Mapping grassland plant communities subjected to nutrient influx and disturbance using UAV remote sensing and machine learning

Authors: Rowley, David; FAY, PHILIP - Retired ARS Employee; Flynn, Kyle; MARTINA, JASON - Texas State University; TREADWELL, MORGAN - Agrilife Research; ROGERS, WILLIAM - Texas A&M University

Submitted to: Society for Range Management Meeting Abstracts
Publication Type: Abstract Only
Publication Acceptance Date: 11/8/2022
Publication Date: 2/13/2023
Citation: Rowley, D.W., Fay, P.A., Flynn, K.C., Martina, J.P., Treadwell, M.L., Rogers, W.E. 2023. Mapping grassland plant communities subjected to nutrient influx and disturbance using UAV remote sensing and machine learning [abstract]. Society for Range Management Meeting Abstracts, Boise, ID, February 12-16, 2023.

Interpretive Summary: Nutrient influx and disturbance are among the most pervasive and severe drivers of global change in grassland plant communities. Future sustainability relies on detecting and understanding changes in these ecosystems. The emergence of unmanned aerial vehicle (UAV) remote sensing technology as a tool to efficiently collect data over large spatial scales suggests its use as a technique to assess changes in grassland community composition. We evaluated multispectral imagery on 25 grassland communities during the 2022 growing season (March – September). Each community was designated as a control plot (C) or a plot subject to nutrient influx (NPKµ), disturbance (D), or nutrient influx + disturbance (NPKµ+D). We compiled three datasets from the five spectral bands and 13 calculated vegetation indices: (1) spectral bands only (B), (2) vegetation indices only (I), and (3) spectral bands + vegetation indices (B+I). Community composition from each dataset was classified using a supervised classification method based on a Random Forest classifier. Classification accuracies were dependent on both the dataset (B, I, B+I) and date. Classification accuracy was best for all three datasets between May – July. This range of months coincided with peak productivity and peak phenological differences between plant functional groups. Classification accuracies for all three datasets were lowering during early growing season months (March and April), as well as late growing season months (August and September). This can likely be explained by difficulties in detecting phenological diversity between plant functional groups during the early growing season and difficulties in detecting differences between living and senesced plant matter during the late growing season while training the Random Forest. The results from our study indicate that UAV multispectral imagery is indeed sufficient at assessing grassland communities subjected to nutrient influx and disturbance, however, attention is warranted when considering timing of image acquisition.

Technical Abstract: Nutrient influx and disturbance are among the most pervasive and severe drivers of global change in grassland plant communities. Future sustainability relies on detecting and understanding changes in these ecosystems. The emergence of unmanned aerial vehicle (UAV) remote sensing technology as a tool to efficiently collect data over large spatial scales suggests its use as a technique to assess changes in grassland community composition. We evaluated imagery on 25 grassland communities during the 2022 growing season (March – September). Each community was designated as a control plot (C) or a plot subject to nutrient influx (NPKµ), disturbance (D), or nutrient influx + disturbance (NPKµ+D). Community composition from each dataset was classified using a remote sensing-based supervised classification method based on a Random Forest classifier. Classification accuracies were dependent on both the remote sensing dataset and date data were aquired. Classification accuracy was best for all three datasets between May – July. This range of months coincided with peak productivity and peak phenological differences between plant functional groups. Classification accuracies for all three datasets were lowering during early growing season months (March and April), as well as late growing season months (August and September). This can likely be explained by difficulties in detecting phenological diversity between plant functional groups during the early growing season and difficulties in detecting differences between living and senesced plant matter during the late growing season while training the Random Forest. The results from our study indicate that UAV imagery is indeed sufficient at assessing grassland communities subjected to nutrient influx and disturbance, however, attention is warranted when considering timing of image acquisition.