Enhancing the prediction of irrigation demand for open field vegetable crops in Germany through neural networks, transfer learning, and ensemble models

Precise irrigation management in vegetable production is key for optimizing water use and ensuring crop productivity. This study develops two types of artificial neural networks (ANNs), multilayer perceptron (MLPs) and long short-term memory (LSTM) networks for the prediction of available water capacity (AWC in %) as target parameter for irrigation scheduling. These ANNs are trained with experimental data from three-year (2020–2023) open field trials with spinach on two sites in Germany, and for three soil layers (0–20 cm, 20–40 cm and 40–60 cm). This data encompassed soil texture, plant signals and plant developmental status derived from vegetation indices based on spectral reflectance along with meteorological variables including mean air temperature, humidity, wind speed, photothermal time, and their cumulative values. Two additional models are pretrained with freely accessible AWC data from 320 stations across Germany and subsequently fine-tuned with the same experimental data as before. An ANN ensemble model consolidates the knowledge from preceding models to enhance robustness and promote transferability to new climatic conditions and soil textures. Methods of explainable AI such as variable importance analysis and sensitivity analysis enhance the model explainability by identifying influential factors for each soil layer. Models trained with additional AWC data and fine-tuned with experimental performed best (R 2 > 0.98, RMSE <1.5 %) across all soil depths. The LSTM models perform slightly better than the MLP equivalent, emphasizing the importance of temporal dependencies in soil moisture prediction. The ensemble model minimized cumulative errors and provided stable results by averaging the outputs of all models. While ANNs provide highly accurate results, implementation requires expertise and resources of IT infrastructures such as the development of interfaces to weather stations and, if applicable, additional sensors. Consequently, deploying the ANN-based IS in practice requires a service provider with specialized knowledge in both IT and vegetable production for effective implementation and maintenance. • Artificial neural networks (ANN) accurately predict available water capacity (AWC) with R 2 > 0.98. • Transfer learning with open data from 320 weather stations enhances adaptability to different climates. • ANN ensemble model enhances irrigation demand robustness by integrating multiple networks. • Explainable AI (variable importance and sensitivity analysis) enhances understanding of model mechanisms. • Dynamic crop development modelled using vegetation indices (biomass, nitrogen) and photothermal time.