Wed, Aug 04, 2021:On Demand
Background/Question/Methods
Subtropical pasture is a significant yet understudied component of global grazing lands and provides numerous ecosystem services including food security and greenhouse gas mitigation via soil carbon storage. Overall, the delivery of ecosystem services from subtropical pasture depends on its productivity and quality (e.g. nitrogen content), both of which exhibit large heterogeneity due to variations in environmental conditions and management, challenging efforts to quantify them at scale. Fortunately, current remote sensing technology has demonstrated promise in measuring spatiotemporal variability within agricultural landscapes, although the applied efficacy of various spectral and spatial resolutions must be established. Here, we report findings from two experiments for estimating key variables of subtropical pasture productivity across spatial scales. In the first experiment, small (1m2) experimental plots were measured with a UAV-borne hyperspectral sensor and sampled for biomass and nitrogen content. In the second, a time series of sampled biomass and nitrogen content from large (0.8 ha) experimental paddocks was compared to coinciding historical Landsat 8 imagery and gridded climate data. Several statistical and machine learning models were developed and evaluated for performance at predicting biomass and nitrogen content, key state variables relevant both to land managers and ecosystem models to quantify and document ecosystem services.
Results/Conclusions Effective models for predicting the response variables were achieved for both spectral datasets. For the hyperspectral data, Bayesian hierarchical models using all spectral bands predicted out-of-sample biomass with a R2 of 0.35 and out-of-sample nitrogen content with a R2 of 0.62. For the Landsat 8 and gridded climate data, a multiple linear regression model using Landsat 8 bands along with average temperature and solar radiation as features predicted out-of-sample biomass with a R2 of 0.72 and out-of-sample nitrogen content with a R2 of 0.78. Despite the large disparity in spatial and spectral resolution, the multi-spectral/climate models outperformed the hyperspectral models, indicating the significant predictive contribution of the climate features likely due to phenology. Satellite-borne remote sensing platforms also benefit from their range of coverage, allowing the monitoring of much larger areas than UAV. We conclude with an overview of our ongoing reconciliation of these data into a process-based model for pasture production and the potential for this tool to improve grazing management and the quantification of ecosystem services at scale.
Results/Conclusions Effective models for predicting the response variables were achieved for both spectral datasets. For the hyperspectral data, Bayesian hierarchical models using all spectral bands predicted out-of-sample biomass with a R2 of 0.35 and out-of-sample nitrogen content with a R2 of 0.62. For the Landsat 8 and gridded climate data, a multiple linear regression model using Landsat 8 bands along with average temperature and solar radiation as features predicted out-of-sample biomass with a R2 of 0.72 and out-of-sample nitrogen content with a R2 of 0.78. Despite the large disparity in spatial and spectral resolution, the multi-spectral/climate models outperformed the hyperspectral models, indicating the significant predictive contribution of the climate features likely due to phenology. Satellite-borne remote sensing platforms also benefit from their range of coverage, allowing the monitoring of much larger areas than UAV. We conclude with an overview of our ongoing reconciliation of these data into a process-based model for pasture production and the potential for this tool to improve grazing management and the quantification of ecosystem services at scale.