Enhancing agricultural automation through weather invariant soil parameter prediction using machine learning.

Area of Science:

• Agricultural Science
• Data Science
• Environmental Monitoring

Background:

• Soil parameters are vital for crop yield, yet research on agricultural automation in Bangladesh is limited.
• Expensive sensors and controlled environments hinder large-scale soil parameter monitoring.
• There is a need for cost-effective solutions for predicting soil moisture and temperature.

Purpose of the Study:

• To develop an inexpensive method for predicting soil moisture and temperature.
• To establish a robust relationship between weather parameters (humidity, temperature) and soil parameters.
• To explore the feasibility of using machine learning for agricultural automation.

Main Methods:

• Utilized a custom dataset of ~9000 data points collected using inexpensive sensors in an uncontrolled agricultural setting.
• Applied machine learning models, including Multilayer Perceptron (MLP) and Random Forest, to predict soil parameters from ambient weather data.
• Evaluated model performance using R-squared, Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE).

Main Results:

• XGBoost regressor demonstrated superior performance, achieving R-squared scores of 0.93 for soil moisture and 0.99 for soil temperature.
• The model exhibited very low errors: RMSE of 0.037 and MAE of 0.015 for soil moisture; RMSE of 0.001 and MAE of 0.0008 for soil temperature.
• A strong correlation was found between weather and soil parameters, validating the prediction approach.

Conclusions:

• It is feasible to accurately predict soil parameters (moisture, temperature) from readily available weather data.
• This approach offers a cost-effective solution for mass agricultural automation, especially in regions like Bangladesh.
• The findings have significant implications for improving agricultural practices and resource management.