Advances in Machine Learning for Sensor Data with Applications in Smart Agriculture and Beyond

This project advances machine learning methods for sensor-driven agriculture, with a focus on forecasting tree- and soil water status. Motivated by challenges like data scarcity, mixed-frequency inputs, and complex physical dynamics, the work develops models that turn raw data into actionable insights for smart irrigation. Key contributions include scalable soil moisture forecasting using deep learning models such as transformers and time-series foundation models, which show strong performance even with limited data. To predict stem water potential, a key but sparsely measured indicator of tree water stress, a multi-task learning framework is introduced, dubbed MultiMix, that leverages highly available data (e.g. soil water potential data) to predict the sparsely available target, stem water potential. The approach generalizes well across crops and geographies. Finally, the project presents a self-supervised framework for anomaly detection in sensor data, enabling detection without labeled anomalies. Together, these contributions address core challenges in smart irrigation and offer broadly applicable methods for learning from complex sensor data.