Publication Date

12-2021

Series Number

419

Publisher

Department of Primary Industries and Regional Development

City

Perth

ISSN

1039-7205

Abstract

The Department of Primary Industries and Regional Development (DPIRD) is developing an integrated monitoring system using remote sensing and on-ground measurements to track pasture condition across Western Australia’s pastoral region. We extended and adapted the methods developed in the Pastoral Lease Assessment Using Geospatial Analysis (PLAGA) project (Robinson et al. 2012), which combined remotely sensed vegetation indices (VIs) with on-ground pasture condition observations to assess the potential of using different vegetation indices in a statewide condition monitoring system.

There were 6 regions in WA’s pastoral rangelands with DPIRD on-ground condition traverse points: Kimberley and Broome, Pilbara, Yalgoo and Sandstone, Goldfields, Nullarbor and Other Rangelands. In these regions, we evaluated 8 Landsat VIs (EVI, GDVI2, GDVI3, GDVI4, LMI, MSAVI2, NDVI and STVI-1) to measure their correlation with the on-ground condition classes (‘good’, ‘fair’, ‘poor’) in 5 stratification levels:

  • land system
  • functional group (land type)
  • pre-European vegetation type
  • pasture (habitat) type
  • broad vegetation group (management unit).

The experiments were done using 3 discrimination and classification strategies:

  • fairs-excluded strategy – 2 classes using good and poor points, and excluding the fair points
  • fairs-included strategy – 2 classes using good and fair points as one class, and poor points as the other class
  • 3-class strategy – 3-class machine learning classification using good, fair and poor points.

The VIs were computed using NBART Landsat 5 (Thematic Mapper [TM]), Landsat 7 (Enhanced Thematic Mapper Plus [ETM+]) and Landsat 8 (Operation Land Imager [OLI]) multispectral imagery downloaded from the Australian Geoscience Data Cube.

Discrimination potential and reliability of VIs varied among geographic regions and stratification levels. In the Nullarbor region, the LMI and STVI-1 produced the highest discrimination potential between condition classes (mean area under curve [AUC] = 0.70–0.80) and the highest overall reliability of 67–100% in all stratification levels. In contrast, the highest discrimination any VI could produce in the Yalgoo and Sandstone region was a mean AUC of 0.56–0.61 with correspondingly low overall reliability (less than 40%) by the EVI and LMI, which means in most of the groups in all stratification levels, no VI produced adequate discrimination. Of all the regions, the discrimination potential and reliability of using a VI to monitor condition was highest in the Nullarbor region.

Maps and tables in the appendixes outline the groups making up each stratification level and where VIs can or cannot be confidently adopted for monitoring pasture condition. The patch-wise use of VIs (rather than simply adopting a particular VI over an entire region) increased reliability, indicating that the use of VIs should be customised for each region.

Our study revealed several issues with using the current condition data to evaluate the potential of remote sensing VIs. No field data was available for many pastoral stations (especially in the Gascoyne, Meekathara, Murchison and Wiluna land conservation districts which were called ‘Other Rangelands region’ in this work) and correlations between VIs and pasture condition ratings in some groups within stratification levels were not adequate. This implies that the on-ground condition assessments need to be modified or that quantitative data needs to be collected separately to calibrate or validate remote sensing products.

Based on our findings and the limitations of available condition data, future research should focus on determining condition indicators (such as vegetation cover patchiness and clumpiness) using the VI with the highest adequate discrimination potential in each group of a stratification level. Monitoring groundcover change over time based on time series analysis is another aspect of condition that could be derived using the suggested VIs in each group of a stratification level.

Number of Pages

202

Keywords

remote sensing, vegetation indices, rangeland pasture condition monitoring, machine learning, Western Australia

Disciplines

Agriculture | Environmental Monitoring | Sustainability

Maps

Maps are not included as part of the complete document download. If this report contains a map, it will be available in the individual parts list below.

This file is 15.2 MB. Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS