Understanding the Drivers of Drought Onset and Intensification in the Canadian Prairies: Insights from Explainable Artificial Intelligence (XAI)

Jacob Mardian aDepartment of Geography, Environment and Geomatics, University of Guelph, Guelph, Ontario, Canada
bAgroClimate, Geomatics and Earth Observation Division, Science and Technology Branch, Agriculture and Agri-Food Canada, Ottawa, Ontario, Canada

Search for other papers by Jacob Mardian in
Current site
Google Scholar
PubMed
Close
https://orcid.org/0000-0002-6103-5392
,
Catherine Champagne bAgroClimate, Geomatics and Earth Observation Division, Science and Technology Branch, Agriculture and Agri-Food Canada, Ottawa, Ontario, Canada

Search for other papers by Catherine Champagne in
Current site
Google Scholar
PubMed
Close
,
Barrie Bonsal cWatershed Hydrology and Ecology Research Division, Environment and Climate Change Canada, Saskatoon, Saskatchewan, Canada

Search for other papers by Barrie Bonsal in
Current site
Google Scholar
PubMed
Close
, and
Aaron Berg aDepartment of Geography, Environment and Geomatics, University of Guelph, Guelph, Ontario, Canada

Search for other papers by Aaron Berg in
Current site
Google Scholar
PubMed
Close
Open access

We are aware of a technical issue preventing figures and tables from showing in some newly published articles in the full-text HTML view.
While we are resolving the problem, please use the online PDF version of these articles to view figures and tables.

Abstract

Recent advances in artificial intelligence (AI) and explainable AI (XAI) have created opportunities to better predict and understand drought processes. This study uses a machine learning approach for understanding the drivers of drought severity and extent in the Canadian Prairies from 2005 to 2019 using climate and satellite data. The model is trained on the Canadian Drought Monitor (CDM), an extensive dataset produced by expert analysis of drought impacts across various sectors that enables a more comprehensive understanding of drought. Shapley additive explanation (SHAP) is used to understand model predictions during emerging or worsening drought conditions, providing insight into the key determinants of drought. The results demonstrate the importance of capturing spatiotemporal autocorrelation structures for accurate drought characterization and elucidates the drought time scales and thresholds that optimally separate each CDM severity category. In general, there is a positive relationship between the severity of drought and the time scale of the anomalies. However, high-severity droughts are also more complex and driven by a multitude of factors. It was found that satellite-based evaporative stress index (ESI), soil moisture, and groundwater were effective predictors of drought onset and intensification. Similarly, anomalous phases of large-scale atmosphere–ocean dynamics exhibit teleconnections with Prairie drought. Overall, this investigation provides a better understanding of the physical mechanisms responsible for drought in the Prairies, provides data-driven thresholds for estimating drought severity that could improve future drought assessments, and offers a set of early warning indicators that may be useful for drought adaptation and mitigation.

Significance Statement

This work is significant because it identifies drivers of drought onset and intensification in an agriculturally and economically important region of Canada. This information can be used in the future to improve early warning for adaptation and mitigation. It also uses state-of-the-art machine learning techniques to understand drought, including a novel approach called SHAP probability values to improve interpretability. This provides evidence that machine learning models are not black boxes and should be more widely considered for understanding drought and other hydrometeorological phenomena.

For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Jacob Mardian, jmardian@uoguelph.ca

Abstract

Recent advances in artificial intelligence (AI) and explainable AI (XAI) have created opportunities to better predict and understand drought processes. This study uses a machine learning approach for understanding the drivers of drought severity and extent in the Canadian Prairies from 2005 to 2019 using climate and satellite data. The model is trained on the Canadian Drought Monitor (CDM), an extensive dataset produced by expert analysis of drought impacts across various sectors that enables a more comprehensive understanding of drought. Shapley additive explanation (SHAP) is used to understand model predictions during emerging or worsening drought conditions, providing insight into the key determinants of drought. The results demonstrate the importance of capturing spatiotemporal autocorrelation structures for accurate drought characterization and elucidates the drought time scales and thresholds that optimally separate each CDM severity category. In general, there is a positive relationship between the severity of drought and the time scale of the anomalies. However, high-severity droughts are also more complex and driven by a multitude of factors. It was found that satellite-based evaporative stress index (ESI), soil moisture, and groundwater were effective predictors of drought onset and intensification. Similarly, anomalous phases of large-scale atmosphere–ocean dynamics exhibit teleconnections with Prairie drought. Overall, this investigation provides a better understanding of the physical mechanisms responsible for drought in the Prairies, provides data-driven thresholds for estimating drought severity that could improve future drought assessments, and offers a set of early warning indicators that may be useful for drought adaptation and mitigation.

Significance Statement

This work is significant because it identifies drivers of drought onset and intensification in an agriculturally and economically important region of Canada. This information can be used in the future to improve early warning for adaptation and mitigation. It also uses state-of-the-art machine learning techniques to understand drought, including a novel approach called SHAP probability values to improve interpretability. This provides evidence that machine learning models are not black boxes and should be more widely considered for understanding drought and other hydrometeorological phenomena.

For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Jacob Mardian, jmardian@uoguelph.ca

1. Introduction

Drought is a recurrent and costly natural disaster with impacts on water resources, food security, ecosystem functioning, human health, and the global economy (Ding et al. 2011; Wilhite et al. 2007). Changes to the global climate have already impacted hydroclimatic extremes such as droughts and floods, with evidence that the amount of land experiencing severe drought has doubled since 1970 (Dai 2011). Drying trends are expected to continue in some regions in the twenty-first century, with many drought-prone areas experiencing even drier conditions under current climate projections (Cook et al. 2020). While many regions experience periodic droughts, the Canadian Prairies are particularly vulnerable due to its semiarid climate and economic dependence on rainfed agriculture. The region has experienced a series of widespread and high-severity droughts in recent years and water deficits are expected to worsen because of a warmer climate (Tam et al. 2019), so there is a need to quantify and improve the understanding of processes underlying drought conditions.

Drought is associated with several land–atmosphere feedbacks that exacerbate existing dry conditions and encourage their persistence. On finer scales, the soil moisture–temperature feedback occurs when low soil moisture decreases evapotranspiration (ET) through water constraints, which raises land surface temperatures and subsequently air temperatures, increases evaporative demand, and further depletes soil moisture (Seneviratne et al. 2010). Similarly, dry soils and high temperatures may increase ceiling heights and reduce cloud cover, further reducing precipitation (Roberts et al. 2006). Large-scale weather systems can move into the region and alleviate these symptoms, but synoptic scale investigations into Prairie drought have repeatedly found a high pressure ridging pattern that blocks moisture pathways from the Pacific Ocean and the Gulf of Mexico (Shabbar et al. 2011; Bonsal et al. 2011; Knox and Lawford 1990; Newton et al. 2014b).

Interannual hydroclimate variability in western Canada is also related to the phases of large-scale atmosphere–ocean teleconnections as these patterns influence moisture pathways and associated temperature anomalies. Asong et al. (2018) found a dominant periodicity of Prairie drought between 8 and 32 months that is associated with the Pacific–North America (PNA) pattern and the El Niño–Southern Oscillation (ENSO). Another study found an increasing influence of the Arctic Oscillation (AO) and Atlantic multidecadal oscillation (AMO) on Canadian drought since the 1980s (Yang et al. 2020). Several other papers have investigated the role of teleconnections in Prairie drought and water availability, noting influence from PNA, ENSO, AO, North Atlantic Oscillation (NAO), and Pacific decadal oscillation (PDO), among others (Jiang et al. 2014; Newton et al. 2014a,b; Bonsal and Shabbar 2008; O’Neil et al. 2017).

However, each drought event is unique in its spatiotemporal characteristics, severity of impacts, and its dynamical drivers (Bonsal et al. 2020). Several studies have investigated the characteristics of individual events. For example, the widespread and persistent drought of 2001/02 has been well documented due to its enormous impacts across the country, and particularly on the Prairies (Wheaton et al. 2008; Bonsal and Regier 2007; Hanesiak et al. 2011). An investigation into the U.S Great Plains drought of 2012 revealed several causes, including high pressure that reduced convection, cyclonic activity, and moisture transport from the Gulf of Mexico (Hoerling et al. 2014). The severe Canadian Prairie drought of 2015 has been partially attributed to anthropogenic climate change that increased late winter temperatures and reduced snowpack, reducing resiliency to dry conditions in the subsequent seasons (Szeto et al. 2016). These studies provide valuable information for understanding and predicting future droughts.

Recently, advances in explainable artificial intelligence (XAI) have enabled understanding of machine learning predictions that could be used for in-depth investigations of drought. Shapley additive explanations (SHAP) are one approach that reverse engineers the output of a machine learning model to explain why predictions are made, providing variable importance metrics for each observation. They are based on Shapley values, which quantify the contribution of each player in a cooperative game, considering some participants may have contributed more or less than others (Shapley 1953). In the context of machine learning, the game reproduces the outcome of the model, and the players are the covariates. In short, SHAP values quantify the marginal contribution of each covariate to the model [see Lundberg and Lee (2017) for the theoretical foundations]. Importantly, SHAP values assess local variable importance (i.e., for each prediction), allowing a more granular investigation into the inner mechanics of each observation. Dikshit and Pradhan (2021) applied SHAP for understanding spatial drought predictions in Australia, but limitations in their experimental design, such as using rainfall inputs to predict running averages of rainfall outputs and a small predictor set, hindered their ability to unveil mechanisms of drought beyond rainfall. Using a response variable that captures the comprehensive impacts of drought, such as the Canadian Drought Monitor (CDM), and a more varied set of predictors may provide more nuanced insights into drought mechanisms.

In this paper, we develop a machine learning model to nowcast drought in the Canadian Prairies from 2005 to 2019 and calculate SHAP values in probability units with a focus on identifying observations most representative of drought onset and intensification. This enables the identification of early warning indicators, drought thresholds, time scales of relevance, and relationships to teleconnections to identify the key determinants of Prairie drought.

2. Data and methodology

a. Data

A variety of climate, satellite, and ancillary data were retrieved, processed, and used as inputs to a machine learning model to predict drought severity and extent. These include predictors related to spatiotemporal autocorrelation, drought indicators that measure water balance anomalies, teleconnection indices describing large-scale ocean–atmosphere patterns, as well as soil moisture, evapotranspiration, water storage, and vegetation parameters obtained from satellite imagery (Table 1).

Table 1.

Summary table of the variables used in this study. Note that all teleconnection indices were calculated for the current month, with a 3-month lag, and with a 6-month lag to capture the lagged effect of large-scale circulation patterns of local climates.

Table 1.

1) Canadian Drought Monitor

The CDM measures drought on an ordinal scale from no drought (ND) to abnormally dry (D0), moderate drought (D1), severe drought (D2), extreme drought (D3), and exceptional drought (D4). These drought classes correspond to probability percentiles of drought occurrence, similar to the approach used by the United States Drought Monitor (USDM) (Svoboda et al. 2002). The extreme and exceptional drought categories (D3 and D4) were merged into one class (from now on D3/D4), as the low sample size of exceptional droughts presented issues for model training.

In this study, the CDM is the validation metric of drought severity and extent as it is a comprehensive drought assessment of all impacts following a robust analytical framework (Svoboda et al. 2002; Rippey et al. 2021). This contrasts with other common validation metrics of drought that only measure one type of drought impact, such as precipitation for meteorological drought monitoring or soil moisture for agricultural drought monitoring. The CDM assessment maps are produced by examining many data sources alongside expert input to reach a convergence of evidence on drought and its wide-ranging impacts. Based on this approach, the CDM assessment is accepted as the best available integrated drought assessment and provides a richer assessment than using a single indicator.

However, there is also a degree of subjectivity and opaqueness in the framework that hinders interpretability for scientists and decision-makers. For instance, it is unclear how expert input from different sectors (e.g., agriculture, water resources, forestry) is utilized, which data-driven indictors are used and how they vary over time, and how this information is weighted by federal, provincial, and academic scientists to produce the CDM. The purpose of using the CDM as the basis for this study is to understand the physical factors driving a wide range of drought impacts based on 20 years of expert analysis. Some of the indicators used in the model are used in the CDM (e.g., station-based drought indicators) while others are not commonly used or consulted at all (e.g., teleconnections, satellite datasets).

2) Spatial and temporal predictors

The CDM categories both vary spatially and temporally but the model used in this research is neither spatially nor temporally implicit, treating each observation as independent. Consequently, temporal (lagged) and spatial (location) predictors were used to capture serial autocorrelation and spatial autocorrelation, respectively. Primarily, the NDProp, D0Prop, D1Prop, D2Prop, and D3/D4Prop variables measure the proportion of each drought category that were observed in a pixel neighborhood the previous month. The pixel neighborhood is defined by the ecoregion it belongs to, which uses a surrounding area with similar landscape characteristics, including geomorphology, soils, vegetation, and climate (Ecological Stratification Working Group 1995). This provides the model with spatiotemporal information about the current distribution of drought that may be useful for predicting future changes. The spatial predictors included the mean latitude and longitude of each grid cell, which generally capture the moisture gradient in the Prairies from west (dry) to east (wet) and from south (dry) to north (wet).

3) Drought indicators

The standardized precipitation evapotranspiration index (SPEI) is a drought indicator that estimates water deficits and surpluses by measuring precipitation against potential evapotranspiration. The latter is estimated using the Thornthwaite method due to limited data availability. Positive values represent wet anomalies and negative values represent dry anomalies (Vicente-Serrano et al. 2010). In this study, SPEI was calculated from gap-filled monthly precipitation and temperature weather station data and spatially interpolated (AAFC 2017). It was calculated on time scales of 1, 3, 6, 9, 12, 24, and 36 months (from now on SPEI1, SPEI3, etc.) because the time scale over which precipitation is measured is an important determinant of its impacts, with shorter time scales corresponding to meteorological and agricultural droughts and longer time scales corresponding to hydrological and ecological droughts.

The Palmer drought severity index (PDSI) is another drought indicator that measures aridity by calculating water balance terms for a two-layer soil model with climatological inputs (Palmer 1965). The index is calculated on a similar scale to SPEI with negative values representing dry conditions and positive values representing wet conditions. We apply a modified version of the PDSI, called the Palmer drought index (PDI) that uses an adjusted water balance model from the Versatile Soil Moisture Budget (VSMB) model and an adjusted regional correction factor tuned for the Canadian Prairies (Akinremi et al. 1996).

4) Teleconnections

Several monthly teleconnection indices were obtained from NOAA’s National Climatic Data Center (NCDC) and Physical Sciences Laboratory (PSL) (NCDC 2021; PSL 2021a). Each teleconnection index is also lagged by 3 months and 6 months as these patterns may influence surface hydroclimate conditions in subsequent seasons.

The El Niño–Southern Oscillation (ENSO) is the dominant mode of global interannual climate variability defined by a periodic fluctuation in sea surface temperatures (SSTs) and air pressure in the equatorial Pacific Ocean every 2–7 years. La Niña triggers upper-atmosphere waves that are associated with a low pressure trough over western Canada that brings cold, wet air, increased snowpack and more cold spells to the region, while El Niño is associated with a high pressure ridge over the continent that blocks Arctic air and brings warm, dry air with less snowpack and less cold spells (Newton et al. 2014b,a; Bonsal and Shabbar 2011). However, the relationship between drought and ENSO phases is less clear, particularly during the summer. ENSO strength is measured using the multivariate ENSO index (MEI) version 2, the first mode of the combined empirical orthogonal function (EOF) of five relevant variables over the tropical Pacific: sea level pressure, SSTs, zonal winds, meridional winds, and outgoing longwave radiation (PSL 2021b).

The Arctic Oscillation (AO) is the dominant mode of interannual winter climate variability in the Northern Hemisphere, defined by the strength of circumpolar winds from the surface to the midstratosphere. The positive phase traps cold air in the Arctic and results in warmer weather in the midlatitudes, while the negative phase is a weakened and meandering polar jet stream that brings cold and dry air to areas of the midlatitudes (Wang and Chen 2010; Baldwin and Dunkerton 1999; Lawrence et al. 2020). AO strength is measured using the AO index, the first EOF mode of the monthly mean 1000-mb (1 mb = 1 hPa) anomaly data in the Northern Hemisphere (NCDC 2021).

The North Atlantic Oscillation (NAO) describes the sea level pressure difference between the Azores high and Icelandic low and therefore the strength and direction of westerly winds and storm tracks across the North Atlantic that directly influences the hydroclimate of eastern North America and Europe. The positive phase represents an increase in the pressure gradient with stronger winds and higher precipitation in the eastern North America and northern Europe, while the negative phase represents a weaker gradient with drier conditions across the Eastern seaboard of North America (Hurrell et al. 2003). NAO strength is measured using the NAO index is the first EOF mode of the monthly mean 500-mb anomaly data in the Northern Hemisphere (NCDC 2021).

The AMO is a low-frequency pattern in the North Atlantic SSTs that is thought to be related to slowly changing heat transport patterns from the Atlantic meridional overturning circulation (AMOC) and changes in radiative forcing (Trenary and DelSole 2016). The phases of AMO are related to regional hydroclimate variability in North America (Enfield et al. 2001). The AMO index is calculated using a weighted average of detrended SSTs in the North Atlantic (PSL 2021c). This research uses the unsmoothed version of the AMO index.

The PDO is a low-frequency pattern of SSTs in the North Pacific with a profound influence on North American climate. The positive phase is associated with warmer temperatures, lower precipitation, and drought conditions in Western North America on interannual to interdecadal time scales, while negative PDO phases are associated with the opposite response (Kerr et al. 2021). The PDO index used here is obtained by regressing the anomalies of NOAA’s extended reconstruction of SSTs (ERSST version 5) against the Mantua PDO Index during their overlap period and mapping the results (NCDC 2021).

The PNA pattern is one of the most influential climate patterns in North America during the winter, describing general atmospheric circulation over the North Pacific and North America (Leathers et al. 1991). The PNA index is calculated by projecting the PNA loading pattern (second EOF mode of monthly mean 500-mb height anomalies from 1950 to 2000 over 0°–90°N) to the daily anomaly 500-mb height anomalies over the same area (NCDC 2021).

5) Soil moisture

The European Space Agency Climate Change Initiative (ESA CCI) long-term surface soil moisture product blends active and passive satellite observations for long-term climate change monitoring (Dorigo et al. 2017). This enables the use of a single calibrated soil moisture product for the entire study period. Each monthly value was then converted to a percent of average scale to better represent departures from normal conditions ideal for drought monitoring. From this product, moving averages were calculated on 1-, 2-, 4-, and 8-month time scales. However, there are no soil moisture observations from December to March due to snow cover, so the moving averages were calculated based on the last valid months. For example, a 2-month moving average (SM2) in April consists of the soil moisture values from April and the preceding November, with the hypothesis that there is a high correlation between soil moisture after winter freeze-up and before spring thaw (Yang and Wang 2019).

6) Evaporative stress

The evaporative stress index (ESI) calculates standardized anomalies of the evapotranspiration to potential evapotranspiration (ET:PET) ratio by driving a land surface model with an atmospheric boundary layer in time-differencing mode using satellite-based land surface temperatures (LSTs) (Anderson et al. 2011, 2016). The result shows differences in water usage rates that may be useful for monitoring drought, including flash droughts with rapid onset and intensification (Otkin et al. 2013). In this paper, monthly averages were computed from weekly data. Given the ability of ESI to monitor rapid onset and termination of drought conditions, the slope of the weekly values was calculated for each month using simple linear regression to monitor the rate of change. The full methodology used for computation of ESI is outlined in Anderson et al. (2011).

7) Groundwater

Gravity Recovery and Climate Experiment (GRACE) uses a pair of satellites to measure variations in Earth’s gravity field over time, with repeat passes roughly every 30 days. These changes in gravity are caused by redistributions of mass in the oceans and the atmosphere, as well as water, snow, and ice over land. Changes in water storage can be estimated by observing mass changes after controlling for ocean and atmospheric contributions (Wahr et al. 2004). This product is an effective indicator of long-term hydrological droughts, as the total water storage (TWS) product is capable of detecting the location of water after precipitation has been redistributed to various aquifers (Cammalleri et al. 2019). It also has the potential of monitoring changes in water availability caused by factors other than precipitation, such as groundwater irrigation. The TWS observations are then used to drive the GRACE Data Assimilation System (DAS), based on the Catchment Land Surface Model (CLSM), to derive weekly groundwater storage percentiles for hydrological drought monitoring (Houborg et al. 2012). This study uses the weekly groundwater percentiles to compute groundwater indices on time scales of 1, 3, 6, 9, and 12 months (e.g., GRACE6). Similar to ESI, the slope was calculated for each month by running a simple linear regression model on the weekly values to evaluate rapid changes in water storage.

8) Vegetation

During periods of drought, vegetation stress can be detected remotely using normalized difference vegetation index (NDVI), a measure of vegetation health or “greenness” on a scale from −1 to 1. Weekly maximum MODIS NDVI composites from the Canadian Ag-Land Monitoring System (CALMS) produced by Agriculture and Agri-Food Canada. This 230-m resolution product was then upscaled to the modeling grid of 5 km through spatial averaging, retaining both the mean (meanNDVI) and standard deviation (sdNDVI) statistics. These values were then averaged monthly capture the mean and variability of vegetation dynamics within each grid cell that month. As with ESI and GRACE, the slopes of weekly values were calculated to monitor the rate of change in the mean (meanNDVI_Slope) and variability (sdNDVI_Slope) of vegetation within the grid cell that month.

b. Methodology

1) Study area

The Canadian Prairies is a socially, economically, and agriculturally important region of Canada (Fig. 1). The area is defined by its flat terrain and semiarid climate primarily dependent on rainfed agriculture that makes it particularly vulnerable to drought impacts. In this study, the entire agricultural extent of the Prairies was the study area. This area is about 50% agriculture with the other 50% being a combination of natural grasslands, forests, wetlands, and urban areas. The study period of 2005–19 was chosen based on data availability.

Fig. 1.
Fig. 1.

The Canadian Prairies study area with land cover (AAFC 2015).

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

2) Machine learning model

The XGBoost algorithm is an ensemble technique that sequentially builds decision trees, with each tree predicting the residuals of the previous tree until no further improvements can be made, using majority voting to obtain the final prediction (Fig. 2). It is widely recognized as a scalable, efficient, and accurate machine learning algorithm (Chen and Guestrin 2016).

Fig. 2.
Fig. 2.

Simplified diagram of the XGBoost framework for classification. The first DT is naïve, and each subsequent DT is trained on the residuals from the DT before it. The models typically have several hundred trees.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

In this paper, each dataset was resampled to a 5-km modeling grid covering the spatial domain for each month from 2005 to 2019. Each grid cell has a corresponding monthly CDM category used as the response variable that all covariates were trained on in a nowcasting approach. The XGBoost model is evaluated using a time series cross validation technique that replicates real-time drought monitoring, sequentially predicting each month from 2005 to 2019 using all historical data for training. This model is repeated many times with different parameter combinations, and the model with the closest agreement to the CDM based on the weighted Cohen’s Kappa is selected (Cohen 1968).

3) Explainable AI

While machine learning ensemble approaches like XGBoost improve model accuracy, they simultaneously decrease interpretability. These more advanced machine learning models are often referred to as “black box” methods because users can observe inputs and outputs but have no knowledge of their internal mechanisms. This trade-off between accuracy and interpretability is a significant challenge for scientific inquiry, as understanding how and why an accurate model makes predictions is critical for understanding the relevant processes driving each prediction. Here, understanding the unique contribution of each covariate to the model may be useful for understanding the key determinants of drought.

These techniques differ from traditional statistical approaches to variable importance, such as principal component analysis (PCA) and analysis of variance (ANOVA). They offer insights into the complex and nonlinear covariance found in big data, which enable machine learning models to perform well but may not be apparent using traditional methods. Furthermore, traditional approaches often rely on assumptions about the data, such as independence, homoscedasticity, and linearity. However, these assumptions are violated due to spatial and temporal autocorrelation, numerous nonstandardized predictors, and the nonlinearity of XGBoost, respectively.

In this study, we employed the interventional SHAP approach (Janzing et al. 2019; Lundberg et al. 2020) to approximate feature contributions to drought predictions using a model retrained on the entire time series and calculating SHAP values on a pixelwise basis. Unlike the conditional SHAP method, which respects correlations between features, interventional SHAP establishes statistical independence by breaking feature dependencies with a background dataset, capturing the true causal effect of each feature. This statistical independence is evidenced in Chen et al. (2020, Fig. 2), showing that interventional SHAP assigns zero importance to dummy variables—those that have no effect but are correlated to other variables. This contrasts with the conditional approach where the presence of a dummy variable significantly altered the variable importance analysis. As a result, the interventional SHAP method accurately attributes feature contributions, independent of other features, demonstrating robustness against the choice of predictor sets. See Laberge and Pequignot (2022) for a comprehensive exploration of this technique.

Next, each SHAP value, is converted from log odds to probabilities using the following approach:

  • Step 1. Obtain the baseline probability for class k, pbase(k), using the shap package in python. This represents the probability of any random observation belonging to that class before incorporating the impact of feature values.

  • Step 2. Convert the baseline probability into odds: obase(k)=pbase(k)/1pbase(k).

  • Step 3. Convert the SHAP value ϕ from log odds to odds o(ϕ) = eϕ, representing the change in odds of observing class k after incorporating the impact of the feature value.

  • Step 4. Calculate the predicted odds of observing class k after incorporating the impact of the feature value: opred(k)=obase(k)o(ϕ).

  • Step 5. Convert step 4 from odds to probability, representing the predicted probability of observing class k after incorporating the impact of the feature value: ppred(k)=opred(k)/[1+opred(k)].

  • Step 6. Finally, obtain the SHAP value in probability units, estimating the change in probability of observing class k after incorporating the impact of the feature value: p(ϕ) = ppred(k) − pbase(k).

These transformed SHAP values are called SHAP probability values. First, SHAP probability values are presented for single pixel to provide an intuitive example for understanding predictions. Next, they are pooled across all grid cells and time steps to present average variable importance metrics for each drought class, therefore addressing the processes responsible for different drought severities. Finally, the observations are filtered to include only dates representative of drought onset/intensification to identify which variables are effective for early warning of worsening drought status.

3. Results and discussion

a. Variable importance for an individual observation

The results for a randomly selected D3/D4 grid cell in July of 2015 are shown in Fig. 3 for demonstration. The baseline probability of a D3/D4 drought is 1.46%, based on its frequency in the training data while the predicted probability for this grid cell is 48%. This difference is driven by the values of the model predictors. Long-term groundwater deficits (GRACE9 < 13th percentile), surface water balance (PDI < −2.5), and precipitation (SPEI12 < −2) substantially increased the probability of a D3/D4, along with the presence of a strong El Niño event (MEI = 1.73) and an above average proportion of D3/D4 drought the previous month (D3/D4Prop = 0.09). Predictors that decreased the probability of D3/D4 drought include long-term surface soil moisture values only slightly below average (SM8 > 92% of average) and the longitude of the pixel, indicating D3/D4 drought did not occur often in the Peace River region of Alberta, Canada during the study period.

Fig. 3.
Fig. 3.

Force plot showing the most impactful predictors of a D3/D4 drought observation of the Peace River region in July 2015. The baseline probability (base value) of a D3/D4 drought is 0.0146, while the predicted probability for this observation is 0.48. The marginal contribution of each variable to this prediction is demonstrated by the size of the bars. Variables in orange increase the probability of a D3/D4 drought, while variables in red decrease the probability. The original feature values are displayed below the bars.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

b. Average variable importance for the study period

The SHAP probability values calculated for each pixel are pooled across all grid cells and time steps to present average variable importance metrics for each drought class to gain insight into the drivers of Prairie drought from 2005 to 2019 (Fig. 4). The most important variables for drought prediction are the spatiotemporal autocorrelation predictors (e.g., NDProp, D0Prop) that provide the model with relevant information of current drought status. More specifically, it contains the proportion of surrounding pixels currently designated under each drought category. These plots demonstrate the value of including spatiotemporal information in a drought prediction model. Drought is an inherently persistent and slow-moving hazard, so capturing this evolution in both space and time is key to anticipating future changes to drought status. In general, higher proportions of a drought category substantially increase the probability of observing that drought class in the future.

Fig. 4.
Fig. 4.

Facet barplot showing average variable importance of Prairie drought from 2005 to 2019 across each drought class. Lagged teleconnection variables are averaged to improve illustration (e.g., AO, AO.3MonthLag, AO.6MonthLag− > AO).

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

Other important variables include PDI, SPEI3, and SPEI6, suggesting that the CDM is most correlated to meteorological drought on a 3–6-month time scale. This is roughly the time scale of 1–2 seasons (3–6 months) and an agricultural growing season (about 4 months in the Prairies). For ND, D0, and D1 categories, SPEI3 tends to be the dominant predictor, while SPEI6 is the dominant predictor for D2 drought, indicating that severe droughts generally accumulate over longer time scales than moderate drought. The high importance of PDI suggests added value using a surface water balance approach over meteorological deficits alone. This is because PDI estimates moisture anomalies by considering the effects of antecedent precipitation and the subsequent response of different land surface characteristics (Shabbar and Skinner 2004; Liu et al. 2015). It is also highly correlated to soil moisture, enabling its use as both a meteorological and agricultural drought indicator (Szép et al. 2005).

Finally, another pattern seen is that ND and D0 conditions are primarily predicted using drought indicators and the autocorrelation predictors. As drought severity increases, so does the complexity of prediction as shown by the higher importance of other variables. For instance, D3/D4 drought is characterized by surface and subsurface impacts after moisture deficits propagate through the hydrological cycle. The more severe a drought, the more it is associated with soil moisture and groundwater storage.

c. Drought indicator thresholds

The CDM categories can be more effectively defined by determining empirically driven thresholds. Each category is broadly defined by a return period, but return periods are challenging to identify under nonstationary climate conditions and their relationship to drought indicators is even less clear. Table 2 shows the value of each drought indicator that maximizes the probability of observing each CDM category. The purpose is to clarify the range of moisture index anomalies associated with each category. The SPEI1, SPEI24, and SPEI36 indicators were excluded because no clear thresholds were found, suggesting that the CDM is not strongly related to water balance deficits on these time scales.

Table 2.

Values that maximize probability of a drought class label for selected drought indicators and GRACE groundwater percentiles. “NA” indicates there is no clear maximum.

Table 2.

For SPEI3, SPEI6, SPEI9, and SPEI12, ND is strongly defined by values above zero, meaning that moisture availability is normal or wetter than normal. The probability of D0 conditions is maximized by weakly negative anomalies in the range from 0 to −0.5 depending on the time scale of interest. D1 conditions occur with values around −1, except for SPEI3 where stronger anomalies of −1.5 maximize probability. This means that D1 conditions are defined by moderate dryness, with an emphasis on worsening conditions in the previous 3 months. Finally, D2 and D3/D4 droughts both occur with values less than −2 with no clear threshold for separation. The exception is SPEI9, where the threshold for D2 is −1.5, providing separation with the D3/D4 category. These values are generally in agreement with the value ranges given in Svoboda et al. (2002) to define the USDM categories, but with some notable differences such as lower thresholds for D0 and higher thresholds for D2. However, given the low sample size of D3/D4 drought in this study, it is difficult to provide clear separation with D2 conditions. Instead, information outside of SPEI and PDI are needed to separate D2 (return period of 1 in 10 years) from D3 (1 in 20 years) and D4 (1 in 50 years), such as the land surface and groundwater response to these anomalies.

The PDI indicator tells a different story. ND is defined by values above 2.5 and the probability of D0 is maximized by values of 1. This suggests that the PDI indicator is biased with a mean above zero. The drought categories of D1, D2, and D3/D4 are characterized by negative PDI values of −1, −2.5, and −2.5, respectively. These values are not in agreement with the value ranges given in Svoboda et al. (2002) to define the USDM categories, likely due to differences in PDI calculation compared with PDSI. For instance, a different water balance model is used which may alter ET and runoff estimates, while available water capacity estimates of soils may contribute to biases.

Finally, the GRACE1 and GRACE3 groundwater percentiles were included to evaluate the utility of this data for estimating drought in the Prairies and to test the hypothesis that D2 and D3/D4 drought require additional information beyond meteorological drought for accurate classification. The results are mixed. ND occurs with groundwater percentiles above 80 and 85 on 1- and 3-month time scales, but D0 conditions are not clearly related to groundwater percentiles. D1 conditions are marginally separated from D2 and D3/D4 droughts on 1-month time scales as the probability of D1 conditions are maximized for groundwater percentiles of 40, while any values below 40 coincide with both D2 and D3/D4 drought and cannot be distinguished. The 3-month groundwater percentiles more effectively separate the different drought severities with some overlap, with D1 corresponding to values of 50, D2 occurring with any values below 30, and D3/D4 occurring with any values below 20.

Overall, these drought indicator thresholds demonstrate that CDM categories are generally correlated to meteorological water balance anomalies on time scales ranging from 3 months to 1 year. The average SHAP probability values from the model show a peak at 3 months for ND, D0, and D1 conditions and a peak at 6 months for more severe droughts (Fig. 4). However, the drought thresholds indicate that the D2 and D3/D4 categories can be most effectively separated using SPEI9, but that using additional information such as GRACE groundwater percentiles may be necessary for differentiating the most severe drought categories.

d. Variable importance for drought onset events

SHAP probability values were filtered to isolate months where drought onset or intensification occurred. This filter was defined as pixels where the was an increase in its drought severity category from the previous month and no other pixels in the same ecoregion were of equal or higher severity in the previous month (i.e., drought of the same severity was not already present within the same ecoregion). For example, if a pixel changes from D0 to D1 and no pixels in the ecoregion were already D1, then it is included as onset. If the pixel changes from D0 to D1 but some pixels in the ecoregion were already D1, they are not included. Drought onset was defined this way to identify the climatological drivers of drought onset in the region as opposed to spatial expansion of existing drought. Sina plots are presented in Figs. 58 to display the relationship between feature values and variable importance for onset of D0, D1, D2, and D3/D4 drought, respectively. While the variables are ordered by average importance, many less important variables have substantial impacts on drought formation throughout the study period. This can be seen by individual points of the sina plots, which represent the impact of the variable for a single observation.

Fig. 5.
Fig. 5.

Sina plot demonstrating variable importance for D0 onset events. Ordered by average absolute SHAP probability value (displayed on the left). Positive SHAP probability values indicate the observation increased the probability of D0 onset, while negative SHAP probability values decrease the probability. Color of points correspond to the original feature value on the scale of the full dataset.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

Fig. 6.
Fig. 6.

Sina plot for D1 onset.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

Fig. 7.
Fig. 7.

Sina plot for D2 onset.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

Fig. 8.
Fig. 8.

Sina plot for D3/D4 onset.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

1) Drought indicators

In general, the relationship between the dominant SPEI time scales and the severity of drought are positively related, while the values between SPEI values and drought categories have a negative relationship.

Transitions from ND to D0 status are driven by changes in short-term precipitation and temperature anomalies as shown by the highest importance of SPEI1 and SPEI3 (Fig. 5). D0 is a CDM category defined by short-term dryness and transitions between the presence and absence of drought, so onset occurs quickly (Wood et al. 2015). As expected, reductions in SPEI1 and SPEI3 increase the probability of D0. However, note that the only moderate values drive this category (shown in orange), as the lowest values of SPEI (that would be shown in purple) and the highest values of SPEI (that would be shown in yellow) never occurred during D0 onset events as these would likely belong to a drought category (D1+) or ND, respectively.

The onset of D1 drought is associated with slightly longer time scales, with the highest variable importance belonging to SPEI3, SPEI6, and PDI (Fig. 6). The higher importance of PDI for D1 onset compared to D0 suggests that D0 onset is primarily driven only by reductions in atmospheric moisture availability, while the surface water balance is more important to the prediction of drought onset. Intensification to D2 status is driven by even longer time scales than D1, with SPEI6 being the most dominant followed by SPEI3 and SPEI12 among the top predictors (Fig. 7). Finally, transitions to D3/D4 status are most associated with the SPEI3 and SPEI9 drought indicators, although many of the most important variables are not drought indicators as other information is required to predict the most extreme drought events (Fig. 8). Interestingly, only moderate values of SPEI and PDI can predict drought onset and intensification. This is because the lowest values of SPEI and PDI do not occur during onset and intensification of any drought category, but during sustained droughts.

2) Satellite groundwater percentiles

Short-term anomalies in groundwater (GRACE1) were a key predictor of D0 onset in some cases, increasing the probability of D0 onset by up to 8% when values were low. There is no clear relationship between D0 onset and longer-term trends in groundwater as abnormally dry conditions are short-term moisture anomalies. GRACE1 is even more important for predicting the onset of drought categories from D1 to D4. For instance, moderate groundwater percentiles increase the probability of D1 onset, while low values increase the probability of D2 and D3/D4 onset. Low values of GRACE3 are also an important predictor of D2 and D3/D4 drought, demonstrating the large impacts of high severity droughts on long-term groundwater levels.

The GRACE1_Slope indicator measures the linear trend of weekly groundwater percentiles over the previous month, and although its average importance was low, it demonstrated value for predicting drought onset in some flash drought cases. It had some of the highest SHAP values of any variable for the D1 and D2 categories, reaching over 10% and 15%, respectively. Moderate weekly decreases in groundwater led to D1 onset, while D2 onset was associated with sharp decreases in groundwater within the month.

Overall, these results show that GRACE is able to capture a range of changes in groundwater conditions that agree with the CDM, including nuanced monthly changes associated with D0 conditions, rapid weekly changes associated with flash droughts, and extreme reductions in groundwater levels over several months that are associated with higher severity drought. This contrasts with a previous investigation that found GRACE groundwater retrievals may not be useful on time scales less than 3 months (Thomas et al. 2017). This demonstrates the effectiveness of the high-resolution and model-driven GRACE DAS data for drought monitoring. The high impact of GRACE1_Slope means that it can also be used as an early warning indicator of drought onset and intensification by utilizing its high temporal resolution to detect rapid changes. These data were useful for identifying drought because they capture the response of surface and subsurface processes to meteorological conditions, including human impacts such as irrigation, therefore providing added value over precipitation anomalies alone.

3) Satellite soil moisture percentiles

Satellite soil moisture percentiles were not among the most important variables for predicting D0 onset, but in some cases SM1 proved to be critically important when dry soils increased the probability of D0 onset by up to 8%. They were not relevant for predicting D1 or D2 onset. This is likely because these data are satellite derived surface soil moisture and not root zone soil moisture that is needed to measure the effects of water deficits on vegetation. While these components of the soil matrix are correlated, surface soil moisture exhibits higher temporal variability due to land–atmosphere interactions that drive wetting and drying events (Champagne et al. 2018). However, SM2 and to a lesser extent SM4 were important predictors of D3/D4 drought onset. This is because the largest soil moisture anomalies on the seasonal time scale indicate persistently dry surface soil moisture conditions. If there is no moisture supply to the surface for long periods of time, the root zone will not be replenished, the two layers will be more closely correlated, and the soil moisture percentiles will capture reductions in vegetation growth. For example, very low values of SM2 indicate parched soils for two months that create a hostile growing environment and reduce surface water availability to drive hydroclimate feedbacks. These findings are consistent with other studies in the Prairies that found that the ESA CCI soil moisture dataset is useful for capturing dry soil moisture extremes (Champagne et al. 2019; Oozeer et al. 2020).

4) Evaporative stress index

ESI is an index that describes temporal anomalies in evapotranspiration on the land surface and has demonstrated the ability to provide early warning of drought, and particularly flash droughts characterized by rapid onset due to a lack of precipitation and high temperatures (Anderson et al. 2013; Otkin et al. 2018; Parker et al. 2021). In this study, large ESI anomalies are useful in predicting D1, D2, and D3/D4 drought onset (Fig. 9). In fact, all negative values are associated with each category, with higher magnitude anomalies generally increasing the probability of onset. In some cases, ESI anomalies increased the probability of observing these D1 onset by up to 5%, D2 onset by 3%, and D3/D4 onset by 4% (Figs. 68). While the ESI has been evaluated for drought detection in other geographic regions, including the United States (Anderson et al. 2011), Brazil (Anderson et al. 2016), and Australia (Nguyen et al. 2019), these results suggest applicability to the Canadian Prairies as well. This is an important demonstration as regional differences in land surface heterogeneity, climate characteristics, and satellite image quality and availability could potentially influence model calibration and parameterization.

Fig. 9.
Fig. 9.

Facet boxplot showing the change in probability of observing each category due to ESI.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

One limitation is that ESI was aggregated to the monthly time scale to match the time step of the model and the CDM, which reduces the effectiveness of ESI as a fast response indicator. Utilizing weekly ESI values may improve results of drought onset prediction further.

5) Teleconnections

Overall, the SHAP probability values were lower for teleconnections than other types of variables. However, the total contribution of teleconnection indices to the model predictions are not representative of how important these processes are in driving drought for three main reasons. First, most of the variability forced by these large-scale processes are captured by the drought indicators (i.e., SPEI, PDI), so the SHAP probability values shown only represent their unique contributions. Second, teleconnection indices only have a strong impact on Prairie climate under certain conditions. Neutral conditions tend to have a low impact on climate conditions, while extremes have a much larger impact that is not well represented using averages. Similarly, the strongest teleconnection signals in the Prairies occur in the cold season. While this can help initiate drought through low snowpack that reduces soil moisture in subsequent seasons, moisture conditions during the warm season may help sustain or intensify a drought but have a weak relationship with teleconnections. Finally, the interaction between different teleconnections can amplify or dampen their response, and their interactions are challenging to quantify due to their different time scales and irregular cycles. In this study, the relationship between teleconnections and drought is understood by examining the relative change in SHAP probability values for different intensities of teleconnection indices, especially during negative (−) and positive (+) phases.

(i) El Niño–Southern Oscillation

MEI.Lag3− or neutral conditions paired with MEI+ values increase the probability of D1 onset, signaling that the beginning of El Niño events are associated with drought onset (Fig. 6). Otherwise, the lagged MEI variables had little effect on drought onset while current MEI conditions were important predictors, suggesting a fast response of Prairie hydroclimate conditions to the tropical Pacific forcing. The relationship is clear: drought onset of all categories is more likely under El Niño conditions (MEI > 1). However, La Niña events (MEI < 1) can also trigger dry conditions (D0) and moderate droughts (D1), but not the higher severity CDM categories (Fig. 10).

Fig. 10.
Fig. 10.

Facet boxplot showing the change in probability of observing each category due to MEI (ENSO strength). Note the blank bins for D2 meaning D2 onset never occurred with MEI values in those ranges.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

Strong ENSO events are known to impact precipitation, temperature, and surface water regimes in the Prairies (Basu et al. 2020; Bonsal and Shabbar 2008), and drought has been associated with both El Niño and La Niña events. For instance, Li et al. (2018) found that the extreme 2015 drought was driven by a strong El Niño in combination with the Madden–Julian oscillation (MJO). Higher SSTs and more active convection in the central Pacific Ocean (140°E) triggered Rossby wave trains that propagated to the Prairies and created a high pressure ridge in the upper troposphere that favors dry conditions. Conversely, a strong La Niña event was associated with the onset of the 1999–2005 Prairie drought, although it was not sustained during the entire period with neutral and El Niño conditions emerging later. These investigations are consistent with our findings that both extremes in ENSO conditions are associated with drought onset in the Prairies, but El Niño is generally associated with onset and intensification of the most extreme droughts. However, Bonsal and Lawford (1999) found that dry spells tend to occur the second summer following an El Niño event during the 1948–91 study period, while this study suggests a much faster response to El Niño forcing from 2005 to 2019.

(ii) Pacific–North American pattern

PNA− and PNA.Lag6− are important for predicting D0 onset (Fig. 6), while PNA+ events, both current and lagged, are associated with D1–D4 onset (Figs. 7 and 8). This agrees with the conventional understanding of PNA impacts on the Prairies. PNA+ is associated with an upper-level ridging pattern that drives warm and dry conditions in western Canada and may provide optimal conditions for drought onset, while the PNA− is associated with an upper-level trough that drives cool and wet conditions (Franzke et al. 2011). More specifically, the ridging pattern associated with PNA+ blocks outbreaks of warm, moist air from the North Pacific, which instead has a stronger meridional component that steers the air mass toward Alaska. This pattern is far more apparent during the cold season that the warm season (Newton et al. 2014b,a).

It is important to note that PNA is a subseasonal to seasonal pattern that is moderately predictable as it is linked to tropical processes. For example, investigations have demonstrated that PNA is modulated by the MJO’s 30–60-day cycle (Johnson and Feldstein 2010; Seo and Lee 2017; Mori and Watanabe 2008), by the interannual variability of ENSO (Soulard et al. 2019; Li et al. 2019), and by PDO (Ge and Luo 2023; Yu and Zwiers 2007). For instance, PNA+ is more frequent during El Niño and PDO+ episodes and vice versa. As a result, while the PNA events are short-lived, the relative frequency of these events is modulated by ENSO and PDO’s interannual to interdecadal patterns with implications for Prairie surface weather on longer time scales. In addition, the relationships between teleconnections and drought demonstrated cannot be understood in isolation due to the close links between different atmosphere–ocean patterns.

(iii) Atlantic multidecadal oscillation

The AMO index for the current month, 3-month lag, and 6-month lag were all important predictors of D0 onset (Fig. 5). AMO.Lag3− and AMO.Lag6− paired with AMO+ appear to drive D0 onset, indicating dry conditions are associated with transitions from cool to warm North Atlantic SSTs. AMO is also a driver of D1–D4 drought onset, with the lagged AMO− phase increasing the probability of all drought categories and the AMO+ phase having little effect (Fig. 11), suggesting a low-frequency drought forcing from cool Atlantic SSTs. This relationship is in contrast to other investigations that find that the warm phase of AMO is associated with drought in the Prairies and the U.S. Midwest (Shabbar and Skinner 2004; Enfield et al. 2001). While there was substantial interannual variation in AMO values (range from −0.1 to 0.6), the entire study period fell within an AMO warm phase, so the results should be interpreted with caution. It can only be said that during the positive AMO phase from 2005 to 2019, temporary deviations to negative AMO values coincided with drought events on the Prairies.

Fig. 11.
Fig. 11.

Facet boxplot showing the change in probability of observing each category due to 3-month lagged AMO index.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

(iv) North Atlantic Oscillation and Arctic Oscillation

NAO and the onset of CDM categories do not have a clear pattern, with a few exceptions. Strong NAO.Lag3 + events (NAO > 2) are related to D1 onset (Fig. 6), while strong NAO.Lag6 + events (NAO > 2) are related to D3/D4 drought (Fig. 8). This provides evidence that strong NAO+ phases can result in drought onset, and when prolonged may lead to intensification. However, in many cases, strong NAO− phases (NAO < 2) are also drivers of D2 and D3/D4 drought (Figs. 7 and 8), suggesting that drought onset may occur during either extreme of NAO.

Previous investigations have found an ambiguous link between NAO and Prairie drought. Bonsal and Shabbar (2008) found a weak but significant correlation between winter NAO and precipitation in the southern Prairies, with NAO+ decreasing precipitation and vice versa. On the contrary, Knox and Lawford (1990) found that spring NAO+ is associated with increased precipitation in the southern Prairies. Chartrand and Pausata (2020) provide a dynamical explanation for this, with a much higher incidence of cyclogenesis on the lee side of the Alberta Rocky Mountains and increased storm track density during winter NAO+ phases. This may increase the frequency of Alberta clipper storm systems and subsequently increase moisture availability in the Prairies. They also found a slight decrease in cyclogenesis and storm track density during NAO−, although the association is weaker.

The AO index was a poor predictor of drought onset for all categories, with low impact and mixed effects. This may be due to the high correlation between AO and NAO that reduces or eliminates the unique contributions of AO to Prairie drought.

(v) Pacific decadal oscillation

The PDO variables showed low importance on drought onset overall, despite a wide range of values ranging from less than −3 to +2 throughout the study period. However, when including all observations (i.e., not restricted to onset events only) it was a much more important predictor (not shown). This suggests its low-frequency variability does not provide the same early warning benefits as other teleconnections, despite its importance in driving Prairie hydroclimate. The strongest effect occurred with a 6-month lag, where PDO+ phases had a small effect on increasing the probability of D1–D4 drought onset (Fig. 12).

Fig. 12.
Fig. 12.

Facet boxplot showing the change in probability of observing each category due to the 6-month lagged PDO.

Citation: Journal of Hydrometeorology 24, 11; 10.1175/JHM-D-23-0036.1

4. Conclusions

A machine learning model was developed to nowcast drought in the Canadian Prairies and SHAP probability values were interpreted to understand the key determinants of drought severity and extent from 2005 to 2019. The use of the CDM as the validation metric enables a unique understanding of expert drought interpretation and its wide-ranging impacts (e.g., meteorological, agricultural, hydrological) while other studies typically measure one of these impacts. The results showed that a variety of data are required for accurate nowcasting, with unique contributions from spatiotemporal autocorrelation predictors, drought indicators derived from climate data, teleconnection indices, and satellite derived parameters related to soil moisture, evapotranspiration, and groundwater.

The most important variables for nowcasting drought were spatiotemporal autocorrelation structures because drought is a slow-moving and persistent phenomenon. Drought indicators were also vital for accurately classifying drought into the different CDM severity categories. In general, there is a positive relationship between the time scale of the precipitation anomaly and the severity of drought. Abnormally dry conditions are best predicted using short-term anomalies from 1 to 3 months, while the most extreme categories are strongly associated with time scales ranging from 6 to 12 months. An analysis of drought indicator thresholds clearly identifies separation between categories using SPEI indicators, but only SPEI9 could differentiate D2 and D3/D4 drought. These high-severity droughts are more complex and driven by numerous factors as they are characterized not only by precipitation anomalies, but the response of the surface and subsurface to those deficits, requiring additional parameters to effectively classify.

An analysis of drought onset revealed the mechanisms most effective in providing early warning of emerging or worsening drought conditions. GRACE groundwater percentiles averaged over one to three months and satellite soil moisture averaged over two months were able to track drought onset and intensification via reductions in soil moisture and groundwater availability, particularly for higher-severity droughts. In some cases, the slope of weekly GRACE groundwater percentiles and ESI anomalies were able to improve predictions of drought onset and intensification due to rapid changes in moisture regimes. The former tracks rapid reduction in groundwater on a weekly time scale, while the latter tracks the onset of vegetation stress due to inadequate soil moisture.

Finally, several teleconnection indices showed strong relationships between atmosphere–ocean forcings and Prairie drought. Most notably, both extremes of ENSO were associated with drought, as La Niña events increased the probability of moderate (D1) drought onset and El Niño increased the probability of drought onset for all severities (D1–D4). The positive phase of the PNA pattern was strongly associated with D1–D4 drought onset due to a high pressure ridging pattern that blocks moisture pathways into the Prairies. Weak but inconclusive relationships were also found between AMO, NAO, and PDO and drought onset, while no relationship was identified with AO.

Overall, this study provides insight into drought mechanisms and predictability in an important agricultural, social, and economic region of Canada. Crucially, the variable importance analysis was conducted for drought onset and intensification events in probability units, providing an intuitive explanation for anticipating future changes to drought status. These early warning indicators may be useful for agricultural and water management through the introduction of adaptation and mitigation measures including the planning of crop selections, irrigation schedules, and water resource allocation. Future research can use this approach to understand the key determinants of drought improvement and termination, as well as how variable importance varies spatially and temporally, such as across seasons and years.

Acknowledgments.

This research was funded in part by Agriculture and Agri-Food Canada, the Natural Sciences and Engineering Research Council of Canada (NSERC), and the Canada First Research Excellence Fund: Food from Thought. The authors wish to acknowledge Trevor Hadwen, Richard Warren, and Tyler Black for assistance with data preparation.

Data availability statement.

Data analyzed in this study were a reanalysis of existing data, which are openly available at locations cited in the reference section.

REFERENCES

  • AAFC, 2015: Annual crop inventory. Government of Canada, accessed 20 July 2023, https://open.canada.ca/data/en/dataset/ba2645d5-4458-414d-b196-6303ac06c1c9.

  • AAFC, 2017: Standardized precipitation evapotranspiration index. Government of Canada, accessed 20 July 2023, https://open.canada.ca/data/en/dataset/d765cc41-8ee0-4aca-be4b-084448e52a58.

  • Akinremi, O. O., S. M. McGinn, and A. G. Barr, 1996: Evaluation of the Palmer drought index on the Canadian Prairies. J. Climate, 9, 897905, https://doi.org/10.1175/1520-0442(1996)009<0897:EOTPDI>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Anderson, M. C., and Coauthors, 2011: Mapping daily evapotranspiration at field to continental scales using geostationary and polar orbiting satellite imagery. Hydrol. Earth Syst. Sci., 15, 223239, https://doi.org/10.5194/hess-15-223-2011.

    • Search Google Scholar
    • Export Citation
  • Anderson, M. C., C. Hain, J. Otkin, X. Zhan, K. Mo, M. Svoboda, B. Wardlow, and A. Pimstein, 2013: An intercomparison of drought indicators based on thermal remote sensing and NLDAS-2 simulations with U.S. Drought Monitor classifications. J. Hydrometeor., 14, 10351056, https://doi.org/10.1175/JHM-D-12-0140.1.

    • Search Google Scholar
    • Export Citation
  • Anderson, M. C., and Coauthors, 2016: The evaporative stress index as an indicator of agricultural drought in Brazil: An assessment based on crop yield impacts. Remote Sens. Environ., 174, 8299, https://doi.org/10.1016/j.rse.2015.11.034.

    • Search Google Scholar
    • Export Citation
  • Asong, Z. E., H. S. Wheater, B. Bonsal, S. Razavi, and S. Kurkute, 2018: Historical drought patterns over Canada and their teleconnections with large-scale climate signals. Hydrol. Earth Syst. Sci., 22, 31053124, https://doi.org/10.5194/hess-22-3105-2018.

    • Search Google Scholar
    • Export Citation
  • Baldwin, M. P., and T. J. Dunkerton, 1999: Propagation of the Arctic Oscillation from the stratosphere to the troposphere. J. Geophys. Res., 104, 30 93730 946, https://doi.org/10.1029/1999JD900445.

    • Search Google Scholar
    • Export Citation
  • Basu, S., D. J. Sauchyn, and M. R. Anis, 2020: Hydrological extremes in the Canadian Prairies in the last decade due to the ENSO teleconnection—A comparative case study using WRF. Water, 12, 2970, https://doi.org/10.3390/w12112970.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B., and M. Regier, 2007: Historical comparison of the 2001/2002 drought in the Canadian Prairies. Climate Res., 33, 229242, https://doi.org/10.3354/cr033229.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B., and A. Shabbar, 2008: Impacts of large-scale circulation variability on low streamflows over Canada: A review. Can. Water Resour. J., 33, 137154, https://doi.org/10.4296/cwrj3302137.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B., and A. Shabbar, 2011: Large-scale climate oscillations influencing Canada, 1900–2008. Canadian Biodiversity: Ecosystem Status and Trends 2010. Canadian Councils of Resource Ministers Tech. Thematic Rep. 4, 20 pp., https://publications.gc.ca/collections/collection_2011/ec/En14-43-4-2011-eng.pdf.

  • Bonsal, B., Z. Liu, E. Wheaton, and R. Stewart, 2020: Historical and projected changes to the stages and other characteristics of severe Canadian Prairie droughts. Water, 12, 3370, https://doi.org/10.3390/w12123370.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B. R., and R. G. Lawford, 1999: Teleconnections between El Nino and La Nina events and summer extended dry spells on the Canadian Prairies. Int. J. Climatol., 19, 14451458, https://doi.org/10.1002/(SICI)1097-0088(19991115)19:13<1445::AID-JOC431>3.0.CO;2-7.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B. R., E. E. Wheaton, A. C. Chipanshi, C. L. Lin, D. J. Sauchyn, and L. Wen, 2011: Drought research in Canada: A review. Atmos.–Ocean, 49, 303319, https://doi.org/10.1080/07055900.2011.555103.

    • Search Google Scholar
    • Export Citation
  • Cammalleri, C., P. Barbosa, and J. V. Vogt, 2019: Analysing the relationship between multiple-timescale SPI and GRACE terrestrial water storage in the framework of drought monitoring. Water, 11, 1672, https://doi.org/10.3390/w11081672.

    • Search Google Scholar
    • Export Citation
  • Champagne, C., Y. Zhang, P. Cherneski, and T. Hadwen, 2018: Estimating regional scale hydroclimatic risk conditions from the Soil Moisture Active-Passive (SMAP) satellite. Geosciences, 8, 127, https://doi.org/10.3390/geosciences8040127.

    • Search Google Scholar
    • Export Citation
  • Champagne, C., J. White, A. Berg, S. Belair, and M. Carrera, 2019: Impact of soil moisture data characteristics on the sensitivity to crop yields under drought and excess moisture conditions. Remote Sens., 11, 372, https://doi.org/10.3390/rs11040372.

    • Search Google Scholar
    • Export Citation
  • Chartrand, J., and F. S. R. Pausata, 2020: Impacts of the North Atlantic Oscillation on winter precipitations and storm track variability in southeast Canada and the northeast United States. Wea. Climate Dyn., 1, 731744, https://doi.org/10.5194/wcd-1-731-2020.

    • Search Google Scholar
    • Export Citation
  • Chen, H., J. D. Janizek, S. Lundberg, and S.-I. Lee, 2020: True to the model or true to the data? arXiv, 2006.16234v1, https://doi.org/10.48550/arXiv.2006.16234.

  • Chen, T., and C. Guestrin, 2016: XGBoost: A scalable tree boosting system. KDD’16: Proc. 22nd ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, San Francisco, CA, Association for Computing Machinery, 785–194, https://doi.org/10.1145/2939672.2939785.

  • Cohen, J., 1968: Weighted Kappa: Nominal scale agreement provision for scaled disagreement or partial credit. Psychol. Bull., 70, 213220, https://doi.org/10.1037/h0026256.

    • Search Google Scholar
    • Export Citation
  • Cook, B. I., J. S. Mankin, K. Marvel, A. P. Williams, J. E. Smerdon, and K. J. Anchukaitis, 2020: Twenty-first century drought projections in the CMIP6 forcing scenarios. Earth’s Future, 8, e2019EF001461, https://doi.org/10.1029/2019EF001461.

    • Search Google Scholar
    • Export Citation
  • Dai, A., 2011: Characteristics and trends in various forms of the Palmer Drought Severity Index during 1900–2008. J. Geophys. Res., 116, D12115, https://doi.org/10.1029/2010JD015541.

    • Search Google Scholar
    • Export Citation
  • Dikshit, A., and B. Pradhan, 2021: Explainable AI in drought forecasting. Mach. Learn. Appl., 6, 100192, https://doi.org/10.1016/j.mlwa.2021.100192.

    • Search Google Scholar
    • Export Citation
  • Ding, Y., M. J. Hayes, and M. Widhalm, 2011: Measuring economic impacts of drought: A review and discussion. Disaster Prev. Manage., 20, 434446, https://doi.org/10.1108/09653561111161752.

    • Search Google Scholar
    • Export Citation
  • Dorigo, W., and Coauthors, 2017: ESA CCI Soil Moisture for improved Earth system understanding: State-of-the art and future directions. Remote Sens. Environ., 203, 185215, https://doi.org/10.1016/j.rse.2017.07.001.

    • Search Google Scholar
    • Export Citation
  • Ecological Stratification Working Group, 1995: A National Ecological Framework for Canada. Agriculture and Agri-Food Canada Tech. Rep., 132 pp., https://sis.agr.gc.ca/cansis/publications/manuals/1996/A42-65-1996-national-ecological-framework.pdf.

  • Enfield, D. B., A. M. Mestas-Nunez, and P. J. Trimble, 2001: The Atlantic multidecadal oscillation and its relation to rainfall and river flows in the continental U.S. Geophys. Res. Lett., 28, 20772080, https://doi.org/10.1029/2000GL012745.

    • Search Google Scholar
    • Export Citation
  • ESA, 2023: ESA CCI soil moisture. Universität Hamburg, accessed 20 July 2023, https://www.esa-soilmoisture-cci.org/data.

  • Franzke, C., S. B. Feldstein, and S. Lee, 2011: Synoptic analysis of the Pacific–North American teleconnection pattern. Quart. J. Roy. Meteor. Soc., 137, 329346, https://doi.org/10.1002/qj.768.

    • Search Google Scholar
    • Export Citation
  • Ge, Y., and D. Luo, 2023: Impacts of the different types of El Niño and PDO on the winter sub-seasonal North American zonal temperature dipole via the variability of positive PNA events. Climate Dyn., 60, 13971413, https://doi.org/10.1007/s00382-022-06393-z.

    • Search Google Scholar
    • Export Citation
  • Hanesiak, J. M., and Coauthors, 2011: Characterization and summary of the 1999–2005 Canadian Prairie drought. Atmos.–Ocean, 49, 421452, https://doi.org/10.1080/07055900.2011.626757.

    • Search Google Scholar
    • Export Citation
  • Hoerling, M., J. Eischeid, A. Kumar, R. Leung, A. Mariotti, K. Mo, S. D. Schubert, and R. Seager, 2014: Causes and predictability of the 2012 Great Plains drought. Bull. Amer. Meteor. Soc., 95, 269282, https://doi.org/10.1175/BAMS-D-13-00055.1.

    • Search Google Scholar
    • Export Citation
  • Houborg, R., M. Rodell, B. Li, R. Reichle, and B. F. Zaitchik, 2012: Drought indicators based on model-assimilated Gravity Recovery and Climate Experiment (GRACE) terrestrial water storage observations. Water Resour. Res., 48, W07525, https://doi.org/10.1029/2011WR011291.

    • Search Google Scholar
    • Export Citation
  • Hurrell, J. W., Y. Kushnir, G. Ottersen, and M. Visbeck, 2003: An overview of the North Atlantic Oscillation. The North Atlantic Oscillation: Climatic Significance and Environmental Impact, Geophys. Monogr., Vol. 134, Amer. Geophys. Union, 1–35, https://doi.org/10.1029/134GM01.

  • Janzing, D., L. Minorics, and P. Blöbaum, 2019: Feature relevance quantification in explainable AI: A causal problem. Proc. 23rd Int. Conf. on Artificial Intelligence and Statistics (AISTATS), Palermo, Italy, PMLR, 2907–2916, https://proceedings.mlr.press/v108/janzing20a.html.

  • Jiang, R., T. Y. Gan, J. Xie, and N. Wang, 2014: Spatiotemporal variability of Alberta’s seasonal precipitation, their teleconnection with large-scale climate anomalies and sea surface temperature. Int. J. Climatol., 34, 28992917, https://doi.org/10.1002/joc.3883.

    • Search Google Scholar
    • Export Citation
  • Johnson, N. C., and S. B. Feldstein, 2010: The continuum of North Pacific sea level pressure patterns: Intraseasonal, interannual, and interdecadal variability. J. Climate, 23, 851867, https://doi.org/10.1175/2009JCLI3099.1.

    • Search Google Scholar
    • Export Citation
  • Kerr, S. A., Y. Andreichuk, and D. Sauchyn, 2021: Warm and cool season reconstruction and assessment of the long-term hydroclimatic variability of the Canadian Prairie provinces through the development of the Canadian Prairies paleo drought atlas. Int. J. Climatol., 41, 35393560, https://doi.org/10.1002/joc.7034.

    • Search Google Scholar
    • Export Citation
  • Knox, J. L., and R. G. Lawford, 1990: The relationship between Canadian Prairie dry and wet months and circulation anomalies in the mid‐troposphere. Atmos.–Ocean, 28, 189215, https://doi.org/10.1080/07055900.1990.9649374.

    • Search Google Scholar
    • Export Citation
  • Laberge, G., and Y. Pequignot, 2022: Understanding interventional TreeSHAP: How and why it works. arXiv, 2209.15123v2, https://doi.org/10.48550/arXiv.2209.15123.

  • Lawrence, Z. D., J. Perlwitz, A. H. Butler, G. L. Manney, P. A. Newman, S. H. Lee, and E. R. Nash, 2020: The remarkably strong Arctic stratospheric polar vortex of winter 2020: Links to record-breaking Arctic Oscillation and ozone loss. J. Geophys. Res. Atmos., 125, e2020JD033271, https://doi.org/10.1029/2020JD033271.

    • Search Google Scholar
    • Export Citation
  • Leathers, D. J., B. Yarnal, and M. A. Palecki, 1991: The Pacific/North American teleconnection pattern and United States climate. Part I: Regional temperature and precipitation associations. J. Climate, 4, 517528, https://doi.org/10.1175/1520-0442(1991)004<0517:TPATPA>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Li, X., Z.-Z. Hu, P. Liang, and J. Zhu, 2019: Contrastive influence of ENSO and PNA on variability and predictability of North American winter precipitation. J. Climate, 32, 62716284, https://doi.org/10.1175/JCLI-D-19-0033.1.

    • Search Google Scholar
    • Export Citation
  • Li, Z., Y. Li, B. Bonsal, A. H. Manson, and L. Scaff, 2018: Combined impacts of ENSO and MJO on the 2015 growing season drought on the Canadian Prairies. Hydrol. Earth Syst. Sci., 22, 50575067, https://doi.org/10.5194/hess-22-5057-2018.

    • Search Google Scholar
    • Export Citation
  • Liu, Y., X. Yang, L. Ren, F. Yuan, S. Jiang, and M. Ma, 2015: A new physically based self-calibrating Palmer drought severity index and its performance evaluation. Water Resour. Manage., 29, 48334847, https://doi.org/10.1007/s11269-015-1093-9.

    • Search Google Scholar
    • Export Citation
  • Lundberg, S. M., and S.-I. Lee, 2017: A unified approach to interpreting model predictions. NIPS’17: 31st Conf. on Neural Information Processing Systems (NIPS), Long Beach, CA, Curran Associates Inc., 4768–4777, https://dl.acm.org/doi/10.5555/3295222.3295230.

  • Lundberg, S. M., and Coauthors, 2020: From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell., 2, 5667, https://doi.org/10.1038/s42256-019-0138-9.

    • Search Google Scholar
    • Export Citation
  • Mori, M., and M. Watanabe, 2008: The growth and triggering mechanisms of the PNA: A MJO-PNA coherence. J. Meteor. Soc. Japan, 86, 213236, https://doi.org/10.2151/jmsj.86.213.

    • Search Google Scholar
    • Export Citation
  • NCDC, 2021: Teleconnections. NOAA, accessed 20 July 2023, https://www.ncdc.noaa.gov/teleconnections/.

  • Newton, B. W., T. D. Prowse, and B. R. Bonsal, 2014a: Evaluating the distribution of water resources in western Canada using synoptic climatology and selected teleconnections. Part 1: Winter season. Hydrol. Processes, 28, 42194234, https://doi.org/10.1002/hyp.10233.

    • Search Google Scholar
    • Export Citation
  • Newton, B. W., T. D. Prowse, and B. R. Bonsal, 2014b: Evaluating the distribution of water resources in western Canada using synoptic climatology and selected teleconnections. Part 2: Summer season. Hydrol. Processes, 28, 42354249, https://doi.org/10.1002/hyp.10235.

    • Search Google Scholar
    • Export Citation
  • Nguyen, H., M. C. Wheeler, J. A. Otkin, T. Cowan, A. Frost, and R. Stone, 2019: Using the evaporative stress index to monitor flash drought in Australia. Environ. Res. Lett., 14, 064016, https://doi.org/10.1088/1748-9326/ab2103.

    • Search Google Scholar
    • Export Citation
  • O’Neil, H. C. L., T. D. Prowse, B. R. Bonsal, and Y. B. Dibike, 2017: Spatial and temporal characteristics in streamflow-related hydroclimatic variables over western Canada. Part 1: 1950–2010. Hydrol. Res., 48, 915931, https://doi.org/10.2166/nh.2016.057.

    • Search Google Scholar
    • Export Citation
  • Oozeer, Y., C. G. Fletcher, and C. Champagne, 2020: Evaluation of satellite-derived surface soil moisture products over agricultural regions of Canada. Remote Sens., 12, 1455, https://doi.org/10.3390/rs12091455.

    • Search Google Scholar
    • Export Citation
  • Otkin, J. A., M. C. Anderson, C. Hain, I. E. Mladenova, J. B. Basara, and M. Svoboda, 2013: Examining rapid onset drought development using the thermal infrared–based evaporative stress index. J. Hydrometeor., 14, 10571074, https://doi.org/10.1175/JHM-D-12-0144.1.

    • Search Google Scholar
    • Export Citation
  • Otkin, J. A., M. Svoboda, E. D. Hunt, T. W. Ford, M. C. Anderson, C. Hain, and J. B. Basara, 2018: Flash droughts: A review and assessment of the challenges imposed by rapid-onset droughts in the United States. Bull. Amer. Meteor. Soc., 99, 911919, https://doi.org/10.1175/BAMS-D-17-0149.1.

    • Search Google Scholar
    • Export Citation
  • Palmer, W. C., 1965: Meteorological drought. U.S. Weather Bureau Research Paper 45, 58 pp., https://www.ncei.noaa.gov/monitoring-content/temp-and-precip/drought/docs/palmer.pdf.

  • Parker, T., A. Gallant, M. Hobbins, and D. Hoffmann, 2021: Flash drought in Australia and its relationship to evaporative demand. Environ. Res. Lett., 16, 064033, https://doi.org/10.1088/1748-9326/abfe2c.

    • Search Google Scholar
    • Export Citation
  • PSL, 2021a: Physical Sciences Laboratory. NOAA, accessed 20 July 2023, https://psl.noaa.gov/.

  • PSL, 2021b: Multivariate ENSO Index version 2 (MEI.v2). NOAA, accessed 20 July 2023, https://psl.noaa.gov/enso/mei/.

  • PSL, 2021c: AMO (Atlantic Multidecadal Oscillation) index. NOAA, accessed 20 July 2023, https://psl.noaa.gov/data/timeseries/AMO/.

  • Rippey, B., B. Fuchs, D. Simeral, D. Bathke, R. Heim, and M. Svoboda, 2021: The drought monitor comes of age. Weatherwise, 74, 2937, https://doi.org/10.1080/00431672.2021.1873000.

    • Search Google Scholar
    • Export Citation
  • Roberts, E., R. E. Stewart, and C. A. Lin, 2006: A study of drought characteristics over the Canadian Prairies. Atmos.–Ocean, 44, 331345, https://doi.org/10.3137/ao.440402.

    • Search Google Scholar
    • Export Citation
  • Seneviratne, S. I., T. Corti, E. L. Davin, M. Hirschi, E. B. Jaeger, I. Lehner, B. Orlowsky, and A. J. Teuling, 2010: Investigating soil moisture–climate interactions in a changing climate: A review. Earth Sci. Rev., 99, 125161, https://doi.org/10.1016/j.earscirev.2010.02.004.

    • Search Google Scholar
    • Export Citation
  • Seo, K.-H., and H.-J. Lee, 2017: Mechanisms for a PNA-like teleconnection pattern in response to the MJO. J. Atmos. Sci., 74, 17671781, https://doi.org/10.1175/JAS-D-16-0343.1.

    • Search Google Scholar
    • Export Citation
  • Shabbar, A., and W. Skinner, 2004: Summer drought patterns in Canada and the relationship to global sea surface temperatures. J. Climate, 17, 28662880, https://doi.org/10.1175/1520-0442(2004)017<2866:SDPICA>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Shabbar, A., B. R. Bonsal, and K. Szeto, 2011: Atmospheric and oceanic variability associated with growing season droughts and pluvials on the Canadian Prairies. Atmos.–Ocean, 49, 339355, https://doi.org/10.1080/07055900.2011.564908.

    • Search Google Scholar
    • Export Citation
  • Shapley, L., 1953: A value for n-person games. Contributions to the Theory of Games (AM-28), H. W. Kuhn and A. W. Tucker, Eds., Vol. II, Princeton University Press, 307–317.

  • Soulard, N., H. Lin, and B. Yu, 2019: The changing relationship between ENSO and its extratropical response patterns. Sci. Rep., 9, 6507, https://doi.org/10.1038/s41598-019-42922-3.

    • Search Google Scholar
    • Export Citation
  • Svoboda, M., and Coauthors, 2002: The drought monitor. Bull. Amer. Meteor. Soc., 83, 11811190, https://doi.org/10.1175/1520-0477-83.8.1181.

    • Search Google Scholar
    • Export Citation
  • Szép, I. J., J. Mika, and Z. Dunkel, 2005: Palmer Drought Severity Index as soil moisture indicator: Physical interpretation, statistical behaviour and relation to global climate. Phys. Chem. Earth, 30, 231243, https://doi.org/10.1016/j.pce.2004.08.039.

    • Search Google Scholar
    • Export Citation
  • Szeto, K., X. Zhang, R. E. White, and J. Brimelow, 2016: The 2015 extreme drought in western Canada [in “Explaining Extreme Events of 2015 from a Climate Perspective”]. Bull. Amer. Meteor. Soc., 97 (12), S42S46, https://doi.org/10.1175/BAMS-D-16-0147.1.

    • Search Google Scholar
    • Export Citation
  • Tam, B. Y., K. Szeto, B. Bonsal, G. Flato, A. J. Cannon, and R. Rong, 2019: CMIP5 drought projections in Canada based on the standardized precipitation evapotranspiration index. Can. Water Resour. J., 44, 90107, https://doi.org/10.1080/07011784.2018.1537812.

    • Search Google Scholar
    • Export Citation
  • Thomas, B. F., J. S. Famiglietti, F. W. Landerer, D. N. Wiese, N. P. Molotch, and D. F. Argus, 2017: GRACE groundwater drought index: Evaluation of California Central Valley groundwater drought. Remote Sens. Environ., 198, 384392, https://doi.org/10.1016/j.rse.2017.06.026.

    • Search Google Scholar
    • Export Citation
  • Trenary, L., and T. DelSole, 2016: Does the Atlantic multidecadal oscillation get its predictability from the Atlantic meridional overturning circulation? J. Climate, 29, 52675280, https://doi.org/10.1175/JCLI-D-16-0030.1.

    • Search Google Scholar
    • Export Citation
  • Vicente-Serrano, S. M., S. Beguería, and J. I. Lopez-Moreno, 2010: A multiscalar drought index sensitive to global warming: The standardized precipitation evapotranspiration index. J. Climate, 23, 16961718, https://doi.org/10.1175/2009JCLI2909.1.

    • Search Google Scholar
    • Export Citation
  • Wahr, J., S. Swenson, V. Zlotnicki, and I. Velicogna, 2004: Time-variable gravity from GRACE: First results. Geophys. Res. Lett., 31, L11501, https://doi.org/10.1029/2004GL019779.

    • Search Google Scholar
    • Export Citation
  • Wang, L., and W. Chen, 2010: Downward Arctic Oscillation signal associated with moderate weak stratospheric polar vortex and the cold December 2009. Geophys. Res. Lett., 37, L09707, https://doi.org/10.1029/2010GL042659.

    • Search Google Scholar
    • Export Citation
  • Wheaton, E., S. Kulshreshtha, V. Wittrock, and G. Koshida, 2008: Dry times: Hard lessons from the Canadian drought of 2001 and 2002. Can. Geogr., 52, 241262, https://doi.org/10.1111/j.1541-0064.2008.00211.x.

    • Search Google Scholar
    • Export Citation
  • Wilhite, D. A., M. D. Svoboda, and M. J. Hayes, 2007: Understanding the complex impacts of drought: A key to enhancing drought mitigation and preparedness. Water Resour. Manage., 21, 763774, https://doi.org/10.1007/s11269-006-9076-5.

    • Search Google Scholar
    • Export Citation
  • Wood, E. F., S. D. Schubert, A. W. Wood, C. D. Peters-Lidard, K. C. Mo, A. Mariotti, and R. S. Pulwarty, 2015: Prospects for advancing drought understanding, monitoring, and prediction. J. Hydrometeor., 16, 16361657, https://doi.org/10.1175/JHM-D-14-0164.1.

    • Search Google Scholar
    • Export Citation
  • Yang, K., and C. Wang, 2019: Seasonal persistence of soil moisture anomalies related to freeze–thaw over the Tibetan Plateau and prediction signal of summer precipitation in eastern China. Climate Dyn., 53, 24112424, https://doi.org/10.1007/s00382-019-04867-1.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., T. Y. Gan, and X. Tan, 2020: Spatiotemporal changes of drought characteristics and their dynamic drivers in Canada. Atmos. Res., 232, 104695, https://doi.org/10.1016/j.atmosres.2019.104695.

    • Search Google Scholar
    • Export Citation
  • Yu, B., and F. W. Zwiers, 2007: The impact of combined ENSO and PDO on the PNA climate: A 1,000-year climate modeling study. Climate Dyn., 29, 837851, https://doi.org/10.1007/s00382-007-0267-4.

    • Search Google Scholar
    • Export Citation
Save
  • AAFC, 2015: Annual crop inventory. Government of Canada, accessed 20 July 2023, https://open.canada.ca/data/en/dataset/ba2645d5-4458-414d-b196-6303ac06c1c9.

  • AAFC, 2017: Standardized precipitation evapotranspiration index. Government of Canada, accessed 20 July 2023, https://open.canada.ca/data/en/dataset/d765cc41-8ee0-4aca-be4b-084448e52a58.

  • Akinremi, O. O., S. M. McGinn, and A. G. Barr, 1996: Evaluation of the Palmer drought index on the Canadian Prairies. J. Climate, 9, 897905, https://doi.org/10.1175/1520-0442(1996)009<0897:EOTPDI>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Anderson, M. C., and Coauthors, 2011: Mapping daily evapotranspiration at field to continental scales using geostationary and polar orbiting satellite imagery. Hydrol. Earth Syst. Sci., 15, 223239, https://doi.org/10.5194/hess-15-223-2011.

    • Search Google Scholar
    • Export Citation
  • Anderson, M. C., C. Hain, J. Otkin, X. Zhan, K. Mo, M. Svoboda, B. Wardlow, and A. Pimstein, 2013: An intercomparison of drought indicators based on thermal remote sensing and NLDAS-2 simulations with U.S. Drought Monitor classifications. J. Hydrometeor., 14, 10351056, https://doi.org/10.1175/JHM-D-12-0140.1.

    • Search Google Scholar
    • Export Citation
  • Anderson, M. C., and Coauthors, 2016: The evaporative stress index as an indicator of agricultural drought in Brazil: An assessment based on crop yield impacts. Remote Sens. Environ., 174, 8299, https://doi.org/10.1016/j.rse.2015.11.034.

    • Search Google Scholar
    • Export Citation
  • Asong, Z. E., H. S. Wheater, B. Bonsal, S. Razavi, and S. Kurkute, 2018: Historical drought patterns over Canada and their teleconnections with large-scale climate signals. Hydrol. Earth Syst. Sci., 22, 31053124, https://doi.org/10.5194/hess-22-3105-2018.

    • Search Google Scholar
    • Export Citation
  • Baldwin, M. P., and T. J. Dunkerton, 1999: Propagation of the Arctic Oscillation from the stratosphere to the troposphere. J. Geophys. Res., 104, 30 93730 946, https://doi.org/10.1029/1999JD900445.

    • Search Google Scholar
    • Export Citation
  • Basu, S., D. J. Sauchyn, and M. R. Anis, 2020: Hydrological extremes in the Canadian Prairies in the last decade due to the ENSO teleconnection—A comparative case study using WRF. Water, 12, 2970, https://doi.org/10.3390/w12112970.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B., and M. Regier, 2007: Historical comparison of the 2001/2002 drought in the Canadian Prairies. Climate Res., 33, 229242, https://doi.org/10.3354/cr033229.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B., and A. Shabbar, 2008: Impacts of large-scale circulation variability on low streamflows over Canada: A review. Can. Water Resour. J., 33, 137154, https://doi.org/10.4296/cwrj3302137.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B., and A. Shabbar, 2011: Large-scale climate oscillations influencing Canada, 1900–2008. Canadian Biodiversity: Ecosystem Status and Trends 2010. Canadian Councils of Resource Ministers Tech. Thematic Rep. 4, 20 pp., https://publications.gc.ca/collections/collection_2011/ec/En14-43-4-2011-eng.pdf.

  • Bonsal, B., Z. Liu, E. Wheaton, and R. Stewart, 2020: Historical and projected changes to the stages and other characteristics of severe Canadian Prairie droughts. Water, 12, 3370, https://doi.org/10.3390/w12123370.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B. R., and R. G. Lawford, 1999: Teleconnections between El Nino and La Nina events and summer extended dry spells on the Canadian Prairies. Int. J. Climatol., 19, 14451458, https://doi.org/10.1002/(SICI)1097-0088(19991115)19:13<1445::AID-JOC431>3.0.CO;2-7.

    • Search Google Scholar
    • Export Citation
  • Bonsal, B. R., E. E. Wheaton, A. C. Chipanshi, C. L. Lin, D. J. Sauchyn, and L. Wen, 2011: Drought research in Canada: A review. Atmos.–Ocean, 49, 303319, https://doi.org/10.1080/07055900.2011.555103.

    • Search Google Scholar
    • Export Citation
  • Cammalleri, C., P. Barbosa, and J. V. Vogt, 2019: Analysing the relationship between multiple-timescale SPI and GRACE terrestrial water storage in the framework of drought monitoring. Water, 11, 1672, https://doi.org/10.3390/w11081672.

    • Search Google Scholar
    • Export Citation
  • Champagne, C., Y. Zhang, P. Cherneski, and T. Hadwen, 2018: Estimating regional scale hydroclimatic risk conditions from the Soil Moisture Active-Passive (SMAP) satellite. Geosciences, 8, 127, https://doi.org/10.3390/geosciences8040127.

    • Search Google Scholar
    • Export Citation
  • Champagne, C., J. White, A. Berg, S. Belair, and M. Carrera, 2019: Impact of soil moisture data characteristics on the sensitivity to crop yields under drought and excess moisture conditions. Remote Sens., 11, 372, https://doi.org/10.3390/rs11040372.

    • Search Google Scholar
    • Export Citation
  • Chartrand, J., and F. S. R. Pausata, 2020: Impacts of the North Atlantic Oscillation on winter precipitations and storm track variability in southeast Canada and the northeast United States. Wea. Climate Dyn., 1, 731744, https://doi.org/10.5194/wcd-1-731-2020.

    • Search Google Scholar
    • Export Citation
  • Chen, H., J. D. Janizek, S. Lundberg, and S.-I. Lee, 2020: True to the model or true to the data? arXiv, 2006.16234v1, https://doi.org/10.48550/arXiv.2006.16234.

  • Chen, T., and C. Guestrin, 2016: XGBoost: A scalable tree boosting system. KDD’16: Proc. 22nd ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, San Francisco, CA, Association for Computing Machinery, 785–194, https://doi.org/10.1145/2939672.2939785.

  • Cohen, J., 1968: Weighted Kappa: Nominal scale agreement provision for scaled disagreement or partial credit. Psychol. Bull., 70, 213220, https://doi.org/10.1037/h0026256.