EGU24-2072, updated on 08 Mar 2024
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Soil moisture forecasting for dryland fields in Australia

Qazi Muqeet Amir and Thomas Bishop
Qazi Muqeet Amir and Thomas Bishop
  • School of Life & Environmental Sciences, Sydney Institute of Agriculture, University of Sydney, Sydney, Australia (

Australia is frequently susceptible to droughts. Major droughts within the last 50 years such as that of 1982-1983 and the Millennial drought of 1997-2010 severely impacted crop growth across the country. Soil moisture can be in deficit during droughts due to a lack of recharge and high evapotranspiration from the soil. Dryland agriculture is particularly sensitive to droughts as there is no irrigation input into the soil. Soil water availability is a critical constraint to agricultural productivity, so the ability to predict its current and future state accurately is key in informing decisions relating to irrigation, fertiliser use, and yield targets. While soil moisture forecasting has been conducted in literature previously, there is limited understanding of the spatial, seasonal, and meteorological patterns that underlie the forecastability of soil moisture in a particular field. Hence this research aims to understand the spatial, seasonal, and meteorological factors that influence the forecast accuracy of soil moisture in dryland fields in Australia. 

Across Australia an increasing number of growers have soil moisture probes, which report current and historic soil moisture. The domain of this work is in the CosmOz probe network, consisting of 26 cosmic ray soil moisture probes across Australia, accounting for various geophysical and climatic regions. The probes measure average soil moisture to depths in the soil between 10 to 50 cm. Forecasting soil moisture requires the addition of various modelled/remotely sensed data such as meteorological, vegetation type, and soil property data. Using this data and lagged soil moisture as predictors, soil moisture has been forecasted at the locations of each CosmOz probe. With up to 13 years of training data, machine learning models have been fitted to forecast soil moisture with high accuracy forecasts of up to 30 days. To improve predictions a neural network autoencoder has been employed to engineer features that account for anomalous periods in the predictors.

A key outcome of this study is identifying patterns in forecast accuracy and predictor importance with respect to region, soil type, meteorological conditions, and time of year. These patterns create a nationwide perspective of soil moisture forecastability and the potential for forecasting in areas with no soil moisture probe data available. 

How to cite: Amir, Q. M. and Bishop, T.: Soil moisture forecasting for dryland fields in Australia, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2072,, 2024.