Machine Learning for Spatiotemporal Prediction of River Siltation in Typical Reach in Jiangxi, China

Abstract

Accurate forecasting of river siltation is essential for ensuring inland waterway navigability and guiding sustainable sediment management. This study investigates the downstream reach of the Shihutang navigation power hub along the Ganjiang River in Jiangxi Province, China, an area characterized by pronounced seasonal sedimentation and hydrological variability. To enable fine-scale prediction, we developed a data-driven framework using a random forest regression model that integrates high-resolution bathymetric surveys with hydrological and meteorological observations. Based on the field data from April to July 2024, the model was trained to forecast monthly siltation volumes at a 30 m grid scale over a six-month horizon (July–December 2024). The results revealed a marked increase in siltation from July to September, followed by a decline during the winter months. The accumulation of sediment, combined with falling water levels, was found to significantly reduce the channel depth and width, particularly in the upstream sections, posing a potential risk to navigation safety. This study presents an initial, yet promising attempt to apply machine learning for spatially explicit siltation prediction in data-constrained river systems. The proposed framework provides a practical tool for early warning, targeted dredging, and adaptive channel management.

1. Introduction

Inland water transport serves as a cornerstone of sustainable and integrated multimodal transportation systems, offering substantial advantages, such as low operational costs, a high freight capacity, reduced energy consumption, and a minimal environmental impact [1]. China, with its long-standing reliance on river-based trade, has developed an extensive inland waterway network that plays a pivotal role in advancing national strategies, such as the Belt and Road Initiative, the Yangtze River Economic Belt, and the construction of modern transport infrastructure [2,3].

Jiangxi Province, located in southeastern China, holds historical and strategic significance in the country’s inland waterway system [4]. With the ongoing expansion of water transport infrastructure, the province has established a progressively high-grade channel network. Navigation channels are vital arteries within this system, and their uninterrupted and efficient operation is essential for ensuring reliable waterborne transport. However, maintenance activities such as dredging and channel excavation—though necessary for meeting navigability standards—have introduced significant alterations to local hydrological regimes [5,6]. These human interventions disrupt the natural flow–sediment equilibrium, increasing sediment transport complexity and intensifying the burden of post-dredging channel maintenance [7]. Among the major challenges in waterway management, siltation remains persistent and spatially heterogeneous. The evolution of sediment deposition in alluvial rivers is influenced by a variety of natural and anthropogenic factors, including river morphology, seasonal flow variation, sediment properties, and human interventions [8,9,10,11]. Particularly in flood seasons, back-siltation can occur rapidly and unevenly, posing risks to channel depth and navigation safety.

To address this issue, the traditional approaches, such as physical modeling and numerical simulation (e.g., MIKE21 and Delft3D), have been widely applied [12,13]. These models are grounded in physical mechanisms and offer robust explanatory power, but are often limited by high computational costs, reliance on detailed boundary conditions, and sensitivity to parameter calibration. In highly dynamic and data-limited environments, their applicability is constrained. As an alternative, machine learning (ML) methods have gained attention for their ability to model complex nonlinear processes using observational data, without relying on explicit physical assumptions [14].

Recent studies have demonstrated the potential of ML techniques, such as artificial neural networks (ANNs), support vector machines (SVMs), and random forests (RFs), in sedimentation prediction for estuaries, reservoirs, and ports [15,16,17]. For example, the SVM has been used to model sediments from the Molinos River in Bogotá, Colombia [16]; RF, the SVM, and the ANN were jointly applied to model suspended sediment in Tigris River, Baghdad [17]. These approaches have shown improved predictive accuracy and reduced dependency on domain-specific parameter calibration. However, most existing studies focus on coarse spatial resolution or temporal aggregation, which limits their ability to capture the spatial heterogeneity and localized dynamics of sediment accumulation. This hinders their practical application in detailed sediment risk assessment and channel maintenance planning.

To address this gap and align with the evolving trend of intelligent waterway management, this study explores the potential of machine learning for fine-scale spatial prediction of sediment deposition. The study area is the downstream channel of the Shihutang navigation power hub on the Ganjiang River, which is characterized by dynamic sedimentation and strong seasonal hydrological variability. An RF regression model is developed using multi-source inputs, including bathymetric surveys, hydrological observations, and meteorological statistics, to predict monthly siltation volumes at a 30 m spatial resolution during the period from July to December 2024. By revealing localized siltation patterns under data-limited conditions, this approach aims to inform more precise dredging and risk mitigation strategies. The results contribute to the broader goal of intelligent and adaptive inland waterway management in the face of increasing environmental uncertainty.

2. Materials and Methods

2.1. Study Area and Data

The study area is situated along the Shihutang navigation power section of the Ganjiang River in Ji’an City, Jiangxi Province, China (Figure 1). This river reach lies within a typical alluvial floodplain, characterized by soft surface sediments and deep fluvial deposits formed through prolonged erosion and deposition processes. The channel exhibits a pronounced meandering “S”-shaped morphology, and the surrounding terrain is predominantly flat. The curved channel geometry induces spatial variability in hydrodynamic conditions. In particular, at meander bends ranging from 30° to 60°, the flow velocity increases by up to 1.3 times relative to the upstream inflow, resulting in zones of enhanced shear stress. These areas are prone to intensified riverbed erosion, sediment resuspension, and localized siltation. Additionally, the Xiajiang Hydropower Project, situated approximately 95 km downstream, has disrupted the natural flow–sediment equilibrium, altering sediment transport pathways and deposition patterns in the reach.

Given these geomorphological and hydraulic complexities, the Shihutang section of the Ganjiang River serves as a representative case for investigating sedimentation dynamics in inland waterways. It provides a practical foundation for modeling siltation processes and developing data-driven approaches to support channel maintenance, sediment risk assessment, and navigation safety.

The spatial extent of the river channel and the study area boundary were delineated using satellite imagery solely for mapping and boundary definition purposes. This study employs three primary categories of data: bathymetric survey data, hydrological observations, and meteorological statistics. A comprehensive overview of all the data types, their sources, collection periods, and analytical applications is provided in

Table 1. Data sources and descriptions.

Data Type	Source	Description
Bathymetric data	Field survey	High-resolution riverbed elevation data collected via single-beam echo sounding at multiple time points; used to calculate grid-scale siltation volumes.
Upstream water level	Shihutang hub operational data	Daily upstream water level at the lock head; influences flow velocity and the sediment-carrying capacity of incoming water.
Downstream water level		Daily downstream water level; affects sediment deposition and flushing capacity in the lower channel segment.
Upstream threshold depth		Water depth at the upstream gate threshold; reflects potential sediment accumulation in critical navigation control zones.
Downstream threshold depth		Water depth at the downstream gate threshold; helps assess the degree of sedimentation impacting downstream navigability.
Inflow discharge		Total daily inflow volume; serves as a proxy for upstream hydrodynamic energy and potential sediment supply.
Outflow discharge		Total daily outflow volume; used to analyze sediment export potential and downstream back-siltation dynamics.
Vertical clearance		Distance between water surface and lock structure; indicates available clearance for vessel passage and reflects siltation impact on navigational safety.
Meteorological data	Public weather websites	Daily maximum temperature, minimum temperature, sunshine duration, and wind speed; influence flow regimes and sediment dynamics.
Remote sensing imagery	Satellite data	Optical imagery used to delineate the spatial extent of the river reach.

.

Bathymetric surveys were conducted using single-beam echo sounding on three occasions—1 April, 17 May, and 3 July 2024. High-density point measurements were systematically collected across the study area to capture detailed riverbed morphology. The raw depth data were subsequently resampled into a 30 m × 30 m grid, generating gridded-scale bathymetric maps. By combining these measurements with concurrent water level observations from the Shihutang hub, grid-based water depth maps were derived. These datasets enabled a spatially explicit assessment of riverbed elevation and its temporal evolution (see Figure 2).

The hydrological data consist of daily measurements aggregated at the reach scale, with one value per day for the entire study area. The variables include upstream and downstream water levels, threshold depths at both lock gates, inflow and outflow discharges, and navigational clearance heights. These parameters jointly characterize the hydraulic conditions governing sediment entrainment, transport, and deposition processes. The meteorological data were manually collected from publicly available weather websites based on the geographic location of the study area. Four variables—maximum temperature, minimum temperature, wind speed, and sunshine duration—were compiled at a daily resolution, with one observation per variable per day for the entire reach. These factors influence river discharge and sediment dynamics by modulating evapotranspiration, runoff, and atmospheric forcing.

2.2. Feature Construction

To enable grid-level prediction of river siltation, a structured set of predictive features was constructed by integrating the spatial bathymetric data with the temporally aggregated hydro-meteorological variables. These features capture both the static morphological characteristics of the riverbed and the dynamic environmental conditions influencing sediment transport and deposition.

The input variables were organized into three categories:

Bathymetric feature: Each 30 m × 30 m grid cell was assigned an initial riverbed elevation value derived from field surveys conducted at the beginning of each sampling period (i.e., 1 April or 17 May 2024). This variable provides essential spatial context related to sediment deposition potential.

Hydrological features: A total of 28 features was extracted from daily measurements of the upstream and downstream water levels, threshold depths, inflow and outflow discharges, and vertical clearance. For each of the seven base variables, four statistical indicators—mean, maximum, minimum, and standard deviation—were computed over the corresponding time window to summarize the temporal variability in hydraulic conditions.

Meteorological features: Four weather variables—maximum temperature, minimum temperature, sunshine duration, and wind force—were similarly summarized using the same four statistical measures, resulting in 16 additional features.

In total, each sample (grid cell) is represented by 45 predictive features, comprising one spatial variable (initial bed elevation), 28 temporally aggregated hydrological descriptors, and 16 meteorological descriptors. This feature design enables the model to learn from both spatial heterogeneity and temporal dynamics under realistic data constraints.

2.3. Modeling of Grid-Level Siltation Using a Random Forest

This study proposes a structured, data-driven framework for grid-level siltation prediction by integrating temporal hydro-environmental dynamics with riverbed topography, utilizing a robust ensemble learning approach. The objective is to estimate sediment accumulation over two distinct periods in 2024. The target variable—siltation volume for each interval—was calculated as the difference in riverbed elevation between the start and end of each period. The predictive features include riverbed topography at the beginning of each time window and the corresponding hydrological and meteorological conditions. For the first period (from 1 April to 17 May 2024), the model used bathymetric data from 1 April, along with environmental observations from that period. Similarly, the second period (from 17 May to 3 July 2024) incorporates bathymetry from 17 May and its corresponding hydro-meteorological data.

To account for the differing lengths of the two periods, the total siltation volumes were normalized by the number of days in each interval, resulting in average daily siltation as the modeling target. This normalization is defined as follows:

$$\overline{S_l}={\displaystyle \frac{S_i}{T_i}}$$

(1)

where $\overline{S_l}$ denotes the average daily siltation during the $i$-th time period, $S_i$ is the total observed siltation volume, and $T_i$ is the number of days in the respective period.

A total of 746 valid grid cells was obtained as model samples, including 185 from the 1 April–17 May period and 561 from 17 May–3 July. These samples represent spatially heterogeneous sedimentation responses, each associated with its initial riverbed elevation and environmental features. The full dataset was randomly split into training (80%) and testing (20%) subsets for model development and evaluation. The selected algorithm is the RF Regressor, an ensemble learning method that constructs multiple decision trees using bootstrap samples and random feature selection at each node. RF was chosen for its well-established effectiveness in handling structured tabular data, its robustness to overfitting, and its capacity to model nonlinear relationships without extensive parameter tuning, making it particularly suitable for environmental prediction tasks with mixed resolution and limited datasets. The final prediction is the average output from all the individual trees:

$$\hat{y}={\displaystyle \frac{1}{N}}\sum _{j=1}^N{T}_j(x)$$

(2)

where $\hat{y}$ is predicted daily siltation, $T_j(x)$ is the output of the $j$-th decision tree, and $N$ is the total number of trees.

To enhance model robustness and predictive accuracy, hyperparameter optimization was conducted using Bayesian Optimization with 5-fold cross-validation. This probabilistic optimization framework efficiently explores the hyperparameter space by modeling the objective function and iteratively selecting promising configurations. The tuned parameters include the number of estimators (n_estimators), maximum tree depth (max_depth), minimum samples per leaf (min_samples_leaf), and the number of features considered when splitting (max_features). Model performance was evaluated using three regression metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²). These are defined as follows:

$$MAE={\displaystyle \frac{1}{n}}\sum _{j=1}^N\left|y_i-\widehat{y_l}\right|$$

(3)

$$MSE={\displaystyle \frac{1}{n}}\sum _{j=1}^N{(y_i-\widehat{y_l})}^2$$

(4)

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-\widehat{y_l})}^2}{{\sum }_{i=1}^n{(y_i-\overline{y})}^2}}$$

(5)

where $y_i$ is observed siltation, $\widehat{y_l}$ is the predicted value, $\overline{y}$ is the mean of observed values, and $n$ is the number of test samples. These metrics collectively evaluate the prediction accuracy, sensitivity to large errors, and the model’s explanatory power, respectively.

2.4. Spatial Prediction of Monthly Siltation

The trained random forest model was further extended to perform monthly-scale spatiotemporal predictions of siltation volumes from 3 July to 31 December 2024, covering the six-month forecast period. The model produced grid-level predictions for each individual month, resulting in a continuous series of monthly siltation distribution maps across the study area (see Figure 3). As the model was trained using average daily siltation volumes, a post-processing step was applied to ensure temporal consistency. Specifically, for each target month, the model first predicted the mean daily siltation, which was then multiplied by the number of days in that month to obtain the total monthly siltation volume. This conversion ensures alignment between the training and inference phases, while producing temporally scaled outputs.

To simulate the evolving riverbed topography over time, an iterative update mechanism was adopted. For each month, the predicted siltation volume was combined with the most recent bathymetric surface to estimate the end-of-month riverbed elevation. For example, the July prediction was applied to the bathymetric map from 3 July 2024 to derive the updated topography for 31 July. This new surface was then used as the input terrain for the August prediction, yielding a bathymetric map for 31 August, and so forth. Through this month-by-month update, a series of temporally evolving riverbed surfaces was generated, enabling dynamic visualization and analysis of sediment deposition processes throughout the forecast period. This iterative modeling strategy offers a more realistic representation of cumulative siltation effects and supports practical applications, such as navigation risk assessment, adaptive dredging planning, and channel evolution analysis, under changing hydro-environmental conditions.

3. Results

3.1. Initial Siltation Dynamics and Environmental Conditions

Based on the bathymetric surveys conducted on 1 April, 17 May, and 3 July 2024 (see Figure 4), the underwater terrain in the study area exhibits distinct spatial patterns. The riverbed is generally shallower along the channel margins, indicating higher elevation near the riverbanks. Additionally, the terrain shows a gradual elevation increase from downstream to upstream, consistent with a natural channel slope and fluvial sediment transport dynamics.

By comparing the elevation changes between the survey dates, the siltation volumes for two specific periods—from 1 April to 17 May and from 17 May to 3 July—were calculated. Overall, the majority of 30 m × 30 m grid cells exhibited siltation depths of less than 1 m. The spatial extent of bathymetric data was more concentrated in the 1 April survey, and a larger proportion of grids exceeded 1 m of siltation during the first period (April–May) compared to the second period (May–July). Spatially, siltation distribution revealed strong heterogeneity, with localized clusters of both high and low values observable in each time window, suggesting the influence of local hydraulic structures or morphological features.

The environmental conditions also varied significantly between the two periods (see Figure 5). From April to May, the average water levels were higher, and the air temperatures were lower than during May to July. The mean water level was 49.59 m in the first period and 48.29 m in the second. Meanwhile, the average maximum air temperature increased from 26.13 °C to 29.66 °C. Furthermore, both the water level and air temperature fluctuations were greater in the first period, indicating more dynamic hydro-meteorological conditions early in the season. After early July, the water levels exhibited a gradual downward trend, with daily fluctuations progressively decreasing. The air temperature continued to show diurnal variability, but a clear seasonal decline was observed heading into the winter months. These seasonal changes in hydro-meteorological conditions are closely linked to the spatial and temporal dynamics of sediment deposition and resuspension observed in the study area.

3.2. Predicted Spatiotemporal Patterns of Siltation from July to December 2024

The siltation prediction model developed using underwater topography and time-varying hydrological and meteorological features demonstrates a moderate predictive performance. The model evaluation yields an MAE of 0.24, an MSE of 0.19, and a coefficient of determination (R²) of 0.58. While the model accuracy could be further improved with more extensive training data, the results indicate that this data-driven approach is effective in capturing the spatiotemporal characteristics of siltation under constrained data conditions. To assess the robustness of the modeling framework, we additionally compared the performance of the RF model with two alternative machine learning algorithms: the ANN and eXtreme Gradient Boosting (XGBoost). The RF model achieved the best predictive accuracy, with an R² of 0.58, outperforming XGBoost (R² = 0.41) and the ANN (R² = 0.32). These findings reinforce the suitability of RF for sediment prediction tasks involving mixed resolution features and limited samples, where ensemble tree-based models tend to generalize more reliably than deep or boosting-based methods.

To further enhance model interpretability, a feature importance analysis based on the RF algorithm was conducted (see Figure 6). The results reveal that the initial bed elevation is by far the most influential predictor, reflecting the dominant role of local channel morphology in sediment deposition. In contrast, the hydrological and meteorological features—especially temperature, wind force, and sunshine duration—exhibit comparatively less, but non-negligible importance, indicating that broader environmental fluctuations also modulate siltation dynamics. These findings align with our understanding of physics and highlight the potential of integrating spatially fine-grained terrain data with regionally aggregated environmental indicators in data-driven sediment modeling.

Based on the trained model, the monthly siltation volumes were forecasted from July to December 2024 (see Figure 7 and

Table 2. Statistical summary of model-predicted monthly siltation volumes, July–December 2024.

Siltation	July	August	September	October	November	December
Mean	0.55	0.50	0.53	0.51	0.51	0.48
Standard deviation	0.41	0.34	0.48	0.33	0.36	0.38
Minimum	0.06	0.10	0.08	0.13	0.11	0.18
Maximum	4.00	4.29	4.30	3.82	3.88	3.77
Total	767.49	685.59	736.13	710.37	699.91	665.12

). The results reveal an overall increasing trend in sediment accumulation during the summer months, with an average siltation depth of approximately 0.5 m per month. In terms of the total monthly siltation volume, the highest values were observed in July and September, while the lowest occurred in December. This temporal pattern is consistent with seasonal hydrological fluctuations, particularly higher inflows and sediment transport activity during the wet season (July–September), followed by reduced flow energy and sediment supply in the dry season (October–December). Despite a relatively stable mean depth (~0.5 m), the variation in total volume reflects dynamic changes in the spatial extent of siltation-prone zones.

Spatially, areas with siltation depths exceeding 1 m are dispersed across the study area, exhibiting no clear spatial clustering. This suggests that localized hydraulic conditions (e.g., eddies, backflows, and riverbed morphology) exert significant influence over sediment deposition. Notably, the maximum predicted depth at certain locations exceeds 4 m, particularly during August and September, indicating critical hotspots for potential navigational hazards or maintenance dredging. In contrast, during the winter months, the total siltation volumes decline, and areas with low siltation levels become more concentrated, especially in upstream segments of the navigation channel. This indicates a period of relative sediment stability or diminished transport dynamics, likely due to less river discharge and sediment supply during the dry season.

3.3. Navigation Risk for Cumulative Siltation and Declining Water Levels

From July to December, the combined influence of cumulative siltation and a gradual decline in water levels results in significant hydro morphological changes within the study area (see Figure 8). The model predictions suggest that these changes can lead to riverbed exposure and bank erosion, particularly in zones where siltation accumulates more rapidly, thereby threatening the navigability of the main channel.

The water levels in July are generally sufficient to maintain an adequate depth and width for safe navigation. However, by August, early signs of bank erosion and sediment emergence become apparent along the edges of the navigation route, leading to the narrowing of the navigable corridor. This trend intensifies over time; the sediment-affected areas gradually expand, and by December, large portions of the upstream channel are projected to become non-navigable, disrupting waterway connectivity and posing risks to vessel operation.

It is important to note that Figure 8 only distinguishes the areas with water levels above or below zero, serving as a basic threshold. However, in reality, navigability requires a minimum effective water depth to ensure vessel clearance. Therefore, the actual area affected by restricted navigation is likely to be larger than what is visually represented in this figure. This underlines the fact that even partial water-level reduction, when compounded by siltation buildup, can significantly impair channel functionality.

This evolving situation highlights the urgent need for timely sediment management and water level regulation measures. Without proactive intervention—such as targeted dredging, flow regulation, and adaptive navigation planning—the risk of seasonal navigation failure in this section of the Ganjiang River will increase, particularly during low-flow periods.

4. Discussion

The findings of this study demonstrate that sedimentation dynamics in the Shihutang reach of the Ganjiang River are highly responsive to both seasonal hydro-meteorological conditions and channel morphological characteristics. The proposed RF regression model achieved moderate predictive accuracy, effectively capturing the magnitude of and spatial variability in river siltation processes. Although the model’s performance could be further enhanced through the inclusion of more extensive training data and higher-resolution features (e.g., real-time flow velocities and sediment concentrations), its ability to generate reliable and interpretable results under limited-data conditions validates the practical value of machine learning approaches in fluvial sediment prediction.

The spatiotemporal patterns predicted by the model reveal clear seasonal contrasts. During the wet season (from July to September), the siltation volumes increased markedly, with a few grid cells showing accumulations exceeding 4 m. While these may include isolated outliers, they indicate the potential emergence of high-risk siltation zones, especially under intensified rainfall and elevated temperatures, which accelerate sediment transport and deposition. In contrast, the dry season (from October to December) was characterized by lower siltation rates and more spatially concentrated low-value zones, especially in the upstream section of the channel. These seasonal differences highlight the heterogeneity and temporal variability inherent in riverine sedimentation processes, driven largely by climatic and hydrological fluctuations.

One of the most critical insights from this study concerns the impact of cumulative siltation combined with declining water levels on navigation safety. The model projections suggest that beginning in August, sediment buildup along the channel margins progressively narrows the effective navigable width. By December, extensive shoaling and bed exposure are expected in the lower reaches, posing a high risk of navigation failure if mitigation measures are not implemented in a timely manner. Notably, the water level thresholds presented in the analysis likely underestimate the actual navigability constraints, as effective channel clearance requires a minimum depth that exceeds zero-value elevation.

From a management and planning perspective, these results underscore the urgent need for proactive sediment monitoring, targeted dredging operations, and flexible water level regulation, particularly in the high-risk zones identified by the model. Moreover, integrating such data-driven prediction models into real-time waterway management systems could significantly enhance the capacity for early warning, risk forecasting, and adaptive long-term planning, especially under climate variability and during increasing anthropogenic disturbances.

Several limitations should be acknowledged. First, although the bathymetric data were collected at a high spatial resolution, the hydrological and meteorological variables were spatially uniform daily aggregates, limiting the model’s capacity to reflect fine-scale hydrodynamics. Second, important physical drivers, such as sediment grain size, concentration, substrate type, and dredging history, were not included due to data unavailability. Third, the model was trained on observations from April to June (relatively high-flow months), but applied to predict July–December (transitional-to-low-flow months), which may introduce the extrapolation uncertainty. These factors may contribute to the model’s moderate R², and its reduced sensitivity to localized or extreme siltation events. Fourth, the six-month predictions were not directly validated against post-survey observations or actual dredging records, limiting the ability to assess real-world accuracy.

Despite these limitations, this study provides an important initial step toward data-driven, spatially explicit sediment prediction in inland rivers. The consistency between the predicted patterns and the known seasonal trends supports the model’s practical utility. Moreover, the framework is transferable and can be adapted to other rivers using locally sourced bathymetric and environmental data from field measurements, remote sensing, and hydrodynamic simulations. Future work will focus on incorporating more physical variables and high-resolution flow indicators to develop physics-informed machine learning models. In addition, expanding the dataset to cover a broader range of hydrological and sedimentary conditions will allow for more robust training, validation, and scenario testing under diverse flow regimes.

5. Conclusions

This study developed a data-driven method using random forest regression to predict monthly siltation in the Shihutang reach of the Ganjiang River. By combining underwater topography with the hydrological and meteorological data, the model captured the key spatial and seasonal siltation patterns with reasonable accuracy. The results indicated a clear increase in siltation from July to September, followed by a decline in winter, with upstream navigability notably affected by the year end. The findings highlight the potential of machine learning for sediment risk assessment and channel management. Future work should consider incorporating additional physical variables and real-time data to improve the prediction performance and practical applicability.