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

Authors: REDOLOZA, FLEFORD - Us Geological Survey; WILLIAMSON, TANJA - Us Geological Survey; HEADMAN, ALEXANDER - Us Geological Survey; Allred, Barry

Submitted to: Journal of Environmental Quality
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/3/2023
Publication Date: 5/27/2023
Citation: Redoloza, F., Williamson, T., Headman, A., Allred, B.J. 2023. Machine-learning model to delineate sub-surface agricultural drainage from satellite imagery. Journal of Environmental Quality. 52(4):907-921. https://doi.org/10.1002/jeq2.20493.
DOI: https://doi.org/10.1002/jeq2.20493

Interpretive Summary: Characterization of sub-surface agricultural drainage in terms of its presence or absence in a basin, location relative to other features of the surrounding landscape, and general structure is critical to understanding how this water management strategy might influence runoff response and water quality in the stream network. Our objective was to develop a machine-learning model (MLm) that could delineate tile drains with an accuracy that enables evaluating the density of drainage in a basin and how these networks might connect to the stream network. We trained a UNet MLm to evaluate panchromatic imagery and delineate sub-surface drainage visible in this satellite imagery: the training library included 107 images. Essentially, this MLm method functions similarly to how a human would approach the task of tracing these linear features: considering the image both as a whole and zoomed in at the pixel scale. No pre-conditioning was done on the imagery beyond subsetting these satellite tiles down to a uniform size (0.5 km x 0.5 km). No additional landscape or meteorological information was provided to the MLm to help retain an ability to look for sub-surface drainage with no dependence on where it might be anticipated as a function of soils, land management, or other physiographic properties. The UNet MLm successfully delineated tile-drains in images from two different validation areas that differed in ecoregion and climatic from the training dataset. However, there is a potential to improve this approach with the incorporation of new images and interpretation of individual images by multiple people. This underscores that development of a library of trained imagery for this type of question will be an ongoing effort that can build on what was used here.

Technical Abstract: A link between sub-surface agricultural drainage (tile drains) and basin-wide streamflow magnitude and water quality has been indicated by several studies. But, without an understanding of how tile-drain extent has changed with time, it is difficult to differentiate how this interaction has changed as a result of spatial extent, tile-drain network characteristics, and changes in climate. We present results from a UNet machine-learning model (MLm) that was trained to delineate tile-drain networks in panchromatic satellite imagery. The UNet MLm is a convolutional neural network designed to outline objects of interest within an image. The “U” in UNet refers to the “U” shaped architecture within the MLm designed to improve the accuracy of outlines traced out by the MLm. The workflow presented includes a library of training images, the accuracy of the MLm, and the performance of the model on imagery from two areas that were not used to train the model. Imagery included collections from GeoEye, WorldView, and QuickBird, with acquisition dates from 2008-2020. Training data are from agricultural areas within the U.S. Great Lakes basin. Validation data included imagery from the upper Maumee River, tributary to western Lake Erie, and a headwater of the White River, Indiana, an Ohio River tributary. Each of these areas of interest are heavily tile-drained, where a better understanding of the movement of water, sediment, and sediment-bound nutrients from fields to downstream water bodies is key to managing harmful algal blooms and hypoxia. Knowing where tile drains are is an integral step in beginning to understand how large areas of soil respond to individual precipitation events, subsequent days of drying, and seasonal fluctuations in evapotranspiration. A time-series of visible tile-drain systems will provide a framework to integrate field-scale knowledge about soil characteristics, land management, and environmental response.