A systematic review of artificial intelligence approaches for sedimentation modeling in hydrological systems

Soil erosion and sedimentation are important environmental challenges, particularly within data-scarce and climate-sensitive regions. Accurate sediment yield prediction is essential for mitigating reservoir siltation and supporting sustainable land and water management. This study systematically reviews and synthesizes recent applications of Artificial Intelligence (AI) in sedimentation modeling, clearly outlining its objectives, screening scope, key findings, and novelty. From 277 screened publications, 16 studies were selected following the PRISMA 2020 guidelines. The results show that hybrid AI models, which integrate physical process-based equations with data-driven algorithms (ANN and LSTM), consistently outperform traditional and standalone machine learning approaches, improving R² by up to 15% and reducing the RMSE by 25-40%. The integration of Google Earth Engine (GEE) enables multi-source remote sensing, climatic, and hydrological data fusion, enhancing spatiotemporal sediment monitoring and reproducibility. Despite these advances, challenges remain regarding data quality, computational cost, and model transferability. This review contributes a standardized and reproducible framework to assess AI-based sedimentation models and to outline future research directions towards scalable, interpretable, and climate-resilient sediment management.

1. INTRODUCTION

Sedimentation in reservoirs and watersheds is a critical global issue that threatens water resource sustainability, ecosystem integrity, and economic stability. Sediment accumulation reduces reservoir capacity, degrades water quality, alters aquatic habitats, and increases maintenance costs through dredging and infrastructure repair(Lee et al., 2022; Wang et al., 2018). These impacts are further exacerbated by human-induced pressures, including deforestation, high-input agriculture, urban expansion, and mining, which intensify soil erosion and increase sediment delivery to rivers and reservoirs, thereby complicating sediment management strategies(Borrelli et al., 2020; Trinugroho et al., 2022; Xiong & Leng, 2024). Furthermore, climate-driven variability in precipitation intensity and spatiotemporal patterns exerts a strong control on sediment dynamics by affecting erosion mechanisms and sediment transport pathways, thereby complicating watershed and reservoir management (Panagos & De Rosa, 2023; Trinugroho et al., 2022).

Traditional sedimentation models, including the Revised Universal Soil Loss Equation (RUSLE), coupled with Sediment Delivery Ratios (SDR), have long provided foundational frameworks for estimating soil erosion and sediment fluxes at the plot and watershed levels(Ahmad et al., 2022; Bassairate et al., 2021; Gourfi et al., 2018; Kanito et al., 2023). However, these empirical and semi-empirical approaches frequently rely on simplified assumptions that inadequately capture the nonlinear and complex interactions governing sediment generation and deposition across diverse landscapes under evolving climatic conditions(Heckmann et al., 2018; Yan et al., 2022). This limitation is particularly evident when scaling predictions for large, heterogeneous watersheds that are experiencing rapid changes in land use and climate(Andualem et al., 2023; Hajigholizadeh et al., 2018). In this context, nature-based solutions, such as agroforestry, vegetative buffer strips, and terracing, have shown promising results in reducing surface runoff and soil erosion while providing economic co-benefits to local communities. Studies have demonstrated that integrating these measures, particularly on steep and very steep slopes, reduces sediment yield and enhances agricultural productivity and long-term land resilience(Karyati et al., 2021; Koch et al., 2019).

Artificial intelligence (AI) offers powerful alternatives for modeling nonlinear sedimentation. Methods like Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forests, Long Short-Term Memory (LSTM) networks, and hybrid frameworks capture complex relationships that are often missed by conventional models(Meshram et al., 2020). For instance, ANN architectures have successfully predicted sediment yields in data-scarce Himalayan basins with significantly greater accuracy than RUSLE-SDR approaches(Sarkar et al., 2017), whereas SVM integrations have reduced uncertainty in sediment load forecasts during Mediterranean flash floods(Asaly et al., 2023). Furthermore, the integration of AI with advanced geospatial analysis platforms, including Google Earth Engine (GEE), enables highly detailed, long-term monitoring of sediment dynamics using satellite-derived indices, including the Normalized Difference Vegetation Index (NDVI) and Modified Normalized Difference Water Index (MNDWI)(Jha et al., 2024; Yang et al., 2022). This powerful combination allows for the continuous identification of erosion hotspots, mapping of sediment sources and sinks, and early detection of landscape changes across continental scales, fundamentally transforming watershed diagnostics and management(Bezak et al., 2024; Uroš et al., 2024).

This review synthesizes insights from 16 key studies selected from a comprehensive pool of 277 peer-reviewed publications, focusing on AI-driven sedimentation modeling, enhanced by satellite remote sensing and high-resolution climate data. Unlike previous reviews that have primarily emphasized algorithmic performance or theoretical development, this synthesis critically evaluates how AI methods, particularly hybrid models that integrate physical sedimentation principles using machine learning algorithms, advance the prediction of sediment yield, timing of sediment mobilization, and spatiotemporal dynamics under changing environmental conditions. Our analysis reveals that these integrated approaches consistently outperform standalone empirical and physically based models, with reported increases in prediction accuracy ranging from 25 to 45 percent, as supported by several quantitative assessments (e.g., Allawi et al., 2023; El Bilali et al., 2020; Meshram et al., 2020; Yadav et al., 2022). These studies indicate that hybrid AI-physical frameworks, such as ANN-GA, LSTM-CEEMDAN, and MUSLE-ANN, reduce the RMSE values by 25-40% and increase R^²by up to 15% compared to standalone Machine learning models. Additionally, this review highlights the importance of the operational scalability and transferability of AI techniques across diverse geographic, climatic, and socio-economic contexts, an aspect that has often been underexplored in earlier syntheses.

Although promising advances have been made, several challenges remain. Key issues include inconsistencies in data quality and availability, high computational resource demands, the need for large and diverse training datasets, and limitations in model transferability across different environmental settings(Zhao et al., 2024). These challenges can be addressed by developing standardized validation protocols and modular frameworks that enable the efficient integration of diverse datasets (e.g., climate projections, land-use maps, and soil data), and improved transparency in model architecture and parameterization(Batista et al., 2021). Moreover, incorporating stakeholder engagement and local knowledge into AI model development and validation processes can enhance their relevance, social acceptance, and practical applicability.

While several reviews have examined AI applications in hydrology, few have specifically focused on sedimentation processes by integrating physical, empirical, and hybrid AI models into a unified analytical framework. Existing reviews often emphasize either algorithmic performance or erosion mapping but rarely analyze how AI-driven hybrid frameworks (e.g., ANN-MUSLE and LSTM-SWAT) contribute to improving sediment yield prediction, physical interpretability, and large-scale operational transferability of the models. Therefore, this review fills a clear research gap by providing the first systematic synthesis linking AI-based sedimentation modeling with remote sensing, climate data, and process-based approaches. Its novelty

.....