Monitoring Pasture Species, Biomass and Canopy Heterogeneity Using Sentinel-1 Synthetic Aperture Radar Data

Abstract

<p>Pasture grasses are an important feed-base for the livestock industry. The ability to identify and characterise pasture type, species composition and quantify the available biomass in fields is invaluable to the sustainability and profitability of our livestock industries. Pasture species composition, biomass and canopy structural variations have been measured at different spatial scales using varied optical methods/tools such as active optical sensors, as well as aerial and spaceborne passive optical sensors. At large spatial scale, optical satellite sensors are often used. However, the utilisation of these sensors is affected by cloudy weather conditions and the fact that they are really only responsive to photosynthetically active biomass.</p> <p>Satellite-based Synthetic Aperture Radar (SAR) sensors, though not popular yet in pasture studies, have the potential to offset this limitation of optical sensing as the microwave energy emitted by these sensors penetrate clouds and that these wavelengths are also sensitive to volumetric scattering processes rendering them, potentially useful to situations involving significant, senesced, plant material (e.g. during drought) . This thesis predominantly focussed on Sentinel-1 C-band SAR with the whole research project comprising of three main components: (i) discrimination of pasture species based on C3 and C4 photosynthetic mechanisms and diversity of the botanical composition; (ii) estimating pasture biophysical variables with emphasis on aboveground biomass; and (iii) detection of surface heterogeneity due to selective grazing in pasture fields.</p> <p>In discriminating pasture species into C3, C4 and mixed C3/C4 classes, Random Forest classification was used and the highest overall classification accuracy (86%) was achieved with a combination of grey-level co-occurrence textural metrics and polarimetric SAR metrics. Moreover, the combined strengths of Sentinel-1 SAR and Sentinel-2 optical information parameterised into K-Nearest Neighbours, Random Forest and Support Vector Machine classifiers, produced the highest overall accuracy estimates of 89%, 96% and 95%, respectively. Regression models such as the generalised additive model estimated pasture biomass with a root mean square error of prediction of 392 kg/ha over AGB estimates between 443–2642 kg/ha. Here pasture LAI ranged from 0.27 to 1.87, and sward height from 3.25 cm to 13.75 cm. In the final study, canopy heterogeneity due to selective grazing was detectable with the Sentinel-1 SAR. Particularly, the range estimates (dispersion measure) of the polarimetric scattering entropy produced the strongest, statistically significant, linear correlation with a metric of patchiness (R² =0.74). Altogether, this thesis has demonstrated that Sentinel-1 SAR on its own as well as when integrated with optical data, could be a useful tool providing data to aid in pasture management.</p>