Optimising Drought
Detection in Plants

Traditional agricultural drought detection relies heavily on visual symptoms of wilting or delayed soil moisture readings, by which point cellular damage and yield loss have already occurred.

This project explores the integration of machine learning models with real-time AgriTech sensor networks to detect physiological plant stress before physical symptoms manifest. By combining micro-climate data (ambient temperature, vapor pressure deficit) with non-invasive thermal imaging and root-zone moisture metrics, we can mathematically model the transpiration rate of crops in real time.

Using recurrent neural networks (LSTMs) and gradient boosting frameworks, the system processes multi-channel time-series data to predict cellular water deficits up to 48 hours earlier than conventional threshold-based automated irrigation systems.

The core algorithm handles significant noise present in real-world agricultural datasets, such as cloud cover anomalies in thermal data and sensor drift in high-salinity soils. A custom Kalman filtering pipeline was implemented to preprocess the raw data streams before feature extraction.

Key findings from initial simulations indicate that predictive irrigation scheduling driven by this model can reduce total water consumption by approximately 22% while maintaining optimal stomatal conductance across test crop cohorts.