Comparing the predictive ability of Sentinel-2 multispectral imagery and a proximal hyperspectral sensor for the estimation of pasture nutritive characteristics in an intensive rotational grazing system

Abstract

Using remote sensing to estimate the nutritive characteristics of pasture in livestock systems is a rapidly developing area that has the potential to improve pasture management and utilisation and in turn improve livestock productivity. The Copernicus Sentinel-2 paired satellites have shown promise in predicting pasture nutritive characteristics and have a pixel size small enough to be able to identify intra-paddock nutrient variation in rotational dairy grazing systems where paddock sizes are small. This study sought to develop nutrient prediction models using a dataset of nine Sentinel-2A and -2B satellite images, by linking imagery to perennial ryegrass dairy pasture samples cut within known pixels of the images in the days directly after each cloud-free overpass. Just before destructive pasture sampling, a handheld, proximal, hyperspectral sensor was used to gather spectral signatures (400–2500 nm) from areas to be sampled. Pasture samples (n = 200) were analysed using wet chemistry techniques for seven nutritive characteristics. The dry matter concentration of the samples was also measured, creating eight variates to be modelled in total. The Sentinel-2 data from within the selected pixels was extracted and used to create input features for Random Forest prediction models. Input features included the raw multispectral bands, common vegetation indices, and independent variables ‘season’ and ‘month’. Predictive models were also developed from the hyperspectral sensor data using Partial Least Squares Regression. For both modelling types, variable selection was used to reduce model inputs to only those sensitive to each nutrient. Results comparing the final models in an independent validation test showed similar predictive performance of the Sentinel-2 satellite and the proximal hyperspectral sensor for all variates of interest, shown by mean absolute errors that were not significantly different. In each case, the dry matter concentration was the best predicted variable (Lin’s concordance correlation coefficient (LCCC) > 0.90) with excellent potential to be quantified by Sentinel-2 satellites. Crude protein, metabolisable energy, neutral detergent fibre, non-fibre carbohydrate, and water-soluble carbohydrate concentrations were also predicted with good to moderate accuracy by both sensors (LCCC between 0.50 and 0.80). These models would be sufficiently accurate for making qualitative predictions (e.g., grouping into high to low categories). Ash and acid detergent fibre concentrations were poorly predicted (LCCC < 0.50) and these models would not be recommended for further use. It was concluded that multispectral data from Sentinel-2 has great potential to estimate pasture nutrient concentrations for intensively grazed pastures such as those used in dairy systems in temperate regions.