Machine-learning model to delineate sub-surface agricultural drainage from satellite imagery
Knowing subsurface drainage (tile-drain) extent is integral to understanding how landscapes respond to precipitation events and subsequent days of drying, as well as how soil characteristics and land management influence stream response. Consequently, a time series of tile-drain extent would inform one aspect of land management that complicates our ability to explain streamflow and water-quality as a function of climate variability or conservation management. We trained a UNet machine-learning model, a convolutional neural network designed to highlight objects of interest within an image, to delineate tile-drain networks in panchromatic satellite imagery without additional data on soils, topography, or historical tile-drain extent. This was done by training the model to match the accuracy of human experts manually tracing the surface representation of tile drains in satellite imagery. Our approach began with a library of images that were used to train and quantify the accuracy of the model, with model performance tested on imagery from two areas that were not used to train the model. Satellite imagery included acquisition dates from 2008 to 2020. Training imagery was from agricultural areas within the US Great Lakes basin. Validation imagery was from the upper Maumee River, tributary to western Lake Erie, and an Indiana, Ohio-River headwater tributary. Our analysis of the satellite imagery paired with meteorological and soil data found that during spring, a combination of relatively high solar radiation, intermediate soil-water content and bare fields enabled the best model performance. Each area of interest was heavily tile-drained, where better understanding the movement of water, nutrients, and sediment from fields to downstream water bodies is key to managing harmful algal blooms and hypoxia. The trained UNet model successfully identified tile drains visible in the validation imagery with an accuracy of 93%–96% and balanced accuracy of 52%–54%, similar to performance for training data (95% and 63%, respectively). Model performance will benefit from ongoing contributions to the training library.