Articles
Neural network and ensemble approaches for predicting irrigation demand in open field vegetable crops
Article number
1468_36
Pages
277 – 280
Language
English
Abstract
Efficient irrigation management is crucial for vegetable production under increasing climate variability.
This study evaluates artificial neural networks (ANNs) for predicting available water capacity (AWC) as a basis for irrigation scheduling.
Open field trials with spinach (Spinacia oleracea) were conducted from 2020 to 2023 at two German sites, combining different irrigation and nitrogen regimes.
Soil texture, meteorological data, and plant-based vegetation indices were used as inputs.
Two ANN types were developed: multilayer perceptrons (MLPs) and long short-term memory (LSTM) networks.
Additionally, transfer learning was applied using freely available AWC data from the German Weather Service.
An ensemble model combined all ANN outputs.
Fine-tuned models achieved the best prediction accuracy (R2>0.98, RMSE < 1.5%), with LSTMs slightly outperforming MLPs due to their ability to capture temporal dependencies.
The ensemble approach minimized cumulative errors and improved robustness across soil types and climatic conditions.
Explainable AI methods identified meteorological and crop-related inputs as key predictors, enhancing model transparency.
This study demonstrates that ANN ensembles are a powerful tool for irrigation scheduling in vegetable production as seen in the prototype decision support system ANNI (ANN for irrigation) developed in Geisenheim.
This study evaluates artificial neural networks (ANNs) for predicting available water capacity (AWC) as a basis for irrigation scheduling.
Open field trials with spinach (Spinacia oleracea) were conducted from 2020 to 2023 at two German sites, combining different irrigation and nitrogen regimes.
Soil texture, meteorological data, and plant-based vegetation indices were used as inputs.
Two ANN types were developed: multilayer perceptrons (MLPs) and long short-term memory (LSTM) networks.
Additionally, transfer learning was applied using freely available AWC data from the German Weather Service.
An ensemble model combined all ANN outputs.
Fine-tuned models achieved the best prediction accuracy (R2>0.98, RMSE < 1.5%), with LSTMs slightly outperforming MLPs due to their ability to capture temporal dependencies.
The ensemble approach minimized cumulative errors and improved robustness across soil types and climatic conditions.
Explainable AI methods identified meteorological and crop-related inputs as key predictors, enhancing model transparency.
This study demonstrates that ANN ensembles are a powerful tool for irrigation scheduling in vegetable production as seen in the prototype decision support system ANNI (ANN for irrigation) developed in Geisenheim.
Authors
S. Rubo, M. Suomi, J. Zinkernagel
Keywords
artificial neural networks, ensemble modeling, irrigation scheduling, soil moisture, spinach, transfer learning
Groups involved
Online Articles (70)