1R. A. Crabbe, 2D. Lamb, 3C. Edwards
1. Precision Agriculture Research Group, University of New England, Armidale NSW Australia
2. Precision Agriclture Research Group, University of New England, Armidale NSW Australia
3. Central Tablelands Local Land Service, Mudgee NSW Australia and Precision Agriculture Research Group, University of New England, Armidale NSW Australia
CAN PASTURE SPECIES COMPOSITION BE DISCRIMINATED FROM SPACE?
Richard Azu Crabbe*,1, David W. Lamb1 and Clare Edwards1,2
*Corresponding author: rcrabbe@myune.edu.au
1Precision Agriculture Research Group, University of New England, Armidale NSW Australia
2Central Tablelands Local Land Service, Mudgee New South Wales Australia
Background:
A mixture of different grass species of a pasture composition (e.g. grass types and legumes) in extensive grazing fields is not uncommon especially in natural grazing communities. Information on the botanical composition of pastures is therefore important to farm managers as it provides feedback on quality of available stock feed as well as, potentially, soil fertility. Such feedback informs farm managers seeking to optimise fertiliser application and manage their stock rotations. Visual monitoring of pasture species composition is typically used, especially at a smaller spatial scale. However, this is difficult to practice in extensive grazing systems owing to the geographical coverage and number of sub-seasonal field visitations required. Satellite remote sensing offers a cost-effective alternative for regular monitoring of large-scale grazing farms. This project seeks to explore the potential of the recently launched Sentinel-1 (C-band synthetic aperture radar imaging) and Sentinel-2 (multispectral imaging) satellite systems to discriminate pasture species composition of an extensive grazing farm.
Methods:
This study involved 20 field calibration sites (30 x 30 m) of varying grass types, such as Rytidosperma spp. (Danthonia), Bothriochloa macra (red grass,), Sporobolus africanus (Parramatta grass) and Poa labillardieri (tussock grass), established on UNE’s 2,700 ha SMART Farm (Armidale NSW Australia), and a field sampling protocol developed around the collection and manual classification of pasture botanical composition. Field sampling dates coincided with Sentinels-1 and -2 overflights. Four different variables derived from the dual polarisation of the Sentinel-1(normalised VV, VH polarisation and their product-VV.VH and ratio-VV/VH) and five optimal vegetation indices estimated from Sentinel-2 spectral bands (Enhanced Vegetation Index, Normalised Water Difference Index, Normalised Difference Index-45, Meris Terrestrial Chlorophyll Index and Sentinel-2 Red Edge Position) were used as features in a random forest (RF) classifier to separate pasture species composition into C3 perennials, C4 perennials and mixed classes. RF models were built separately for Sentinel-1 and Sentinel-2 derived features on their own and then merged these into a general model. To test the portability of the model across different stages of pasture growth, the model was trained with data from one period of the growing season and validated using a data from different growth stages.
Results:
The classification model built using features derived from the multispectral Sentinel-2 imagery performed better in predicting pasture species classes than C-band SAR Sentinel-1 data alone (overall accuracy=0.79, Kappa=0.68 and overall accuracy=0.49 and Kappa=0.23 for exclusive Sentinel-2 and Sentinel-1 models, respectively). The integration of the Sentinel-1 data, which provides structural and textural information of the plant canopies did not significantly improve the overall predictive power of the model over just the exclusive use of Sentinel-2 vegetation indices; this might suggest that biochemical properties (e.g. chlorophyll and water) of sward canopy mask the contributions of the canopy’s structural and textural variations in discriminating species composition type.
