Improving classification of humid-region irrigation using the red-edge band of Sentinel-2: Comparing irrigated and non-irrigated corn and soy in southwestern Michigan

Irrigation can have profound effects on the surrounding hydrologic cycle. Despite this, very little data is available on the distribution and extent of irrigation across much of the world. Sustainable management of agricultural water use, and better understanding of changing agricultural systems, requires better data. While previous research using multi-spectral data from Landsat sensors has successfully been able to differentiate irrigated and non-irrigated farmland in more arid parts of the United States, humid regions experience significantly higher false positives in irrigation classification efforts. Spectral information in the "red-edge," a region of dramatic shift in reflectance by plants between the red and near-infrared wavelengths, can effectively discriminate vegetation chlorophyll content and overall plant health, but this has historically not been available on high resolution space-based sensors. The relatively recent operation of ESA's Sentinel-2 provides the opportunity to leverage information in the red-edge for sub-field scale classification. Here, we tested the ability of this new spectral information to improve the classification of irrigation in southwestern Michigan, a humid region with recent expansion in irrigated fields. We collected ground truth data to map and record changes in irrigation and crop type. We used Google Earth Engine to derive vegetation indices from available Landsat and Sentinel-2 imagery, including the Sentinel-2 Red-Edge Position (S2REP) index. We then used a random forest algorithm to classify irrigation both with and without red-edge indices to quantify the improvement in classification accuracy. We found that including the red-edge improved classification accuracy. This information can be incorporated into large-scale efforts to map irrigation dynamics in humid regions.