A flexible and practical approach for real-time weed emergence prediction based on Artificial Neural Networks

CHANTRE, GUILLERMO R.; VIGNA, MARIO R.; RENZI, JUAN P.; BLANCO, ANÍBAL M.

BIOSYSTEMS ENGINEERING

ACADEMIC PRESS INC ELSEVIER SCIENCE

Año: 2018 vol. 170 p. 51 - 60

1537-5110

Most popular emergence prediction models require species-specific population-based parameterstomodulate thermal/hydrothermal accumulation. Such parameters are frequentlyunknown and difficult to estimate. Moreover, such models also rely on hardly available anddifficult to estimate soil site-specific microclimate conditions, which in turn depend on soilheterogeneity at a field spatial level. On the other hand, modern agriculture benefits fromeasily available real-time information, in particular on-line meteorological data generatedby forecasts and automatic local weather stations. In this context, Artificial Neural Networks(ANN) provide a flexible option for the development of prediction models, especially to studyspecies which show a highly distributed emergence pattern along the year. In this work, anANN approach based on easily obtainable meteorological data (daily minimum andmaximum temperatures; daily precipitation) is proposed for weed emergence prediction.Relative Daily Emergence (RDE), expressed as a proportion of the total emergence, was theadopted output variable. Field emergence data recorded on a weekly basis were used togenerate RDE patterns through linear interpolation. Results for three study cases from theSemiarid Pampean Region of Argentina (Lolium multiflorum, Avena fatua and Vicia villosa), which show irregular and time-distributed field emergence patterns, are reported. In all cases, ANN model selection was based on the Root Mean Square Error of the test set which showed better consistency than other typical Information Theory performance metrics. The combination of large ANN with a Bayesian Regularization Algorithm generated satisfactory estimations based on the RMSE values for independent Cumulative Emergence data.