Data-driven forecasting of soil moisture based on enhanced pre-processing of data from on ground sensor networks

Accurate forecasting of soil moisture is essential for hydrology, yet real-world sensor networks are often affected by challenges including noise, missing data, and hardware limitations. This study proposes a preprocessing pipeline designed to identify and extract high-quality, gap-free intervals from a complex in-situ dataset collected from a sensor network deployed at the Parque Agrourbano de Valdebebas (Madrid, Spain). State-of-the-art deep learning models, namely Long Short-Term Memory (LSTM) networks and Transformer architectures, were benchmarked against a simple baseline, with all models trained on carefully curated data segments. Results indicate that both the LSTM and Transformer models substantially outperform the baseline across medium- to long-term horizons, demonstrating strong generalization on previously unseen intervals and sensors. These findings emphasize the importance of rigorous data preparation for reliable forecasting in noisy environmental contexts and demonstrate the potential of machine learning (ML) techniques for robust soil moisture prediction under realistic field conditions.