Mapping agricultural tile drainage in the US Midwest using explainable random forest machine learning and satellite...

Area of Science:

• Agricultural Science
• Environmental Science
• Remote Sensing

Background:

• Agricultural tile drainage is expanding globally due to farming intensification and climate change.
• Accurate, spatially explicit maps of tile drainage are lacking, hindering hydrologic modeling and nutrient management.
• Existing methods struggle to provide high-resolution tile drainage data across large agricultural regions.

Purpose of the Study:

• To develop and validate a machine-learning model for creating a high-resolution (30-m) Spatially Explicit Estimate of Tile Drainage (SEETileDrain) across the US Midwest.
• To identify key satellite-derived and environmental features influencing tile drainage.
• To provide accurate data for improving agricultural water and nutrient management strategies.

Main Methods:

• Developed a machine-learning model (SEETileDrain) using 31 satellite-derived and environmental features.
• Trained the model with over 60,000 ground truth points on the Google Earth Engine platform.
• Employed feature importance metrics and Accumulated Local Effects for model interpretation.

Main Results:

• The SEETileDrain model achieved high accuracy (96% correct classification, F1 score of 0.90).
• County-aggregated tile drainage area showed strong agreement with Ag Census data (r² = 0.69).
• Median summer nighttime Land Surface Temperature (LST) and soil moisture were the most influential features.

Conclusions:

• Satellite remote sensing effectively maps agricultural tile drainage at a large scale with high spatial explicitness.
• The SEETileDrain model provides valuable data for land use monitoring, hydrologic, and nutrient models.
• The developed algorithms have potential for broader remote sensing mapping applications in agriculture.