetically active radiation (PAR) irradiance. Fleeces were added to a subset of the plants to generate additional within-polytunnel variation. Trigonometric models transformed weather station data, which showed a relatively low agreement with polytunnel air temperature (R2 = 0.6) and RH (R2 = 0.5), into more accurate polytunnel-specific predictions for temperature and RH (both R2 = 0.8). Cumulative fruit yields followed logistic growth curves and the coefficients of these curves were dependent on micro-climatic conditions. After 10,000 iterations, machine learning adequately optimised the coefficients of these curves, including RH and air temperature into the fitted equation. Dataloggers measuring environmental data in-situ could infer model parameters using iterative training for novel fruit cultivars growing in different locations without a-priori phenological information. Reliance on manually measured yield data is a current limitation but if high-throughput technologies emerge then this process could be entirely automated. We have demonstrated that this framework can be used to predict fruit timing. Predictions could be refined and updated as frequently as new data becomes available, which in this case would be every eight minutes. This approach represents a step-forward in developing bespoke phenological predictions to inform grower decisions.

Files

This is a metadata only record.