Machine-learning model to delineate sub-surface agricultural drainage from satellite imagery

Area of Science:

• Environmental Science
• Remote Sensing
• Agricultural Engineering

Background:

• Understanding subsurface drainage (tile-drain) extent is critical for analyzing landscape responses to precipitation and drying.
• Tile-drain networks significantly influence streamflow and water quality, complicating assessments based on climate variability or conservation efforts.
• A time series of tile-drain extent is essential for improving hydrological models and land management strategies.

Purpose of the Study:

• To develop and validate a machine-learning model for delineating tile-drain networks from satellite imagery.
• To assess the model's performance without relying on soil, topography, or historical tile-drain data.
• To provide a tool for mapping tile-drain extent to better understand agricultural water management impacts.

Main Methods:

• A UNet convolutional neural network was trained to identify tile-drain networks in panchromatic satellite imagery.
• Model training involved expert-annotated imagery, with performance validated on independent datasets.
• Satellite imagery spanned 2008-2020, focusing on agricultural areas in the US Great Lakes basin and Ohio River headwaters.

Main Results:

• The UNet model achieved high accuracy (93%-96%) in identifying visible tile drains in validation imagery.
• Optimal model performance was observed in spring, correlating with high solar radiation, intermediate soil moisture, and bare fields.
• The model's accuracy was comparable to human expert performance on similar datasets.

Conclusions:

• Machine learning, specifically the UNet model, offers a robust method for mapping subsurface tile-drain extent using satellite data.
• Accurate mapping of tile drains is vital for managing water resources, mitigating nutrient runoff, and addressing issues like harmful algal blooms.
• Continued expansion of the training dataset will further enhance model performance and applicability.