Development of long short-term memory models using rainfall and soil moisture to predict soil moisture dynamics

Area of Science:

• Agricultural Science
• Hydrology
• Data Science

Background:

• Accurate soil moisture prediction is crucial for efficient irrigation in orchards.
• Understanding water movement at different soil depths is key to water management.
• Long Short-Term Memory (LSTM) models offer potential for complex time-series prediction in agriculture.

Purpose of the Study:

• To develop and evaluate LSTM models for predicting soil moisture at various depths in an Asian pear orchard.
• To compare the effectiveness of hourly versus event-based rainfall data as model inputs.
• To assess the impact of prediction time horizons on model accuracy.

Main Methods:

• Utilized hourly and event-based rainfall data, alongside soil moisture measurements at 20, 40, and 60 cm depths.
• Developed LSTM models to predict soil moisture at future time steps (t+1, t+3, t+6, t+12).
• Evaluated model performance using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).

Main Results:

• Short-term soil moisture predictions were more accurate than long-term predictions.
• Soil moisture increased with rainfall, with deeper layers responding more gradually.
• Models using event-based rainfall data showed significantly lower prediction errors compared to hourly data.
• Topsoil moisture fluctuated more than subsoil moisture during and after rainfall events.

Conclusions:

• Event-based rainfall data is a more effective input for LSTM models predicting soil moisture in this orchard setting.
• LSTM models utilizing event-based rainfall can support precision irrigation systems for optimized water use.
• The findings highlight the importance of input data type in hydrological modeling for agricultural applications.