A Remote Sensing Method for Estimating Productivity Measures in Guayule Using High-Resolution Spectral and Structural Data

Abstract

The agricultural sector, with its market-driven crop economics and evolving strategies toward resource and pest management, has entered the age of “precision” or “digital” farming. The ongoing efforts to commercialize guayule (Parthenium argentatum Gray) as an alternative source of natural rubber requires creative solutions to estimating crop productivity if the adoption of guayule rubber should expand and do so sustainably. Satellite remote sensing products such as the Normalized Difference Vegetation Index (NDVI) are often linked to plant phenology and common measures of plant development such as crop biomass (i.e., fresh or dry weight) or volume for large regions at coarse spatial resolutions (pixels that are tens-of-meters to kilometers in size). Similarly, commercially-available multispectral sensors mounted to unmanned aircraft systems (UAS; i.e., drones) now offer spatial resolutions well below one meter, effectively allowing for individual plant or leaf-level observations. Structure-from-Motion (SfM) photogrammetry, which re-creates 3-dimensional (3-D) scenes from dense collections of true-color images, is also being linked to biomass and volume measures via crop surface models which incorporate canopy height (CH) information. Until recently, these data were largely viewed independently. We assessed the performance of regression models integrating both NDVI and CH information for estimating measures of crop productivity, including fresh weight (FW), dry weight (DW), fresh volume (FV), fresh-weight-density (FWD; the fresh weight of plant material adjusted by its freshly harvested volume), and dry-weight-density (DWD; the dry weight of plant material adjusted by its freshly harvested volume). Model parameters included mean pixel NDVI, SfM-derived mean canopy height (CH), a term representing the interaction between NDVI and CH, and categorical variables representing the variability of resource allocation within the vertical profile (i.e., at multiple levels, or tiers) in guayule. The FWD model incorporating NDVI, CH, NDVI:CH interaction, and tier parameters reported a mean absolute percentage error (MAPE) between 9 and 13% when comparing field measurements and model predictions, the best-performer of all response variables considered. The full FWD model incorporating all independent variable terms was reduced to a model inclusive of only NDVI and tier parameters for comparison against model predictions based on Sentinel-2 satellite data, which lack canopy height information. MAPE between FWD model predictions based on UAS and satellite imagery were below 3% across all UAS surveys and corresponding satellite scenes evaluated, suggesting model predictions scaled to medium spatial resolution data sources are achievable and reliable.