Debris flows are episodic, high-magnitude geohazards that predominantly occur in mountainous catchments. Their inherent unpredictability and destructive force pose formidable threats to human safety, socio-economic assets, and the operational integrity of critical infrastructure (Baggio et al., 2024). Central Asia, specifically Tajikistan, represents a prototypical high-risk zone characterized by complex orogeny, distinct hydro-climatic regimes, and intensifying anthropogenic interventions, which collectively catalyze frequent debris flow occurrences (Erokhin et al., 2018; Leinss et al., 2021). Recently, the perturbation of climate patterns has exacerbated extreme precipitation events, leading to an escalation in both the frequency and severity of debris flows. This trajectory presents systemic challenges to regional socio-economic development and ecological security (Sorg et al., 2012). For example, on 12 May 2021, continuous heavy rainfall triggered debris flows in Khatlon Province, Tajikistan, causing at least eight deaths and damaging multiple houses, farmlands, bridges, and highways (Xinhua News Agency, 2021). On 27 August 2023, intense rainfall induced debris flows killed 11 people in Vahdat City and 2 people in Rudaki District, Tajikistan, buried 15 vehicles, and forced the evacuation of hundreds of residents (Safarov et al., 2025).

However, the systematic characterization of the spatial distribution and deterministic drivers of debris flows in Tajikistan is nascent, largely constrained by monitoring lacunae and a dearth of comprehensive baseline datasets. This evidentiary gap impedes the efficacy of early warning systems and regional risk governance. To bolster disaster mitigation and resilience, there is an urgent need to develop high-fidelity susceptibility assessment models and execute rigorous scientific risk zonation.

In recent decades, debris flow susceptibility has attracted increasing scholarly attention due to its systemic threat to human safety, infrastructure, and the sustainable development of mountainous regions (Yılmaz et al., 2023; Shahgedanova et al., 2024). Traditional methods for debris flow susceptibility assessment can be broadly categorized into the following groups. First, methods based on empirical knowledge and historical disaster records. Researchers delineate susceptible areas according to the frequency, magnitude, and spatial distribution of debris flows, often combined with field surveys and remote sensing interpretation (Carrara et al., 2008; Kappes et al., 2011; Horton et al., 2013). However, these approaches rely heavily on long-term observational data and are therefore difficult to apply in data-scarce regions, such as high-mountain valleys, given the complex and variable nature of debris-flow environments. Second, multi-criteria evaluation methods based on topographic, geological, and hydrological factors, including analytic hierarchy process (Chen et al., 2015), fuzzy logic models (Zhang et al., 2022), and composite index approaches (Rogelis and Werner, 2014). These methods can integrate information from multiple sources but are strongly dependent on expert judgment, introducing subjectivity. Third, methods based on statistical and probabilistic theory, such as frequency ratio (Angillieri, 2020), information value (Xu et al., 2013), weight-of-evidence (Ilia and Tsangaratos, 2016), and logistic regression models (Bregoli et al., 2015). These approaches are more objective and operationally feasible, but they often assume independence among variables, making it challenging to capture nonlinear relationships and handle high-dimensional data effectively. Fourth, numerical simulation methods based on mathematical-physical processes, including finite element (Naef et al., 2006), finite difference (Shen et al., 2018), discrete element (Li et al., 2012), and coupled hydrodynamic models (Wei et al., 2024). Such methods can realistically represent the physical processes of debris flows, but their complexity, high parameter requirements, and computational demands usually limit their application to small-scale or typical catchments for hazard assessment and zoning.

With advances in geospatial data and computing technology, quantitative approaches integrating Geographic Information System (GIS), remote sensing, and statistical models have been widely applied to delineate debris flow-prone areas. However, traditional approaches exhibit several limitations. Empirical equations and regression-based models often fail to capture nonlinear interactions among conditioning factors, making them unsuitable for regions with strong spatial heterogeneity (Meng et al., 2024; Wang et al., 2025). Likewise, physically based numerical simulations, although theoretically rigorous, are highly sensitive to parameter uncertainty, boundary conditions, and computational requirements, which restricts their applicability for large-scale susceptibility assessment (Ma et al., 2022). In contrast, machine learning models can effectively handle high-dimensional, multi-source, and nonlinear datasets, providing improved predictive performance compared with traditional empirical and statistical methods (Meng and Zhu, 2024; Xie et al., 2025b). More recently, algorithms such as Random Forest (RF) (Huang et al., 2022b), Support Vector Machine (SVM) (Zhao et al., 2024), and Multi-layer Perceptron (MLP) (Yang et al., 2024) have been increasingly adopted for susceptibility mapping. Despite their advantages, challenges remain in transferring machine learning models across different regions, incorporating environmental change scenarios, and improving interpretability of the underlying physical mechanisms.

Selecting appropriate conditioning factors is essential for model accuracy. A widely accepted assumption is that higher precipitation correlates with increased debris flow frequency. However, recent observations in arid and semi-arid regions suggest that severe events may also occur under drought-dominated conditions (Zhou et al., 2024). Study in the Tianshan Mountains has shown that large-scale debris flows can be triggered even with limited precipitation input (Hu et al., 2017), a phenomenon reported in other dry mountainous regions globally (Handwerger et al., 2019). This implies that prolonged drought may influence slope stability by altering soil properties and promoting the accumulation of loose deposits. Despite this, drought conditions have rarely been incorporated into susceptibility assessments, particularly in Central Asia. Considering the arid climate of Tajikistan, integrating a drought index into susceptibility modeling is both necessary and scientifically justified.

Building upon these research gaps, this study incorporates a drought index along with 13 other environmental factors to develop a comprehensive debris flow susceptibility assessment for Tajikistan. Using remote sensing, GIS, and multiple machine learning models, we perform susceptibility modeling, spatial differentiation analysis, and disaster risk zoning. The findings aim to provide a scientific basis and technical support for debris flow prevention, control, and emergency management in mountainous countries of Central Asia.

Tajikistan is located in the southeastern part of Central Asia (36°40′-41°05′N, 67°31′-75°14′E). It borders China to the east, Kyrgyzstan to the north, Uzbekistan to the west, and Afghanistan to the south, with a total land area of approximately 143,100 km². The terrain generally decreases in elevation from east to west (Fig. 1), encompassing three major mountain systems: the Pamir Plateau, the western margin of the Hindu Kush Mountains, and the southern Tianshan Mountains. It is one of the most rugged countries in the world, with mountains accounting for more than 93.00% of the total area. Major river valleys, such as the Vakhsh and Panj rivers, develop along fault zones. Their erosive action has deeply incised the mountains, forming characteristic gorges and alluvial fans. During periods of intense rainfall or snowmelt, these geomorphic and hydrological settings make the region highly susceptible to debris flow hazards.

Debris flow disasters in Tajikistan are mainly concentrated in areas where steep mountainous terrain and deeply incised river valleys converge (Fig. 1). Typical examples include the Zeravshan River and its upper reaches in the Gissar Mountains (such as Panjakent City, Ayni District, and Ghafurov District), as well as the Rasht Valley zone (including Rasht District and Nurobod District). These regions are generally characterized by high elevation, strong relief, and a dense gully network. These river valleys are characterized by steep slopes that provide abundant loose material. When subjected to heavy rainfall or snowmelt, the resulting surface runoff readily converges and mobilizes sediment, leading to debris flow disasters.

Compelling evidence from a recent study indicates that climate change has directly amplified the threat of debris flows in Central Asia’s mountainous regions, particularly within Tajikistan (Shahgedanova et al., 2024). Rising air temperatures and accelerated cryosphere degradation have altered hydrological regimes, increased meltwater contributions, and enlarged the supply of loose debris on steep slopes, thereby enhancing the likelihood of channelized debris-flow formation (Lin et al., 2025; Shen et al., 2025). Concurrently, climate-driven perturbations in precipitation patterns—specifically the increased periodicity of high-intensity, short-duration rainfall—serve as pivotal triggers for debris flows in high-relief catchments. These evolving hydro-geomorphic dynamics substantially exacerbate the susceptibility of the fragile Pamir Plateau and Gissar Mountain valleys, where inherent geological instability and an abundance of lithogenic debris predispose the terrain to mass wasting. Such climate-related impacts underscore the urgency of conducting updated susceptibility assessments to better understand debris flow dynamics under ongoing environmental change.

The stratigraphy of Tajikistan is predominantly composed of sedimentary rocks and displays pronounced lithological diversity (Fig. 2). Quaternary units are mainly represented by unconsolidated deposits composed of mud, clay, sand, gravel, alluvium, and silt. These materials are widely distributed across intermontane basins, plains, and major river valleys, where they are closely associated with modern fluvial systems and surface geomorphological processes. Paleozoic strata are extensively exposed in the eastern and southwestern mountainous regions. These formations are dominated by carbonate rocks, particularly limestone and dolomite, and are commonly interbedded with slate, siltstone, and sandstone. In some areas, Paleozoic sequences also include volcanic and intrusive rocks such as diabase, dacite, and tuff, reflecting complex tectono-magmatic activity. Mesozoic formations are primarily concentrated in western Tajikistan and some intermontane basins. Their lithological assemblages mainly consist of sandstone, siltstone, conglomerate, and limestone, with localized occurrence of evaporitic minerals such as gypsum, consistent with sedimentation under variable depositional environments. Cenozoic deposits, including Paleogene and Neogene units, are widely distributed in tectonically active foreland and basin regions. These strata are characterized by thick clastic successions composed predominantly of conglomerate, sandstone, siltstone, and clay, reflecting intensive uplift, erosion, and sediment accumulation processes. In addition, extensive Quaternary glacial deposits occur in high-altitude regions, where glaciers currently occupy approximately 6.70% of the national territory. Meltwater from these glaciers constitutes an important hydrological source during the warm season and plays a critical role in the generation of floods and debris flows.

The methodological framework, encompassing data processing, computational procedures, and modeling techniques, is illustrated in Figure 3. The research workflow is organized into five primary stages: (1) compiling a comprehensive debris flow inventory based on existing datasets, supplemented by satellite imagery interpretation and field surveys; (2) conducting multicollinearity diagnostics to select characteristic variables relevant to debris flow initiation; (3) constructing predictive susceptibility models using machine learning algorithms; (4) evaluating and validating model performance through comparative analysis; and (5) integrating risk assessments for exposed elements, specifically road density and population density.

By normalizing the contribution weights across the machine learning algorithms, the relative importance of environmental covariates was elucidated (Fig. 13). Elevation was consistently identified as the predominant deterministic factor across all three algorithms, underscoring that topographic relief provides the requisite gravitational potential energy for debris flow initiation. In contrast, variables such as catchment area, NDVI, and precipitation exhibited relatively lower importance. The secondary drivers were identified as the AI, fault density, slope, and lithology. Fault density is intrinsically linked to geological friability and rock mass fragmentation, as tectonic activity generates fractured substrates, thereby ensuring a profuse sediment supply. While slope primarily governs mass transport kinematics, lithology dictates the weathering intensity, both of which are directly coupled with the sediment budgets required for debris flow mobilization.

While extant literature frequently identifies rainfall intensity as the primary trigger for debris flows (Zhao et al., 2021), the influence of antecedent drought remains critically under-researched. However, post-drought debris flows are increasingly recognized as a global phenomenon. Similar to observations in southwestern China and Colombia (Nyman et al., 2019; Chen et al., 2020a), Tajikistan is a hotspot for drought-conditioned mass movements. This region’s semi-arid climate, characterized by desiccated summers followed by intense autumnal convective storms (Podgórski et al., 2023), facilitates high runoff coefficients and severe soil erosion.

Our findings suggest that drought exerts a more profound structural control on debris flow occurrence than instantaneous precipitation. We posit that while precipitation provides the immediate hydrological impulse, initiation is fundamentally governed by sediment availability (Liu et al., 2020). Prolonged water deficits induce desiccation cracking, enhance mechanical weathering, and attenuate soil cohesion, thereby compromising the structural integrity of hillslopes. These drought-induced fissures act as preferential flow paths, allowing rapid infiltration during subsequent storms, which triggers abrupt spikes in pore-water pressure (Hu et al., 2019; Zhong et al., 2021). Concurrently, drought-mediated vegetation degradation weakens root reinforcement and promotes the accumulation of loose proluvial materials. When extreme, short-duration precipitation follows a protracted dry spell—a characteristic climatic pulse in Tajikistan—the convergence of weakened slope stability, abundant lithogenic debris, and rapid runoff concentration creates an optimal environment for debris flow initiation. The transition from drought-induced soil weakening to rainfall-triggered mobilization (Chen et al., 2020b) represents a critical sediment-supply shift that is often overlooked in traditional susceptibility frameworks.

The catchment-scale susceptibility maps identify the Zeravshan and Vakhsh river basins as national-level hazard hotspots, providing a scientific basis for government agencies to prioritize targeted monitoring and engineering mitigation. Furthermore, the demonstrated impact of hydro-climatic coupling highlights the necessity of incorporating soil-moisture deficits and drought indices into early-warning systems, rather than relying exclusively on rainfall thresholds. Integrating real-time aridity monitoring with high-resolution meteorological forecasts would enable authorities to identify “windows of heightened vulnerability” more accurately (Chen et al., 2022; Li et al., 2024).

Several caveats warrant consideration. First, the AI used here represents a long-term climatological baseline rather than a dynamic, temporally resolved measure of antecedent moisture states. Consequently, it reflects the background aridity regime rather than event-specific hydrological fluctuations. Future research should integrate multi-temporal metrics (e.g., SPI time series or satellite-derived soil moisture anomalies) to better quantify how drought-rainfall transitions modulate hazards (Ren and Leslie, 2020). Second, while machine learning models effectively capture spatial risks, challenges regarding sample imbalance and hyper-parameter sensitivity persist (Dikshit et al., 2021). Finally, data resolution remains a decisive constraint (Zhang et al., 2019). Although 30 m resolution was utilized, it may bypass fine-scale geomorphic nuances. As higher-resolution datasets for both topography and precipitation become more accessible, the predictive fidelity and operational reliability of machine learning-based assessments are expected to improve significantly.

This study executed a systematic regional-scale assessment of debris flow hazards in Tajikistan, utilizing hydrological catchments as the fundamental mapping units. By integrating three machine learning algorithms— SVM, RF, and MLP—we evaluated debris flow susceptibility and further synthesized these outputs with population and transportation data to construct a multi-dimensional risk zonation framework. This research aims to elucidate the deterministic drivers of debris flows and provide a robust scientific foundation for disaster risk reduction in data-scarce mountainous environments.

The empirical findings demonstrate that while geological and geomorphological configurations provide foundational controls on debris flow initiation, hydro-climatic factors modulate their spatio-temporal variability. Furthermore, human activities, specifically road density and population density, exert a significant footprint on regional hazard profiles. Among the evaluated models, SVM model exhibited superior predictive fidelity and spatial discrimination, successfully capturing 56.46% of historical events within a mere 8.04% of the total land area. High-risk hotspots were predominantly clustered in the Zeravshan and Vakhsh river basins—regions characterized by the convergence of steep topography, lithological complexity, and intense anthropogenic pressure. Notably, beyond traditional topographic variables, antecedent drought conditions emerged as a pivotal factor, underscoring their critical role in preconditioning slope failure through material weakening.

Based on these outcomes, we recommend that priority should be accorded to high-risk catchments in the formulation of monitoring protocols, early warning systems, and structural mitigation strategies. Future research should prioritize the integration of dynamic hydro-climatic indicators and multi-temporal remote sensing observations to enhance model robustness. Such advancements will be essential for supporting adaptive disaster risk management and fostering socio-ecological resilience amidst accelerating environmental transformations in Central Asia.

JIA Wenjun: data curation, formal analysis, methodology, resources, software, validation, visualization, and writing - original draft; CHEN Ningsheng: supervision, project administration, and funding acquisition; XUE Yang: investigation, resources, supervision, and writing - review & editing; WANG Zhihan: data curation, software, methodology, and validation; WEN Tao: project administration, resources, and supervision; GUO Ru: validation; Safaralizoda NOSIR: investigation and resources; and Aminjon GULAKHMADOV: investigation and resources. All authors approved the manuscript.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

This study was supported by the National Natural Science Foundation of China (42361144880), the Science and Technology Program of Xizang Autonomous Region, China (XZ202402ZD0001), and the Qinghai Province Basic Research Program Project, China (2024-ZJ-904). The authors would like to sincerely thank Dr. Ishfaq GUJREE for his professional assistance in improving the English language, clarity, and readability of the manuscript.