Shani Friedman (Goldfarb)

Precision agriculture relies on translating complex sensor data into actionable insights for water management and stress detection. This study leverages a massive unique dataset of high-frequency measurements (3-minute intervals) collected from over 600 plants across a seven-year period (2018–2025), combining physiological plant metrics with micro-meteorological data.
Our previous published work demonstrated that traditional ML models (Random Forest, XGBoost) could successfully predict daily transpiration rates. To capture rapid physiological responses, we advanced to high-frequency time-series forecasting, benchmarking models such as Prophet, LSTM, and Residual CGRU. While the Residual CGRU and XGBoost regression achieved the lowest error rates (MAE 0.12), these “black-box” models lack the ability to explain the reason behind a plant’s behavioral shift.
We present a novel AI pipeline designed not just for prediction, but for reasoning. We developed a multimodal Large Language Model (LLM) framework (utilizing GPT-5.2 and Gemini 3 as the underlying models) to perform reasoning-based segmentation of time-series data. We developed a multi-stage AI pipeline that combines automated signal segmentation, retrieval-augmented physiological rules, and multimodal reasoning over tabular and visual data in inference, and augmented ingestion of domain knowledge during tuning and its presentation as a semantic layer into the context window during inference. This approach is particularly effective for imbalanced classification and tail performance.
Using foundation models for structured segmentation and explanation, we compare LLM-based methods with classical statistical and deep-learning approaches for detecting irrigation events, transpiration regime shifts, and technical artifacts.
Our results show that LLM-based segmentation achieves performance comparable to established algorithms, while offering substantially improved interpretability and contextual explanation, as well as superior adaptability to rare conditions found in field environments when compared to traditional ML models. We further examine computational costs and trade-offs between accuracy, scalability, and explainability.
This work demonstrates how large language models can transform high-frequency biological sensor data into transparent, explainable, context-depended, physiologically grounded insights. By enabling automated reasoning under uncertainty, our approach supports precision irrigation, early stress detection, and data-driven crop management.
Discussion points include: How reliable are LLMs for scientific signal interpretation? When does explainability justify higher computational cost? How can domain knowledge be integrated into foundation models? How can multi-year, high-frequency time-series data be efficiently represented and integrated into large language models for long-term reasoning? How can language models be used to address rare and under-represented conditions? What standards are needed for trustworthy AI in environmental monitoring?