NEW YORK – LONG ISLAND Utilizing Short-Term Memory (LSTM) Machine Learning Algorithm to Create Soil Moisture Prediction Models and Improve Water Productivity in Southern California

Overview

Due to changes in the climate and poor water management practices, water scarcity is rampant (“Water Scarcity,” 2019). Climate change will further increase variability in rainfall, with more dry spells, droughts, and floods, increasing the threat of a lack of fresh water and limiting available natural resources (Rockstorm et al., 2007). Most current agricultural systems use outdated irrigation methods and overuse water, demonstrating a lack of automated and efficient irrigation scheduling (Aguilar et al., 2015). Mass water wastage significantly reduces crop yields, severely impacting the world food supply. This project created an irrigation scheduling model using Long Short-Term Memory networks to accurately predict soil moisture using various environmental factors. Python code was written in Jupyter Notebook using the Tensorflow Deep Learning library. Through testing hyperparameters (batch size, epoch), the model was optimized, ensuring accurate and robust predictions which were validated through a comparison of actual vs. predicted values in a t-test, where a p-value of 0.47 was obtained, demonstrating there was no significant difference (actual and predicted values were statistically similar). Subsequent statistical tests resulted in a low mean square error (0.213) and a high r2 value (0.852), demonstrating model accuracy. In conclusion, LSTM algorithms were able to create an accurate soil moisture prediction model for Southern California’s agricultural output. Such prediction models will be useful in water conservation, helping improve the water productivity and water use efficiency in a given system, and being cost- effective and efficient in creating a decision support system for irrigation scheduling.