Abstract
Invasive para grass (Urochloa mutica), is widespread on monsoonal wetlands of Northern Australia, including World Heritage listed Kakadu National Park (KNP). A rational spatial framework is required to manage continuing invasions and set conservation objectives and priorities.Multispectral satellite imagery is applied to monitor para grass at medium and high spatial resolution on the 225 km2 Magela Creek floodplain of KNP. Knowledge- and object-based image analysis (OBIA) methods are developed to map vegetation, water depth and fire and used to assess: 1) the spatial vulnerability of native vegetation to invasion by para grass; and 2) the spatio-temporal dynamics of para grass on the floodplain.
Changes in para grass distribution are mapped over a 42 km2 of the floodplain, biennially from 2002 to 2010 using high spatial resolution imagery. Accuracy was optimised by: 1) applying a novel method to standardise the scale of image objects used in classifications; and 2) testing different classification models built using Stochastic Gradient Boosting. Models were trained using samples from single- or multi-date sources and three different OBIA predictor variable sets. Single-date models were most accurate but not transferrable, possibly due to high temporal variability in wetland features and insufficient number of images. Nett para grass cover, patch connectivity and range increased over the decade, although spatial distribution was highly variable between years. Para grass distribution patterns were defined along a depth habitat gradient and change dynamics was also associated with site-specific fire histories. In years following fire, cover was reduced in ‘shallow’ and ‘moderate’ depth habitats but accelerated in ‘deep’ habitats. ‘Moderate’ depths showed the greatest rate of increase, more persistent increases occurred within the ‘shallow’ depth and no significant increase was detected in the ‘deep’ habitat. Native rice and Hymenachne grassland communities were most vulnerable to invasion.
A trade-off is required between the scale, accuracy and precision of data and the need to cover extensive and remote areas in a timely and cost-effective manner. Also, extrapolation and validation at one scale requires comparative information sub sampled at finer scales. Adherence to a dedicated protocol for sampling of imagery and field information will reduce inconsistencies contributing to map uncertainty. Improved coordination of monitoring and controlling weeds over multiple scales will also improve the efficacy of management decisions.
| Date of Award | Oct 2015 |
|---|---|
| Original language | English |
| Supervisor | Penny Wurm (Supervisor), Karen Joyce (Supervisor) & Guy Stuart Boggs (Supervisor) |