Pasture biomass mapping in hill country using remote sensing and geospatial tools

Abstract

Measuring pasture biomass in hill country is challenging. Our objective was to demonstrate how the fine-scale spatial pattern of pasture biomass in a highly heterogenous grassland landscape can be quantified using multispectral remote sensing and spatial machine learning. Images derived from the Sentinel 2 satellite and topographical indices (e.g., slope, aspect), were used as predictor variables. These variables could all be captured remotely, meaning minimum requirement for ‘manual’ data provision by the land manager. Pasture biomass samples were collected from 43 pre-selected spatially balanced sites across the longterm phosphorus (P) and sheep grazing experiments located on AgResearch Ballantrae Research Station to train and validate the prediction model. The spatial pasture biomass model achieved a moderate prediction performance (R2 ~ 0.6, Root mean squared error = 581 kg dry matter/ha). This is a significant achievement, comparable to others, despite addressing the most diverse grassland landscapes at a finer scale. Our study provides insight into the pattern of pasture biomass in heterogenous landscapes, showing that biomass can be highly variable within a slope class, an aspect, or single paddock. Integrating remote sensing with spatial machine learning can improve pasture biomass estimates and advance our ability to routinely update pasture cover in feed budgets for diverse landscapes.