This review employed a structured methodology to identify and analyze studies on AI-based sediment-yield prediction. A focused research question guided the selection process, followed by a comprehensive literature search using predefined inclusion and exclusion criteria to determine each study’s relevance. The following step involved data extraction from the selected studies, focusing on the key variables and methodological aspects. Finally, the findings were synthesized to highlight the contributions, limitations, and trends of AI applications in sediment modeling. In line with this methodology, we examine the specific AI techniques used, along with their performance and limitations. Figure 1 illustrates the overall review process, from formulating the research question to synthesizing the findings to be discussed.

This workflow followed the PRISMA 2020 guidelines to ensure transparent and reproducible screening and selection of studies. Each step is numerically reported in the corresponding flowchart (Fig. 1).

To identify relevant research articles, a systematic review was conducted using two databases: Google Scholar and Scopus, over an eight-month period from February to October 2024. A comprehensive set of search terms (Fig. 2) was employed to ensure inclusion of all pertinent resources. These terms were initially chosen based on the authors' expertise and subsequently refined by incorporating keywords extracted from relevant studies.

The research methodology was systematic, and targeted searches were performed using keywords that comprised three or four terms. The study began with a thorough analysis of various studies on modeling and forecasting sedimentation in reservoirs and watersheds, with a particular focus on approaches that utilize artificial intelligence (AI). This approach is more sophisticated than simple sediment yield estimation and encompasses various methodologies, including artificial neural networks (ANN), support vector machines (SVM), and random forests (RF) (Jooshaki et al., 2021).

The primary objective was to ensure that all AI-based methods used to model sediment dynamics, hydrological and climatic impacts, and water resource management were included in the systematic review. Articles examining the correlations among precipitation, soil erosion, and discharge variations were also included. These factors are directly linked to sediment modeling because they influence key phenomena, such as the transport and deposition of solid particles.

Furthermore, the consideration of spectral indices such as NDVI, MNDWI, and AWEI, combined with the use of geospatial tools like Google Earth Engine (GEE), provides a more comprehensive representation of the input data used in the models. The integration of these environmental variables with clustering approaches (e.g., k-means clustering) and temporal analysis offers a valuable framework for understanding the complex dynamics of hydrological systems.

The initial search yielded 277 publications, which were subsequently screened for evaluation purposes. A comprehensive review of the reference lists of the selected publications was then conducted. The reviewed documents included journal articles, conference proceedings, books, review papers, theses, and online databases. This approach ensured that the analysis was based on a comprehensive and diverse range of sources.

The adopted strategy was designed to ensure the reproducibility and transparency of the results. Exact search strings and combinations were documented and consistently applied across both databases to minimize bias. Specifically, the search in Scopus was conducted using the following query:

TITLE-ABS-KEY (("sediment yield" OR "sediment load" OR "reservoir siltation") AND ("artificial intelligence" OR "machine learning" OR "deep learning" OR "neural network" OR "hybrid model") AND ("watershed" OR "reservoir" OR "catchment")), while the same Boolean operators were used in Google Scholar searches. Duplicate records were automatically removed through metadata matching (title, DOI, and year) prior to manual screening.

The screening and selection processes followed the PRISMA 2020 framework. Figure 1 presents a PRISMA-style flow diagram summarizing the number of records identified, screened, excluded, and retained. At each stage, the number of studies was reported as follows: 277 records were initially identified, 45 duplicates were removed, 232 records were screened based on titles and abstracts, 53 were excluded due to irrelevance, and 179 full-text articles were assessed for eligibility. Finally, 16 research papers met all inclusion criteria and were retained for synthesis. This transparent workflow ensures that each exclusion step can be replicated using the same criteria. Table 1 summarizes the exclusion criteria applied during the study selection process, which ensured the inclusion of only AI-based sedimentation studies.

Following a rigorous selection process, 16 articles were retained for analysis. These included original research or systematic reviews covering various approaches, such as the use of satellite data (NDVI, MNDWI) and cloud-based geospatial processing platforms such as Google Earth Engine (GEE), to advanced modeling techniques, including artificial neural networks (ANN), support vector machines (SVM), and random forest (RF). This selection provides a comprehensive framework for analyzing the contributions and trends of sedimentation forecasting and management. Table 2 summarizes the selected studies by outlining their study areas, study types, key findings, and data sources.

In accordance with PRISMA 2020 guidelines, screening and data extraction were performed to ensure methodological transparency and reproducibility. Consistent inclusion and exclusion criteria were applied throughout the process, focusing on studies that utilized artificial intelligence–based approaches for sedimentation modeling.

Data were extracted from 16 studies meeting the inclusion criteria for this review, covering key aspects of sedimentation modeling and prediction, including publication year, study site or watershed, and principal research objectives. The extracted information also included the artificial intelligence methods applied—such as Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forests (RF), and Long Short-Term Memory (LSTM)—along with the model architectures developed, data preprocessing techniques (including the use of spectral indices such as NDVI, MNDWI, and AWEI), model inputs (e.g., precipitation, NDVI, and streamflow), and model outputs (such as sediment production in tons). In addition, the time periods covered by the datasets, the limitations identified in each study, and the data sources used, including Landsat satellite imagery, climate databases, and watershed data, were systematically recorded to provide a comprehensive synthesis of the selected studies.

Table 3 summarizes the main objectives, methodologies, and key results of selected studies. Overall, hybrid and deep learning models exhibited the strongest predictive capabilities, followed by ensemble and classical ML algorithms. Specifically, physically coupled hybrid frameworks (e.g., SWAT-LSTM and MUSLE-ANN) achieved the highest average R^² values (≈0.94-0.97) and the lowest RMSE (<0.12), highlighting their robustness in modeling complex sediment processes. Ensemble models (e.g., RF–bagging and stacking) also demonstrated high accuracy and stability across various climates.

In contrast, classical models, such as ANN and SVM, although widely applied, exhibited lower stability under heterogeneous hydrological conditions. To streamline the results, only the aggregated performance metrics are discussed below, and detailed numerical values are reported in Table 4. Table 4 summarizes the hierarchy of model performance, ordering the algorithms from most to least accurate and grouping them by environmental context to provide a concise comparative overview.

These results highlight that although ANN remains widely applied, hybrid and LSTM-based models deliver the highest accuracy and stability, particularly in complex or data-scarce hydrological environments. Comprehensive data extraction provided a complete overview of the scope of application of AI-based methods in sediment management and hydrological processes in the field. Furthermore, it assists in assessing the various methodologies, model performance, and remaining gaps in this field of study. Figure 3 illustrates the number of studies conducted for each model used in sedimentation modeling, showing the distribution across various AI methods.

The occurrence category is related to the frequency and patterns of sedimentation events. Sedimentation events are influenced by several factors, including precipitation intensity, soil type, vegetation cover, and topography. Most studies analyzed defined occurrences in a binary manner (yes/no), linking them to critical conditions, such as intense runoff episodes or specific land-use practices. These analyses are essential for identifying areas at high risk of sediment accumulation and for developing appropriate management strategies, such as soil conservation measures.

Severity analysis enables a more comprehensive assessment of the magnitude and effects of sedimentation relative to the occurrence category. AI-based approaches, including artificial neural networks (ANNs) and support vector machines (SVMs) (Jooshaki et al., 2021), have been employed to assess maximum sediment loads, deposition rates, and reductions in reservoir capacity. The key variables were the maximum sediment load (tons), sediment deposition rate, and the impact on the capacity of the hydraulic infrastructure. This information is crucial for planning dredging operations and assessing the economic and environmental consequences of sedimentation.

This period category is concerned with forecasting the likelihood of significant sedimentation events at specific times. This includes forecasting seasonal periods of high sediment discharge or peaks associated with major climatic events. Models based on LSTM networks and time series have proven effective in predicting these periods using historical precipitation and flow data. Such forecasts enable proactive management, including the preparation of reservoirs for significant sediment input and the initiation of impact-reducing interventions.

The objective of the spatial distribution category was to identify areas within a watershed or reservoir where sediment deposition is most likely to occur. A combination of AI methods with geospatial data (NDVI, MNDWI, the AWEI indices, and digital elevation models) was employed to map erosive zones and sediment deposition areas. These studies employed platforms including Google Earth Engine (GEE) and geographic information systems (GIS) to integrate data on land use, topography, and hydrology. Table 5 categorizes the methodologies applied to address different sedimentation themes, such as sediment yield prediction, erosion mapping, and reservoir volume-loss estimation. The results are presented in the form of detailed maps that highlight priority areas for targeted interventions, such as reforestation or the construction of control dams.

Figure 4 illustrates the evolution of research priorities within these categories. Initially, studies focused on the occurrence and severity of sedimentation. However, since 2010, there has been growing interest in the period category, driven by the availability of high-resolution temporal data and advanced AI models.

Furthermore, around 2015, there was a notable shift towards emphasizing spatial distribution, highlighting the increasing role of geospatial approaches in sedimentation management. More recently, studies on sedimentation occurrence have resurged due to the heightened frequency of extreme climatic events and their associated impacts on watersheds and reservoir systems.

Figure 5 presents the conceptual structure linking the major components of AI-based sediment modeling. Input datasets, including climatic (precipitation and temperature), topographic (DEM and slope), and spectral indices (NDVI, MNDWI, and AWEI), were fed into three major AI model categories: (1) classical machine learning (e.g., ANN, SVM, RF), (2) hybrid physical-AI models (e.g., MUSLE-ANN, SWAT-LSTM), and (3) ensemble approaches (e.g., RF-bagging, stacking). These models generate diverse outputs, such as sediment yield (SY), spatial sediment deposition, and reservoir capacity loss, which inform management strategies and decision support systems.

Google Earth Engine (GEE) is a cloud-computing platform that offers an efficient environment for processing and integrating multi-source datasets, including Landsat, Sentinel, and climate archives. Rather than focusing on technical steps, this review emphasizes how GEE-enabled processing transforms sedimentation modeling through improved spatiotemporal scalability, reproducibility, and open-access integration (Amani et al., 2020; Kazemi Garajeh et al., 2024; Tamiminia et al., 2020).

GEE facilitates large-scale analysis by automating the preprocessing of satellite images (e.g., cloud masking and compositing) and generating environmental indices such as NDVI, MNDWI, and AWEI over multi-decadal periods. This capability ensures consistent and reproducible datasets that are comparable across different basins and timeframes (Chen et al., 2023; Pakdel-Khasmakhi et al., 2022).

Cloud-based computation allows near-real-time monitoring of reservoir dynamics, enabling the continuous detection of sedimentation trends and land-use changes without requiring high-performance local infrastructure. Researchers can reproduce analyses directly through shared scripts and datasets, promoting transparency and collaboration in science (Amani et al., 2020; Kazemi Garajeh et al., 2024).

In summary, GEE enhances sedimentation modeling by automating data preprocessing and spectral index generation, integrating climatic, hydrological, and topographic data across multiple spatial scales, improving spatiotemporal scalability and reproducibility, and supporting open-access, collaborative modeling frameworks.

Figure 6 illustrates the conceptual framework of the GEE workflow employed in sedimentation modeling. The Google Earth Engine (GEE) is the foundation of this methodology, offering a robust platform for processing, analyzing, and integrating satellite data from various sources (Tamiminia et al., 2020).

The key benefits of GEE include large-scale data processing, as it provides a comprehensive solution for managing and preprocessing vast volumes of satellite imagery, such as data from the Landsat 5, 7, and 8 missions covering the period from 1990 to 2021. Through automated procedures, GEE enhances imagery by removing artifacts such as clouds and shadows, thereby simplifying and streamlining complex preprocessing steps.

The platform enabled the rapid calculation of spectral indices, as GEE facilitates the efficient generation of essential indices for identifying and monitoring water surfaces (Chen et al., 2023). These include the Normalized Difference Vegetation Index (NDVI) for distinguishing vegetation from other land-cover types, the Modified Normalized Difference Water Index (MNDWI) for accurate water surface detection, and the Adjusted Water Index (AWEI) for reducing the effects of shadows and mixed pixels. The cloud-based computing capabilities of GEE ensure the consistent and rapid production of these indices, thereby supporting coherent spatiotemporal analyses over multiple decades.

Automated spatiotemporal analysis tools in Google Earth Engine (GEE) facilitate continuous observation of water surface dynamics, allowing the detection of trends and anomalies in reservoirs (Pakdel-Khasmakhi et al., 2022).

The generation of training samples for classification purposes was facilitated using GEE through the implementation of K-means clustering, which produced robust training samples subsequently used to train supervised models. These included Random Forest (RF) models, which effectively capture complex relationships between spectral indices and water surfaces, and Support Vector Machine (SVM) models, which are fully compatible with the platform and enable accurate classification of water areas. By leveraging the computational capabilities of GEE, this approach optimizes data processing while improving both efficiency and accuracy of the classification results.

Furthermore, in addition to satellite data, climate and hydrological variables were integrated to enable a more comprehensive analysis of water reservoirs. These parameters included Mean Annual Precipitation (MAP), which directly influences reservoir inflow and sediment transport processes, and Mean Annual Temperature (MAT), which affects evaporation rates and overall hydrological variability. The GEE facilitates the integration of these data with generated spectral indices, thereby facilitating the analysis of complex interactions between climate, spatiotemporal trends, and water surface variations.

To effectively develop and validate hydrological models, it is essential to have comprehensive information on the spatiotemporal variations of reservoirs, including dates, locations, magnitudes of surface changes, and bathymetric data. The integration of these datasets with modern platforms, such as Google Earth Engine (GEE), enables a comprehensive and detailed analysis of large-scale water surface dynamics.

Satellite data, particularly from the Landsat missions, play a critical role in the long-term monitoring of water surface variations in reservoirs, as the Landsat archive available through Google Earth Engine (GEE) offers extensive spatiotemporal coverage over several decades. These data enable the production of accurate water surface maps and spectral indices, including the Normalized Difference Vegetation Index (NDVI) to monitor surrounding vegetation dynamics, the Modified Normalized Difference Water Index (MNDWI) for reliable identification of water surfaces, and the Adjusted Water Index (AWEI) to minimize interference from shadows and vegetation. Through GEE, these indices can be computed efficiently at large spatial and temporal scales, supporting precise temporal analyses of water surface variations.

The inputs were classified into two main categories: raw data and derived data. Raw data consisted of fundamental datasets such as precipitation, air temperature, water level, and river discharge, while derived data included variables calculated from these primary datasets, notably vegetation and water indices obtained from satellite imagery.

Several types of raw data are commonly used in water reservoir studies. Air temperature data are key inputs, as temperature variations directly influence evaporation rates and water inflow, particularly during dry seasons or snowmelt periods. Precipitation data collected from weather stations are essential for estimating the volume of water entering reservoirs, and climate models integrated within GEE use these data to better assess the impacts of climatic events on reservoir water management. In addition, discharge and water level data measured at hydrometric stations are used to estimate reservoir inflows and outflows, with discharge variability being a critical factor in reservoir management due to its direct effect on water availability.

The derived data obtained from the raw data include various indices that are essential for predictive models. These indices were calculated from parameters such as temperature, precipitation, and water level, providing valuable insights into the reservoir behavior.

The most commonly used indices include cumulative freeze degree days (AFDD) and thaw degree days (ATDD), which are derived from air temperature and are widely applied to assess ice and snow dynamics in watersheds, as these processes directly influence reservoir water levels during melting periods. Snowfall and Snow Water Equivalent (SWE), derived from precipitation data, are also key variables, particularly in regions where snowmelt contributes significantly to reservoir inflow. In addition, the Antecedent Precipitation Index (API), which represents cumulative rainfall prior to a given event and serves as a proxy for soil moisture conditions, is frequently used to forecast reservoir water levels by providing an effective measure of water accumulation before major hydrological events such as floods or abrupt water-level changes.

These indices are employed in AI models to forecast reservoir dynamics, particularly to identify periods of flooding or drought risk. By integrating these indices with satellite data in GEE, dynamic maps of reservoirs can be created, and their evolution can be simulated under various climate scenarios.

In addition to climatic and hydrological data, bathymetric data are essential for accurately predicting reservoir evolution. These data reveal the depth and shape of the reservoir's bottom, allowing the calculation of the water volume stored at any given time. Digital Elevation Models (DEMs) derived from these data were used to estimate water volumes in reservoirs and predict the impacts of sedimentation and topographic changes. Integrating these bathymetric data with climatic and hydrological data in GEE enables the simulation of long-term reservoir evolution and the prediction of scenarios under different climatic and hydrological conditions.

As illustrated in Figure 7, hydrometric parameters such as discharge, water level, and precipitation are the most commonly used raw inputs for modeling reservoir evolution. The introduction of climate data, such as air temperature and precipitation indices, has marked a significant shift in this field of study. These climate variables are crucial for assessing reservoir variations because of their direct impact on the filling and draining of reservoirs. Another significant development in reservoir dynamics modeling is the integration of bathymetric and watershed topography data using Digital Elevation Models (DEMs) and other geospatial variables, facilitated by software like the Google Earth Engine (GEE).

The selection of input variables plays a critical role in AI models applied to water reservoir dynamics predictions. Many studies have used correlation coefficients to select input variables because of their simplicity and effectiveness. However, alternative approaches, such as regression methods, including stepwise and backward regression, are also commonly used to refine the selection of input variables. More recent strategies include techniques such as the two-step filter rapper (Chen & Chen, 2015) and the Weight of Evidence method, which incorporates geospatial approaches to further refine data selection, particularly for variables such as vegetation indices, air temperature, and water volume stored in reservoirs. These methods offer new ways to improve forecasts by combining historical and real-time satellite data.

Data normalization is an important step in ensuring the stability and performance of AI models (Sujon et al., 2024). The most commonly used method is standard normalization, which converts data to a uniform scale, reduces redundancies and inconsistencies, and thus improves the accuracy of water reservoir dynamics forecasts. This technique is particularly useful in studies dealing with climate and hydrological data collected from different sources, such as satellites and weather stations. Normalization also reduces the impact of outliers, which can distort the data distribution and lead to errors in model predictions. Additional approaches, such as time-series normality and homogeneity methods, were used to ensure data quality and forecast robustness (Sujon et al., 2024).

Outliers and gaps in datasets are major challenges in data preprocessing, particularly when dealing with large climate and hydrological datasets. These anomalies can bias the model results and increase the prediction uncertainty. Several methods are used to detect these anomalies, such as box plots and the F-test, which help identify and deal with outliers. A more modern approach is to replace outliers with values calculated based on historical data to ensure data consistency, while minimizing bias. Various techniques have been developed to fill gaps in datasets, such as the use of the robust Kendall-Theil line, application of regularized expectation-maximization algorithms, and integration of remote sensing data from satellites, such as MODIS. Autoregressive algorithms and interpolation from other time series or neighboring station data are also used to deal with gaps, especially in areas where measurements are incomplete or missing.

To ensure that the time series data were reliable and of high quality, additional tests were conducted to verify criteria such as normality, homogeneity, and independence of the data. These checks ensured that the data used to model the reservoir dynamics met the necessary standards for robust and reliable analysis.

Machine learning (ML) is the most commonly used approach for predicting reservoir dynamics, particularly water levels, flow management, and sedimentation (Nourani, 2014). Popular machine-learning models include Artificial Neural Networks (ANN), Support Vector Machines (SVM), and algorithms (Jooshaki et al., 2021), such as k-nearest neighbors (KNN) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS). These models can analyze complex time series and predict critical events for reservoir management by relying on historical data, meteorological measurements, and flow data. In particular, ANN can capture complex nonlinear relationships and has been used to estimate flooding and sedimentation events with high accuracy, as shown in recent studies on reservoir sediment yield prediction.

Hybrid approaches use a combination of machine learning models and metaheuristic algorithms to create more robust and accurate models. These approaches are used particularly for complex tasks, such as predicting ice jam breakage in reservoirs and sedimentation. Particle Swarm Optimization (PSO) combined with ANN has achieved high predictive performance, with models reaching up to 95% accuracy in predicting flood events. For instance, while ANNs excel in handling nonlinear relationships in sediment yield prediction, hybrid approaches have been shown to be more effective in complex scenarios, achieving high accuracy. However, their computational requirements limit their scalability in data-scarce regions. Another interesting hybrid method is the combination of Long Short-Term Memory (LSTM) networks with Adaptive Signal Decomposition Techniques (CEEMDAN), which has shown significant improvements in modeling complex reservoir dynamics.

However, these hybrid models remain sensitive to data quality and availability, particularly in specific geographical or climatic contexts. Sites with low data availability or heterogeneous data can pose challenges in obtaining generalizable results, which remains a significant limitation of hybrid approaches.

Ensemble approaches improve the overall accuracy and reduce errors by combining the predictions from multiple models. Random forests (RFs) and bagging are commonly used in this context and have shown good performance in predicting water levels and sedimentation in reservoirs. These models combine multiple decision trees into a random forest, allowing them to handle complex and nonlinear data that characterize reservoir hydrological dynamics. However, despite their power, ensemble approaches have limitations in predicting event locations, which remains a challenge for current models.

Framework approaches integrate physical and AI models to improve reservoir forecasting, particularly in complex environments where hydrological and climatic data interact. For example, a hybrid model combining conceptual models with neural networks was used to assess the flows associated with ice jams and sedimentation. Recently, studies have combined Landsat and TIGGE data with hydraulic models, demonstrating improved capabilities for simulating flooding and sedimentation risks in reservoirs. However, these approaches are still under development and have not been widely applied to predict event occurrence or timing, which remains an area of exploration.