Drought is a diverse and complicated natural disaster that causes significant global economic damage and can occur in any climate regime (Liu et al., 2022). It is an extremely hazardous natural calamity with major consequences for nature, human creation, and life (Hussain et al., 2019; Quiring and Papakryiakou, 2003; Wang et al., 2020; Yao et al., 2018). A drought is a natural disaster caused by prolonged periods of below-average precipitation. Drought poses a significant and intricate challenge, demanding attention owing to its multifaceted nature, which has substantial economic, social, and environmental consequences (Wilhite et al., 2007, 2014; Van Loon, 2015). The complexity of this phenomenon is highlighted by its profound impact, as evidenced by various studies (Belayneh et al., 2014; Sierra-Soler et al., 2016; Saadat et al., 2011). The pivotal role of drought monitoring in addressing this challenge cannot be overstated. Drought monitoring involves the systematic collection, analysis, and dissemination of data on drought conditions, serving as a fundamental basis for informing decision-makers and facilitating the development of effective early warning systems (Fadhil, 2011; Pei et al., 2018; Feng et al., 2019). This comprehensive approach is indispensable for minimizing the risks associated with drought events and enhancing preparedness strategies. The significance of effective drought monitoring is underscored by its crucial role in mitigating the far-reaching impacts of drought across various sectors, ensuring informed decision-making, and proactive measures to enhance resilience and reduce vulnerability.
The importance of evaluating drought goes beyond understanding its mechanisms and devising appropriate strategies. This involves scrutinizing its societal, economic, and ecological consequences (Al-Quraishi et al., 2021; Dejene et al., 2023, Dejene et al., 2023). More research is imperative in light of the emergence of drought, as forecasting droughts continues to pose significant challenges. A comprehensive understanding of the impacts of drought is still lacking, and the growing complexity of human-environment systems further complicates the effective management of drought situations (Wilhite et al., 2000). Remote sensing (RS) is a valuable approach for drought detection and prediction, where it can be used to identify and monitor the physical properties of an area from a distance by measuring reflected and emitted radiation (Wilhite et al., 2000; Palchaudhuri and Biswas, 2016; Jindoet al. 2021; Bello and Aina, 2014; Purkey et al., 2008; Bojer et al., 2023). Changes in vegetation health, soil moisture, and surface water levels can be tracked using remote sensing data (Gaznayee et al., 2022; Al-Quraishi and Negm, 2020). RS data are being utilized to create new drought prediction models, identify patterns and trends to predict future drought events, and provide timely and accurate information for drought response and preparedness.
Millions of people living in the drought-prone Horn of Africa face an increasing threat from the lack of safe, reliable, and affordable water year-round as droughts become more severe and frequent (Thomas et al., 2020). Africa experienced its lengthiest drought period towards the conclusion of the 21st century, lasting for a substantial duration of 197 months (O. E. Adeyeri et al., 2023). To thoroughly examine extreme drought levels, we propose the following research questions to analyze the corresponding risk levels: What is the role of rainfall performance in determining drought occurrence in both low- and high-rainfall areas? How does climate change contribute to the increased frequency and impact of droughts in the water-stressed regions of Africa, particularly in Ethiopia? What historical and recent impacts have frequent and severe droughts on society in Ethiopia, including crop failures, livestock deaths, and food insecurity? How has Ethiopia’s status as a low-income country with a large population in sub-Saharan Africa influenced its vulnerability to drought, as suggested in the historical context? To what extent can protracted droughts in Ethiopia be attributed to powerful El Niño events? What is the recent impact of drought in the Horn of Africa?
Drought occurrence is mostly determined by rainfall performance at a given location, with droughts occurring in low and high rainfall (Wilhite and Glantz, 1985). Droughts are projected to become more frequent and have a greater impact because of climate change in areas of Africa that are already water-stressed (Dai, 2011; Hulme, 1992; Field and Barros, 2014; Masson-Delmotte, 2018; Gaiballah and Abdalla, 2016). Ethiopia has experienced frequent and severe droughts, which have historically and recently caused significant harm to society, leading to crop failures, livestock deaths, and food insecurity. Historically, Ethiopia is a low-income country with the second highest population in sub-Saharan Africa. Numerous reports have shown Ethiopia’s protracted drought results from powerful El Niño events (Haile, 1988). While this appears to be partially correct, the truth is more complicated and should be recognized as such. It has a long history of droughts, which have become more severe, frequent, and devastating since the shocking events of the 1970s (Dai, 2011; Mekonen et al., 2020; Mohammad et al., 2018). More than 4.5 million people who required food support were left by the most recent drought in the Horn of Africa (Philip et al., 2018), in addition to 7.5 million people who had previously required food aid. In the south and southeast of the country, significant cattle mortality has been caused by dry pastures and water shortages (UN-OCHA, 2022). Pastoral and agro-pastoral populations in the country’s east and south have suffered significant losses of animal assets due to consecutive droughts (Abdulkadr, 2019; FAO, 2021).
In the Borena Zone, Ethiopia, consecutive failures of the rainy seasons in 2021 and 2022 have led to severe and widespread droughts, which are considered the most impactful in the last 40 years. This persistent drought has resulted in 3.5 million livestock fatalities and put an additional 25 million at risk, as reported in the first three months of 2022 (Dejene et al., 2023). These consequences have been exacerbated by insufficient time between droughts for communities to recover. Alarmingly, local authorities have noted a more than 10% increase in severe malnutrition cases in the Borena Zone, with projections indicating a continued upward trend (Yisehak et al., 2021). The vulnerability of the zone to natural disasters is attributed to factors such as reliance on rain-fed agriculture, high animal stocking rates, soil degradation, population pressure, and an inadequate administrative system (Dejene et al., 2023). To effectively address the issue of drought risk in a zone, leveraging machine learning is crucial for predicting and evaluating the level of vulnerability in the area (Gonzalez de Andres et al., 2022; Wang et al., 2016; Chen et al., 2017; O.E. Adeyeri et al., 2023). Machine learning is a rapidly evolving discipline with diverse applications and one of its notable applications is drought modeling and forecasting (Azizi et al., 2019). Numerous studies have explored various methodologies to assess drought, highlighting the complexity of developing reliable prediction models owing to this natural phenomenon’s intricate and destructive nature (Thomas et al., 2020). A promising approach involves implementing a drought prediction system based on machine learning, which employs multiple ML models to anticipate drought occurrences. Each model was trained on a distinct dataset, and their predictions were integrated to generate a comprehensive forecast. This multifaceted approach is more accurate than relying on a single machine learning model because it considers diverse factors contributing to drought (Bouaziz et al., 2021).
In recent years, the popularity of ML technologies has led to their widespread adoption in environmental science and disaster management, as evidenced by numerous studies (Nguyen et al., 2020; cX Zhang et al., 2019, Zhang et al., 2019, Zhang et al., 2019; Li et al., 2020). A variety of ML techniques, including Support Vector Machines (Razavi-Termeh et al., 2021; Khosravi et al., 2019), Artificial Neural Networks (ANN) (Bullock et al., 2020b), Random Forests (RF) (Du et al., 2020b; Nhu et al., 2020), Decision Trees (DT) (Rahimzad et al., 2021), and Long Short-Term Memory (LSTM) (Mahadevan et al., 2019; Liu et al., 2022; S Zhang et al., 2019), have been prominently employed to analyze large and intricate datasets. These techniques contribute significantly to assessing drought severity, predicting disasters, aiding in post-disaster response and recovery, and delivering accurate information to decision-makers. Common indicators, such as the Evaporative Demand Drought Index (ETDI), Palmer Drought Severity Index (PDSI), Enhanced Soil Water Index (ESWI), and Standardized Precipitation Index (SPI), play crucial roles in measuring the severity of hydrological and meteorological drought (Vicente-Serrano et al., 2010; Hobbins et al., 2020; Iturbide et al., 2022). By estimating the duration of historical data, LSTM enhances its capability to forecast current drought conditions.
However, using ML algorithms for prediction faces several difficulties, including choosing the best modeling techniques from a broad array of options because each method yields a different set of outcomes (Shafizadeh-Moghadam et al., 2018). The precision of ML predictions is influenced by the availability and quality of historical data, and the developed models may be sensitive to variations in environmental conditions over time. Single-step LSTM prediction is simpler and more precise than multistep prediction. We employed hyperparameter tuning utilizing the GridSearch algorithm to overcome these limitations and identify optimal settings. This study developed regression models by combining dynamic models, ML, RS, and deep learning (DL). Integrating RS, ML, and DL is a promising new drought-monitoring and prediction method. In addition, the unique method for drought risk prediction developed in this study is based on a combination of RS, tree-based ensemble algorithms (ML), and DL.
ML-based drought forecasting methods have several advantages over the typical RS methods. First, it can capture complex spatial and temporal drought patterns more effectively. RS methods typically rely on simple statistical models to relate satellite data to drought conditions, whereas ML-based methods use more sophisticated models, such as deep learning models, to learn complex relationships between satellite data and drought conditions. Second, it is more robust to uncertainty. Drought forecasts can be inaccurate because RS data are susceptible to issues such as cloud cover, scanline error, and sensor malfunctions; however, ML-based methods can be trained to handle uncertainty in the data and produce more accurate forecasts. Third, it can be used to predict droughts at different spatial and temporal scales. Ultimately, ML algorithms demonstrate an impressive capacity for self-learning and improvement through data-driven processes. In contrast, DL employs neural networks to interpret complex patterns and relationships embedded within complex datasets. In contrast to RS, where human intervention is integral to the learning process, ML and DL, once exposed to sophisticated datasets, autonomously develop the ability to forecast future outcomes by drawing insights from historical and multifaceted data.
Furthermore, one of the most difficult challenges in drought risk assessment and prediction is integrating diverse spatiotemporal data from multiple sources, including high-resolution, medium-resolution, and low-resolution satellite data. Integrating these different spatiotemporal datasets from various sources is a difficult task; however, it is critical to developing accurate and reliable drought risk assessment and prediction models (Fung et al., 2020; Guo et al., 2021; Saha et al., 2022; Gaitan et al., 2020). Drought, a complex phenomenon fraught with significant uncertainty, poses a substantial hurdle to creating accurate ML models. Evaluating data accuracy across markedly disparate spatiotemporal resolutions and scales has been explored in existing studies, recognizing that the suitability of various satellite RS products varies with different applications and regions. Despite data limitations, ongoing research endeavors have persisted in enhancing drought prediction through diverse machine-learning algorithms. This underscores the necessity for an in-depth comprehension of the physical causes and potential consequences of droughts, coupled with the capability to gather, access, and analyze extensive datasets from diverse Earth observation platforms. A thoughtful selection of ML algorithms is strongly recommended to address this challenge effectively. ML algorithms exhibit heightened flexibility and adaptability to the non-linear nature of data, rendering them well-suited for predicting droughts of varying durations, frequencies, and intensities. Regression machine learning methods, including cat boosting, Extreme Gradient Boosting (XG), Random Forest (RF), and Light Gradient Boosting Machine (LGBM), have successfully addressed these intricacies within diverse machine-learning frameworks.
The efficiency of these four ML and two DL models for forecasting drought conditions in the Borena zone was examined in this study. These models have a variety of advantages for predicting drought risk. They can handle large datasets, are noise-resistant, predict dryness at various spatial and temporal dimensions, and incorporate uncertainty into their predictions. Therefore, they are an effective technique for managing drought risk. Furthermore, numerous methods are available for tracking drought’s development and status. In this study, our primary objective is to conduct a comprehensive time series-based analysis of drought risk in the Borena zone by leveraging a multisource remote sensing (RS) dataset and advanced machine learning (ML) methods. Specifically, our goals are to: (1) Contribute scientific insights to enhance the precision of drought assessments in the Borena region. (2) Investigate spatiotemporal patterns of drought risk, with a focus on understanding the influence of precipitation and temperature factors across historical instances. (3) Evaluate drought risk by integrating meteorological data into the analysis framework. (4) Explore the application of two widely recognized deep-learning techniques, Long Short-Term Memory (LSTM) and RS, for the assessment and prediction of drought events. (5) Introduce novel regression deep learning techniques, capitalizing on the capabilities of LSTM, to specifically enhance forecasting accuracy for the Borena zone over the period from 1993 to 2022.
