Using large-scale climate drivers to forecast meteorological drought condition in growing season across the Australian wheatbelt

Highlights

• A machine learning-based model was developed to predict drought.

• Lagged large-scale climate indices were adopted as input predictors.

• Growing season drought across all Australian cropping areas was predicted.

• Forecasted drought maps matched well with observed drought maps.

• NINO3.4 SST and MEI were identified as the most influential indices.

Abstract

Recurring drought has caused large crop yield losses in Australia during past decades. Long-term drought forecasting is of great importance for the development of risk management strategies. Recently, large-scale climate drivers (e.g. El Niño-Southern Oscillation) have been demonstrated as useful in the application of drought forecasting. Machine learning-based models that use climate drivers as input are commonly adopted to provide drought forecasts as these models are easy to develop and require less information compared to physical-based models. However, few machine learning-based models have been developed to forecast drought conditions during growing season across all Australian cropping areas. In this study, we developed a growing season (Apr.-Nov.) meteorological drought forecasting model for each climate gauging location across the Australian wheatbelt based on multiple lagged (past) large-scale climate indices and the Random Forest (RF) algorithm. The Standardized Precipitation Index (SPI) was used as the response variable to measure the degree of meteorological drought. Results showed that the RF model could provide satisfactory drought forecasts in the eastern areas of the wheatbelt with Pearson's correlation coefficient r > 0.5 and normalized Root Mean Square Error (nRMSE) < 23%. Forecasted drought maps matched well with observed drought maps for three representative periods. We identified NINO3.4 sea surface temperature and Multivariate ENSO Index as the most influential indices dominating growing season drought conditions across the wheatbelt. In addition, lagged impacts of large-scale climate drivers on growing season drought conditions were long-lasting and the indices in previous year could also potentially affect drought conditions during current year. As large-scale climate indices are readily available and can be rapidly used to feed data driven models, we believe the proposed meteorological drought forecasting models can be easily extended to other regions to provide drought outlooks which can help mitigate adverse drought impacts.

Introduction

Australia is a major global wheat (mainly bread wheat and durum wheat) producer and exporter (GLNC, 2020). It produces 3% of the world's wheat production but accounts for nearly 15% of the world's annual global wheat trade (AEGIC, 2019). Thus, Australian wheat production is of central importance to ensure global wheat supply and food security. Wheat in Australia is mainly grown under rain-fed conditions, thus the wheat industry is sensitive to climate disasters, especially drought. Recurring drought events have resulted in large yield losses in the past decades. For example, the 2018 drought resulted in a 53% reduction in winter crop production in eastern Australia compared to the average of past 20 years (ABARES, 2019). Mitigating the impacts of drought on crop production has been a major research focus as early and reliable drought forecasting can assist farmers to undertake management decisions (e.g. sowing time) which will have financial implications. However, forecasting drought events remains a challenge to the scientific community because their triggers are complex and their features are variable in time and space.

Drought is a normal part of the climate in almost all regions of the world, which results from prolonged absence or shortage of rainfall in comparison with normal years (Zarch et al., 2015; Yu et al., 2020). Australia is the second driest continent on earth because it is located under the subtropical high-pressure belt, which prevents the lift of air required for rain (Williams and Stone, 2009). Australia is also a country prone to drought, as rainfall shows great inter-annual variability throughout the continent. It is well established that the variability of Australian rainfall is linked to the climatic anomalies originating from its surrounding oceans, including the Pacific, Indian, and Southern Oceans (Risbey et al., 2009). These climatic anomalies have been described by several dominant large-scale climate drivers. They include the Indian Ocean Dipole (IOD) which is expressed by the sea surface temperature (SST) gradient between the western and eastern tropical Indian Ocean (Saji and Yamagata, 2003), the El Niño Southern Oscillation (ENSO) defined by the SST anomalies in the central-eastern equatorial Pacific Ocean (McBride and Nicholls, 1983), the Southern Annular Mode (SAM) which refers to the atmospheric circulation in the mid- to high-latitudes of the southern hemisphere on interannual timescales (Marshall, 2003).

The teleconnections between large-scale climate drivers and Australia's rainfall conditions are among the strongest in the world (Kirono et al., 2010). Previous studies have suggested that Australia's extreme hydroclimatic events in the past two decades largely resulted from the anomalies of these climate drivers and their interactions (Xie et al., 2019). For example, the 2002–2009 ‘Millennium drought’ was caused by a combination of long-term upward trend of SAM and prolonged lack of negative IOD phase (Ummenhofer et al., 2009). Following the Millennium drought, extreme wet periods (2010−2011) were found to be mostly driven by a sustained strong La Nina event (Luo et al., 2017), together with a concurrent positive SAM event (Gergis et al., 2012). The 2015 drought event originated from a strong El Niño event but was further enhanced by SAM and IOD variability (L'Heureux et al., 2017; Power and Delage, 2018). The drought events of recent decades in Australia occurred either across the entire continent or in specific regions, resulting in severe adverse impacts on crop production (Dijk et al., 2013). Thus, many studies have focused on empirical relationships between rainfall and large-scale climate drivers in order to provide timely and reliable climate outlooks and mitigate drought impacts (Risbey et al., 2009).

It is generally accepted that large-scale climate drivers have varying influences on rainfall in Australia which is dependent on geographic locations and seasons (Risbey et al., 2009). For example, the Nino 3.4 SST variability has been demonstrated to have a predominant impact on austral autumn rainfall in eastern Australia (van Rensch and Cai, 2014). SAM has been shown to mainly influence southern parts of Australia (Meneghini et al., 2007). Positive phases of SAM tend to result in an increase in spring rainfall in southwest Western Australia and New South Wales (King et al., 2014). There is generally, a negative correlation between IOD and rainfall from June to October in Western Australia, Victoria, South Australia, and southern New South Wales (Steptoe et al., 2018). It should also be noted that these relationships do not act independently, and each climate driver usually accounts for <20% of rainfall variability (Gallant et al., 2012; Risbey et al., 2009). Rainfall conditions throughout the Australian continent are generally the result of the synchronization of these climate drivers (Cleverly et al., 2016).

The teleconnections between large-scale climate drivers and Australian rainfall provide the scientific basis of data-driven drought forecast models (Abbot and Marohasy, 2014). Lagged values of large-scale climate drivers can be adopted as potential predictors of future drought conditions (Mera et al., 2018; Zhang et al., 2019). In the past few decades, a number of data-driven statistical models were developed within different regions of Australia for the forecasting of rainfall or drought conditions in the next month or season. For example, Mekanik et al. (2016) developed eight adaptive network-based fuzzy inference systems models based on lagged values of single or multiple climate drivers (ENSO, IOD, or Inter-decadal Pacific Oscillation) to forecast spring rainfall in Victoria. Their results suggested that the best performing models were able to forecast spring rainfall in a 10-year test period with acceptable correlation coefficients (r = 0.29–0.66) and low errors (RMSE = 10.9–25.0 mm) in 9 locations of Victoria. Abbot and Marohasy (2014) used artificial neural networks and lagged climate variables to forecast monthly rainfall with 1-month lead time in 3 sites of Queensland and achieved r values of >0.55. The two studies also demonstrated that statistical models developed using climate drivers performed better than the state-of-art dynamical model, the Predictive Ocean Atmosphere Model for Australia (POAMA) developed and run by the Australian Bureau of Meteorology. Physics-based dynamical weather forecasting models are normally considered as the mainstream approach by scientific community. However, dynamical models are usually expensive to operate and implement and rely overly on initial conditions. Despite of substantial technological advances and research efforts, dynamical models still have similar performance on seasonal rainfall forecasts in comparison to simple statistical models (Abbot and Marohasy, 2014). Thus, data-driven statistical models based on large-scale climate drivers can still be used to provide seasonal rainfall forecasts effectively in various regions of Australia.

The Australian wheatbelt produces about 25 million tons wheat per year (ABS, 2019). However, wheat yield varies greatly from year to year and is totally constrained by growing season rainfall. Reliable rainfall forecast across the wheatbelt is a critical first step to help growers reduce yield losses from drought. However, from our knowledge, most of previous studies were conducted in specific locations of a certain region, usually a state, and none has ever taken the Australian cropping areas as the domain. Cropping areas are critical areas that will directly benefit from reliable rainfall or drought forecasts. Furthermore, most studies focused on the prediction of next month or next 3 months and few have ever concentrated on the specific growing season (e.g. Apr.-Nov.) of the Australian cropping areas (Sacks et al., 2010). Growing season drought forecasts will provide more preparation time in comparison with seasonal forecasts for farmers to develop drought mitigation strategies. In addition, most previous studies developed linear models considering up to 9-month lead-time impacts of the climate drivers to find the predictability (Hossain et al., 2018). Given the complex effects of climate drivers on Australian rainfall, those linear models might fail to account for the climate drivers with more lead times. The dominant driver in a certain location and its nonlinear effect on rainfall also remain unknown.

Therefore, from the motivation of better understanding the relationship between growing season rainfall and various lagged large-scale climate drivers, the present study investigated growing season rainfall in the Australian wheatbelt as a case study. Instead of forecasting the absolute rainfall amount, we adopted the Standardized Precipitation Index (SPI) (Deo et al., 2017) as the dependent variable, which made the rainfall conditions comparable across different geographic regions. Moreover, it is also a commonly used meteorological drought index evaluating the degree of aridity for a certain location. We implemented a popular machine learning method, RF, as the regression technique to build forecasting models, instead of traditional linear models. The primary objectives of this study are to 1) develop growing season SPI forecasting models using various lagged large-scale climate drivers for each climate station throughout the Australian wheatbelt, 2) identify the best forecasting model for each location and compare their performance across the wheatbelt, 3) quantify the effects of the dominant climate drivers for each station on determining growing season SPI.