Leveraging Microclimate Data, Proximal Canopy Sensing, and Machine Learning for Precise Soil Water Potential Prediction

Introduction

Accurately predicting soil water potential is crucial for optimizing irrigation practices and minimizing water waste in agriculture. This project delves into the critical relationship between high resolution, in-field microclimate measurements, canopy surface temperature, time of day, and soil water content and potential. Building upon our previous work in biophysical modeling, we employ advanced machine learning techniques to investigate the impact of measurement timing on prediction accuracy.

Enhanced Prediction with Machine Learning

This project utilizes a range of machine learning algorithms, including:

• Linear Regression: Establishes a linear relationship between features and target variables.

  
  

• Decision Tree Regressor: Makes decisions based on a series of if-else questions.

  
  

• Random Forest Regressor: An ensemble method combining multiple decision trees for improved accuracy and reduced overfitting.

  
  

• Gradient Boosting Regressor: Iteratively builds a model, focusing on correcting errors from previous models.

  
  

• Support Vector Regression (SVR): Maps data to a higher-dimensional space for finding a linear relationship.

  
  

• Multi-Layer Perceptron (MLP) Regressor: A neural network model suitable for complex patterns.

  
  

• Custom TensorFlow/Keras Neural Network: A neural network model built using TensorFlow/Keras.

We experimented with various models, evaluated performance using machine learning metrics, and visualized the results.

Data-Driven Insights from Microclimate Sensors

By analyzing data collected from multiple locations using IoT-enabled sensor networks, we've uncovered valuable insights into the correlation between soil water content and potential (a more precise indicator of plant water stress than soil moisture) and various environmental factors.

Our findings demonstrate that:

• Data Type Matters: The choice of data type (daily average, late morning average, or 15-minute data) significantly impacts prediction accuracy.

  
  

• Time of Day is Key: Introducing "time of day" as an input variable significantly improves the models' ability to predict water content and potential.

  
  

• Morning Data for Optimal Prediction: While daily data is essential for soil water content, morning average data (9:00 AM - 11:00 AM) appears optimal for estimating soil water potential. Further data collection is needed for confirmation.

Explore the Project on GitHub

We've created a comprehensive Github repository showcasing the application of machine learning and neural networks for predicting soil water content and potential using microclimate data (including canopy surface temperature).

The repository includes:

• Representative Microclimate Data:  Data is collected from six apple orchards in the United States, including wind speed, solar radiation, relative humidity, air temperature, canopy surface temperature, and soil temperature. One file includes "time" data.

  
  

• Sample Dataset for Model Testing: Over 8,000 records of microclimate data collected every 15 minutes during a growing season allow you to see how the models perform with real data. You can also experiment with your own structured data.

  
  

• Enhanced Code Features:

  • Streamlined Code and Readability: A new data_analysis_utils package houses data processing, machine learning, and neural network model handling libraries.

    
    

  • Sample Microclimate Data:  A dedicated microclimate_data directory provides clear project structure.

    
    

  • Diverse Machine Learning and Neural Network Models: Integration of various models for training and evaluation.

    
    

  • Clear Data Visualization: Plots and tables for improved data representation.

    
    

  • Model Persistence: Allows saving, loading, and evaluating trained models for re-use and performance monitoring.

    
    

  • Optimized Neural Networks: Hyperparameter tuning, adjusted settings, and proper layer design (number of hidden layers and nodes) for improved performance.

Conclusion

This project, available on GitHub (https://github.com/envitronicslab/Soil_Water_Prediction), contributes to a deeper understanding of soil water dynamics and provides valuable tools for agricultural professionals and researchers. By employing machine learning techniques and leveraging microclimate data, we can achieve more accurate and efficient water management practices.