Monitoring standing herbaceous biomass and thresholds in semiarid rangelands from harmonized Landsat 8 and Sentinel-2 imagery to support within-season adaptive management

Abstract

Adaptive management requires rangeland managers to respond to changing forage conditions (e.g., standing herbaceous biomass) within the grazing season at the scale of individual pastures. While within-season biomass can be measured or estimated in the field, it is often impractical to make field measurements in extensive rangeland systems with adequate frequency and spatial representation for responsive decision-making by rangeland managers. We sought to develop a single model to accurately predict daily herbaceous biomass across seasonally and annually varying conditions from the Harmonized Landsat-Sentinel satellite time series. We also sought to assess how information about plant community composition derived from a high-spatial resolution map would improve model performance. We used herbaceous biomass data from 1764 ground observations collected over 8 years in North American shortgrass steppe for training in a cross-validated model selection approach to evaluate (1) predictive performance of candidate models both spatially and temporally, (2) relative variable importance of individual spectral bands, vegetation indices, and recently developed broadband spectral angle indices, and (3) the degree to which including plant community composition improved model performance. All 11 candidate models identified in the model selection procedure contained a band angle index and an individual spectral band, and 6 contained one of each input feature type, demonstrating the benefit of combining spectral features for predicting herbaceous biomass across varying conditions. The spatial and temporal cross-validation and selection procedures produced the same top model with similar performance (mean absolute error = 151–182 kg ha−1; R2 = 0.75–0.79), suggesting that a single model performs well over space and time. Including plant community composition in the model reduced mean absolute error by 11–13%. Bootstrapping revealed that –six to seven years of training data were required to achieve consistent model performance across years with varying environmental conditions (e.g., wet, average, dry). The top model could accurately detect (70–87% accuracy) the week that biomass dropped below management-related thresholds as low as 450 kg ha−1 in an independent dataset (n = 950) with modest commission error (10–26%). We discuss how maps showing the probability that herbaceous biomass is below a given threshold can support adaptive management in extensive semiarid rangelands across differing scenarios of risk perception and avoidance. In addition to producing maps to support precision rangeland management strategies, this study demonstrates the importance of combining complementary vegetation indices and acquiring long-term training datasets to achieve reliable predictions of herbaceous standing biomass in highly variable systems.