Mapping fourteen key vegetation species on an annual basis using satellite imagery, environmental data, and machine learning to support environmental monitoring.
The Challenge
The Murray–Darling Basin Authority (MDBA) works with Basin state governments, industries and communities to manage water in the Murray–Darling Basin. An important element of this is to monitor ecosystems that depend on the river system, including native vegetation. In particular, the MDBA must report on the extent of three native vegetation species every five years as part of reporting against key outcomes of the 2019 Basin-wide Environmental Watering Strategy.
In 2013, a method was developed to map the three key vegetation species. However, the method required ten years of satellite imagery to produce an accurate result, making it unsuitable for the MDBA’s five-year reporting timeframe. With the advent of higher resolution and more frequent satellite imagery captures, as well as advances in machine learning and cloud computing, there was a great opportunity to produce a new method that could produce vegetation species extent maps annually, providing much finer detail for reporting than previously possible.
Partners
The project was a collaboration between the Murray–Darling Basin Authority, FrontierSI, and NGIS.
The Solution
The project successfully developed a machine learning model that achieved an overall accuracy of 80.3% on 14 relevant vegetation species, a significant advance on the three species accurately modelled by the previous method. In addition, the new method produced both individual species likelihood layers and a final classification extent map, providing additional information to the MDBA about the level of model confidence in any given area. The likelihood maps can also be used as the basis for customised vegetation extent maps, which could show extent with different levels of model confidence or incorporate multiple existing land use and land cover datasets to better map true vegetation extent. The project delivered the likelihood maps and a classification extent map for 2022 and demonstrated a change detection approach by running the approach for a small area comparing classifications from 2021 to 2022.
The project made efficient use of existing ground survey data by taking satellite measurements from multiple years for each species presence data point. This more than doubled the number of training points for the machine learning approach and produced a far more accurate model than could be achieved by only using satellite imagery from a single year. This approach also had the advantage of including inter-annual variability in the distribution of model features, supporting the model to make stable predictions across different years.
Impact
The MDBA can now use the trained machine learning model to annually map fourteen vegetation species across the Basin from freely available satellite and environmental datasets. This will underpin their five-year reporting on vegetation extent. The MDBA has successfully deployed the model and source code through Google Earth Engine and VertexAI, allowing them to extend the approach with new training data in future years, further refining the model and potentially adding more species as additional field data are collected across the Basin.
Contact
To learn more, contact FrontierSI at contact@frontiersi.com.au.
