Improved remote sensing methods to detect northern wild rice (Zizania palustris L.)

Authors: DENNIS, KRISTEN - National Aeronautics And Space Administration (NASA); LAROE, JILLIAN - Colorado State University; VORSTER, ANTHONY - Colorado State University; YOUNG, NICHOLAS - Colorado State University; EVANGELISTA, PAUL - Colorado State University; MAYER, TIM - Colorado State University; Carver Jr, Daniel; SIMONSON, ELI - National Aeronautics And Space Administration (NASA); ARIAS, VANESA - University Of Alabama; KERN, ANTHONY - University Of Minnesota; Khoury, Colin; RADOMSKI, PAUL - Minnesota Department Of Natural Resources; KNOPIK, JOSHUA - Minnesota Department Of Natural Resources

Submitted to: Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/9/2020
Publication Date: 9/16/2020
Citation: Dennis, K., Laroe, J., Vorster, A., Young, N., Evangelista, P., Mayer, T., Carver Jr, D.P., Simonson, E., Arias, V.M., Kern, A., Khoury, C.K., Radomski, P., Knopik, J. 2020. Improved remote sensing methods to detect northern wild rice (Zizania palustris L.). Remote Sensing. 12(18). Article e3023. https://doi.org/10.3390/rs12183023.
DOI: https://doi.org/10.3390/rs12183023

Interpretive Summary: Wildrice (Zizania palustris L.) is a culturally and ecologically important cereal that grows in shallow lakes in the upper midwestern USA and into Canada. It is considered to be in decline but given its wide range and difficult habitat to monitor through field work, gaps in information on its distribution constrain conservation planning. We developed improved remote detection methods for identifying wildrice from satellite imagery and mapping its distributions at relatively high resolution. These methods can be used to inform conservation efforts.

Technical Abstract: With populations of northern wildrice (Zizania palustris L.) declining during the last century, new and innovative techniques to support monitoring and conservation are needed for this culturally and ecologically significant plant. Current techniques include non-comprehensive, time-intensive field surveys and thresholding of spectral and physical datasets. We trained three wildrice detection models using data from annual aquatic vegetation surveys wildrice in northern Minnesota. These training datasets, which varied in definition of wildrice presence, were combined with Landsat 8 OLI and Sentinel-1 C-band SAR imagery to map wildrice in 2015 using the random forest algorithm. Spectral predictors were derived from phenologically important time periods of emergence (June - July) and peak harvest (August - September). The range of the VV polarization between the two time periods was consistently the top predictor. Model outputs were evaluated using both point and polygon validation; additionally, all predictions were overlaid and compared using an agreement map. While all models performed well in the point validation with percent correctly classified ranging from 83.8% to 91.1%, we found polygon validation necessary to assess wildrice detection accuracy at a level relevant for management. Generally, models had the greatest area of overlap with mixed wildrice stands while only overlapping half of the surveyed monotypic stands. The agreement map highlighted model consistency when detecting larger and denser stands and discrepancies throughout sparse areas of wildrice presence. Bulrushes (Schoenoplectus spp) were the species most commonly misclassified as wildrice, and misclassification rates differed by model. Our practical approach, which highlights a variety of applications, can be applied by resource managers and researchers to guide field excursions and estimate the extent of occurrence. Further testing and validation of the methods outlined to map wildrice using remotely sensed data may support multi-year monitoring efforts.