The MYRB is located between the Toudaoguai hydrological station and the Huayuankou hydrological station in the Yellow River Basin, China (Fig. 1), covering an area of approximately 34.5 × 10⁴ km², with an elevation above sea level ranging from 85 to 3917 m. The basin is influenced by a warm temperate monsoon climate. Precipitation is unevenly distributed over space and time, with an average annual rainfall of 320 mm in the northwest to 840 mm in the southwest, mostly occurring from May to September (Sun et al., 2020). The region is predominantly covered by the Loess Plateau, which is notable for its severe soil erosion and unprecedented soil and water conservation efforts. Consequently, this region serves as a major sediment source, contributing approximately 90 % of the total sediment in the Yellow River (Chang et al., 2022; Sun et al., 2020). Since the 1970s, the Chinese government and local farmers have extensively constructed check dams in the gullies across the region. According to Zeng et al. (2024), there are now over 40 000 check dams in the MYRB, with most located in hilly and gully areas. Furthermore, a series ecological restoration projects, especially the “Grain for Green” Program (conversion of cropland on slope to forestland and grassland), have been conducted to improve vegetation cover.

In this study, we collected observed monthly SY data from 17 hydrological stations during 1970–2020 in the sub-basins of the MYRB from the Yellow River Conservancy Commission of the Ministry of Water Resources (YRCC) (Fig. 1; Table S1). The total area of the 17 sub-basins accounts for 67.9 % of the MYRB, and they have typical land cover, landform and hydrological characteristics of the region. The daily rainfall dataset at resolutions up to 0.1° sourced from a new high-quality gridded precipitation dataset (called CHM_PRE), which was developed by Han et al. (2023) using daily observations from 2839 gauges across China and surrounding areas (1970 to present). Soil data including soil organic carbon and soil particles such as sand, silt, and clay, was obtained from the SoilGrids dataset, produced by the International Soil Reference and Information Centre (ISRIC) with a resolution of 250 m (Poggio et al., 2021). The Digital Elevation Model (DEM) data, with a spatial resolution of 30 m, was sourced from the Shuttle Radar Topography Mission (SRTM-30 m) (Table 1).

The Normalized Difference Vegetation Index (NDVI) data were retrieved from two sources: the Global Inventory Modeling and Mapping Studies third-generation NDVI dataset (GIMMS-NDVI-3g) and the Terra Moderate Resolution Imaging Spectroradiometer (MODIS) Vegetation Indices (MOD13Q1). GIMMS-NDVI-3g, produced using the Advanced Very High Resolution Radiometer (AVHRR) sensor, provides NDVI data every 15 d with a spatial resolution of 8 km from 1982 to 2015 (Pinzon and Tucker, 2014). MOD13Q1 provides NDVI data every 16 d at 250 m spatial resolution from February 2000 to present. To extend time availability and ensure data comparability, the two NDVI datasets were harmonized based on overlapping periods (February 2000 to December 2015) using spatiotemporal stability analysis and statistical downscaling techniques (Huang et al., 2024) (Sect. S1). Because the purpose of this procedure was to construct a temporally consistent NDVI record, the full overlapping period was used for correction to maximize the use of available seasonal, interannual, and spatial variability. The corrected GIMMS-NDVI-3g was then assessed against MOD13Q1 during the same overlapping period. The corrected GIMMS-NDVI-3g showed good consistency with MOD13Q1 (R2=0.94, Root Mean Square Error (RMSE) = 0.053, Mean Absolute Error (MAE) = 0.038). Compared with the original GIMMS-NDVI-3g, RMSE and MAE of corrected GIMMS-NDVI-3g decreased by 46.9 % and 47.2 %, respectively (Fig. S1). These results support the use of the corrected GIMMS-NDVI-3g for deriving the C factor in RUSLE before MOD13Q1 became available in February 2000.

Land use data of five years (1975, 1990, 2000, 2010, and 2020) were obtained from the Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, with a spatial resolution of 30 m. The overall accuracy of the land use data was 86 % (Wu et al., 2024). We reclassified land use into cropland, forest, grassland, built land, waterbody, and other land use types (Table S2). The 30 m-resolution terrace dataset across China for 2000, 2010, and 2020 was generated using a two-stage random forest classification framework that integrates time-series Landsat imagery with Copernicus DEM data (Zhang et al., 2025). This model utilized the texture features of terraces, achieving a classification accuracy of 91.7 %. The vectorized check dam dataset was sourced from YRCC and Zeng et al. (2024). The dataset integrates high-resolution (0.3–1 m) Google Earth imagery from May 2016 to 2020 and an object-based classification method, achieving an overall accuracy of 94.4 % (Zeng et al., 2024). This dataset comprises spatial location and storage capacity of check dams. Although the delineation relies on recent imagery, most terraces and check dams in the basin were constructed in the 1970s–1980s, as documented in regional conservation reports.

The model simulates soil erosion, sediment delivery, and sediment trapping by check dams within watersheds through sequential processes. First, the RUSLE model was used to simulate soil erosion at a specific location on the hillslope. Second, the eroded sediment delivered to the river, namely SY, was estimated through the “RUSLE-IC-SDR” approach. Finally, in watersheds controlled by check dams, the sediment intercepted by cascading check dam systems was calculated based on the trapping efficiency (Fig. 2). On account for these processes, the watershed was divided into two parts: areas with and without check dam regulation. In areas with check dams, sediment interception is explicitly considered. In contrast, areas without check dams include only the erosion and sediment delivery processes (Fig. 2).

The conversion from IC to SDR requires calibrating the parameters IC₀ and KIC. Due to variations in sub-basin characteristics such as topography, soil, and climate (Hao et al., 2022), the optimal combinations of IC₀ and KIC differ across sub-basins. Each sub-basin was represented by 100 m grid cells, which served as the basic computational units for soil erosion, IC, SDR, and SY estimation. Sub-basins were used for calibration, goodness-of-fit assessment, and aggregation of grid-cell-level results. In this study, we conducted individual parameter calibration for 17 sub-basins controlled by hydrological stations. Initially, 70 combinations of IC₀ and KIC were tested, with IC₀ ranging from -7 to -1 (in steps of 1) and KIC from 0.5 to 5 (in steps of 0.5). Each parameter set was applied to the proposed integrated SY model in each sub-basin. This process was repeated for all the 70 combinations, yielding SY values with accounting for trapping of check dams. The observed and simulated SY were compared to evaluate the model's predictive capability using the Nash-Sutcliffe Efficiency coefficient (NSE) (Huang et al., 2024). If the optimal parameter combination fell outside the initial range, the parameter steps were adjusted based on the trend of NSE variation until the optimal combination was identified. For the remaining areas outside these 17 sub-basins, we used the average of the optimal parameters of nearby sub-basins to calculate SDR, which is used to complete the SY estimation for the entire MYRB. The simulated SY was compared with observations from 17 hydrological stations to evaluate whether the parameterized model could reproduce long-term SY dynamics at the sub-basin outlet scale. Model accuracy was quantified using NSE and R2 (Huang et al., 2024). Several SY outliers were identified using the Bonferroni-adjusted studentized residual test, Q-Q residual diagnostics, and Cook's distance, and they were excluded from the goodness-of-fit statistics. In addition, to independently assess the sediment trapping module, the simulated sediment accumulation was compared with field-surveyed sediment accumulation data from 61 check dams reported by Fang et al. (2023).

Previous applications of the traditional RUSLE-IC-SDR framework do not explicitly account for the sediment trapping of the dams (Lan et al., 2023; Schürz et al., 2020; Yin et al., 2025). Building on this context, the key focus of our study was the explicit incorporation of TE-based check dam trapping, which represents sediment interception and storage within each dam along cascading dam systems. To evaluate the impact of check dams trapping within the RUSLE-IC-SDR framework, we compared two model configurations. The first configuration activates the full TE-based trapping scheme, while the second follows the traditional RUSLE-IC-SDR setup in which sediment is routed through the landscape without sediment trapping by check dams. The combinations of IC₀ and KIC in this scenario remained identical to that in the actual check dam-regulated scenario. This design ensures that the only difference between the two simulations is whether check dam trapping is represented. The resulting SYs are denoted as SY_Trap (with sediment trapping) and SY_noTrap (without sediment trapping), and the Sediment Reduction Contribution by check dams (SRC_dam) was computed as: 10SRCdam=SYnoTrap-SYTrapSYnoTrap×100%

The check dams were primarily distributed in the central part of the MYRB, forming highly complex cascading systems (Fig. 4). The area controlled by check dams is 4.68 × 10⁴ km², accounting for 13.55 % of the MYRB. The total designed storage capacity of all check dams is 6.36 × 10⁹ m³, predominantly comprising small dams (< 10 × 10⁴ m³), which account for 68.0 % of the total number (Fig. 4a). There was 68.78 % (n=32595) of the check dams within a cascading system. Within these systems, ≥ 6-tier check dams constitute 16.65 % (n=7892) of the total count, while a subset of 0.56 % (n=265) exhibit exceptionally high class (16–27 tiers), indicating 16–27 sequentially constructed check dams along river channels from upstream to downstream (Fig. 4b).

The proposed model framework considered the effects of cascading check dams on SY. The optimal combinations of IC₀ and KIC in the model varied across sub-basins (Fig. S2). Our analysis demonstrates that models incorporating check dam sediment trapping exhibit superior model performance compared to counterparts excluding check dams (Fig. 5). Model goodness-of-fit assessment using observed SY from 17 hydrological stations in the MYRB showed that incorporating the sediment trapping of check dams could improve the NSE to 0.713 (R2=0.71) (Fig. 5a), representing a 20.0 % improvement compared to the traditional model (NSE = 0.594, R2=0.72) (Fig. 5b). In addition, the simulated sediment trapping by check dams agreed well with the observations from field survey (NSE = 0.651, R2=0.75) (Fig. S3).

From 1970 to 2020, decadal soil erosion rates in the MYRB were 52.65 ± 11.20, 56.46 ± 8.17, 52.11 ± 10.33, 47.61 ± 15.06, and 29.97 ± 7.54 t ha⁻¹ yr⁻¹ in 1970s, 1980s, 1990s, 2000s, and 2010s (Fig. S4). Overall, soil erosion rate decreased across 85.95 % of the basin. Among the 17 sub-basins, average soil erosion rates ranged from 21.19 (Zhujiachuan River basin) to 192.05 t ha⁻¹ yr⁻¹ (Qingjian River basin) (Fig. S5). The mean IC in the MYRB was -3.90 ± 1.45. IC is relatively high in the central basin, covering the hilly-gully regions (about -4 on slopes to around -2 in gullies). However, the hilly-gully region exhibited a rapid decline in IC (<-0.006 yr⁻¹) during 1970–2020. In contrast, the southern and eastern valley plains showed comparatively low IC (<-4.5), and an increasing trend of IC was observed in some areas (Fig. S6). The SDR derived from optimal IC₀–KIC combinations varied substantially among sub-basins. The Fen River basin exhibited the lowest average SDR (0.02), while the Kuye River basin (0.83) and the Huangfuchuan River basin (0.89) reported higher SDR values. Overall, SDR exhibited a declining trend (Fig. S7). In the densely dammed central basin, many check dam systems achieved a TE above 0.80. With the decreases of the effective storage capacity in these check dams, TE generally showed a declining trend (Fig. S8).

Check dams have effectively prevented sediment from being transported out of the watershed (Fig. 6). Under the influence of check dams, the 50-year average SY in MYRB was 15.42 ± 4.56 t ha⁻¹ yr⁻¹, which represents a 12.6 % reduction compared to scenario without sediment trapping by check dams (17.64 ± 5.55 t ha⁻¹ yr⁻¹) (Fig. 6). A comparison of Fig. 6a, b reveals that in the central and northern basin with relatively high SY, there are numerous patches characterized by lower SY (i.e., watersheds controlled by check dams) (Fig. S9). In these watershed areas controlled by check dams, the 50-year average SY was 16.38 ± 6.19 t ha⁻¹ yr⁻¹, representing a 50.01 % reduction compared to the scenario without sediment trapping by check dams (32.76 ± 11.97 t ha⁻¹ yr⁻¹) (Fig. 6).

During 1970–2001, the multi-year average SY in MYRB was 17.41 ± 3.61 t ha⁻¹ yr⁻¹, showing no significant changing trend (P=0.184) (Fig. 7). After 2002, the average SY decreased to 11.89 ± 3.93 t ha⁻¹ yr⁻¹, representing a 31.71 % reduction compared to 1970–2001, with a statistically significant downward trend during 2002–2020 (P < 0.001) (Fig. 7). Furthermore, the difference in SY between scenarios with and without sediment trapping by check dams has been gradually narrowed, decreasing from 14.42 % during 1970–2001 to 7.45 % in the post-2001 period (Fig. 7).

In the MYRB, the overall trend in sediment reduction contribution by check dams (SRC_dam) showed a gradual decline, with a 50-year average value of 11.80 ± 4.56 % (Fig. 8). In the first decade (1970–1979), the SRC_dam remained above 18 %. During 1980–2001, the SRC_dam was 12.56 ± 2.62 %, whereas after 2002, it significantly declined to 7.17 ± 1.19 % (P < 0.001) (Fig. 8). The SRC_dam varied widely across sub-basins, ranging from 1.00 % in Fen River basin to 52.46 % in Jialu River basin in the first decade (Fig. S10). During 1980–2001, SRC_dam in sub-basins ranged from 0.91 % (Fen River basin) to 43.65 % (Jialu River basin). After 2002, the range of SRC_dam narrowed to 0.93 % (Fen River basin) – 37.41 % (Fenchuan River basin).

The accumulated sediment trapped by all check dams in the MYRB over the past 50 years was 3.84 × 10⁹ t, occupying 41.49 % of the total designed storage capacity (Fig. 9). During the first decade, sediment accumulation rates in check dams were highest, averaging 126.24 × 10⁶ t yr⁻¹, accounting for 13.62 % of the total storage capacity (Fig. 9). Between 1980–2001 and post-2001, the average sediment accumulation rates decreased to 90.29 × 10⁶ and 33.03 × 10⁶ t yr⁻¹, respectively. Accumulated sediment during these two periods represented 21.44 % and 6.42 % of the total storage capacity, respectively (Fig. 9).

In the first decade (1970–1979), check dams were filled rapidly, with only 27.13 % of check dams retaining more than 95 % of their storage capacity, while 8.30 % had already lost over 90 %. By 2001, sediment accumulation further reduced storage capacity, with just 9.48 % of check dams maintaining over 95 % capacity and 29.92 % experiencing losses of exceeding 90 % capacity (Fig. 9). Until 2020, 55.08 % of the check dams lost more than 50 % of their storage capacity due to sediment deposition (Fig. 9). Among these, 37.47 % of the check dams lost over 90 % of their storage capacity, and only 7.43 % of the check dams accumulated less than 5 % capacity (Fig. 9).

The degree of sediment accumulation in check dams varied across different sub-basins in the MYRB (Fig. S11). Over the past 50 years, the total sediment accumulation in check dams across the 17 sub-basins ranged from 1.06 % (Fen River basin) to 78.12 % (Huangfuchuan River basin) of the total storage capacity (Coefficient of Variation, CV = 63.89 %) (Fig. S11). Additionally, within these sub-basins, 0.54 % (Qingjian River basin) to 87.94 % (Fen River basin) (CV = 135.55 %) of check dams experienced less than 5 % losses of storage capacity, while 2.78 % (Zhujiachuan River basin) to 76.39 % (Kuye River basin) (CV = 75.12 %) of check dams suffered exceeding 90 % storage capacity losses (Fig. S11).

The temporal dynamics of SRC_dam were influenced by check dam characteristics (such as number, density, controlled area, and storage capacity) and sediment yielding status (such as the magnitude of SY and its spatial distribution) (Bai et al., 2020; Sun and Wu, 2023). Prior to the implementation of vegetation restoration projects, the vegetation coverage was only 25.08 % during 1980s in the MYRB (Zhang et al., 2022), and severe soil erosion occurred in the basin (Fig. S4). The resulting high SY, combined with the initially high effective storage capacity of newly check dams, jointly contributed to a high SRC_dam (Verstraeten and Poesen, 2000). As sediment accumulates in check dams, their effective storage capacity diminishes, resulting in a decreasing TE (Fig. S8b) and SRC_dam (Fig. 8). With the implementation of “Grain for Green” program since 2000, the vegetation coverage has increased to 58.68 % recently (Zhang et al., 2022). The improved vegetation has significantly curbed hillslope soil erosion by intercepting rainfall, enhancing infiltration, and consolidating the soil (Ebabu et al., 2022). As a result, SY has markedly decreased, leading to a more stable and further lower SRC_dam over time (Sun et al., 2020).

The SRC_dam and proportion of accumulated sediment to storage capacity of check dams (SAR_dam) differed substantially among 17 sub-basins (Figs. S10 and S11). Figure 10 shows the relationships of SRC_dam and SAR_dam with check dam related factors (number of check dams, check dam density, the proportion of the watershed area controlled by check dams to the total basin area (Area_ratio)) and SY_noTrap. The SRC_dam displayed a significant linear relationship with Area_ratio (R2 = 0.80, P < 0.001) and check dam density (R2=0.76, P < 0.001) (Fig. 10a, b). In contrast, the SAR_dam was only significantly correlated with sub-basin SY_noTrap (R2=0.62, P < 0.001) (Fig. 10h), but not closely tied to factors associated with check dams. This means that sediment reduction contribution of check dams is mainly controlled by their distribution characteristics, whereas the sediment accumulation efficiency was primarily affected by the sediment yield from upstream of the check dams.

Check dams play an important role in sediment trapping within channels (Abbasi et al., 2019; Esteban Lucas-Borja et al., 2021; Piton et al., 2017). From 1970 to 2020, sediment accumulation in check dams of the MYRB reached peak rates in the 1980s and slowed significantly thereafter (Fig. 9). This was due to the rapid decline in erosion rates or the improper construction of check dams (Gao et al., 2024). The conversion of slopes to terraces and the “Grain for Green” program in 1999 significantly increased vegetation cover, leading to a significant reduction in hillslope soil erosion (Lan et al., 2023). Additionally, enhanced vegetation cover has also reduced hydrologic connectivity (Fig. S6b), consequently lowering the SDR (Abebe et al., 2023; Borselli et al., 2008; Zhao et al., 2020) (Fig. S7b). These changes decreased sediment delivered from hillslopes to channels by reducing both the sediment source and transport capacity. This facilitates sediment deposition, further regulating sediment transport processes (Piton and Recking, 2017).

We used the trade-off between SRC_dam and SAR_dam as a diagnostic indicator to compare sediment reduction effectiveness and storage occupation among sub-basins (Sect. S2). This comparison can provide useful information for the management of check dams. A relatively high SAR_dam combined with a low SRC_dam, as observed in the Kuye River, Huangfuchuan River, and Jing River basins, indicates high storage pressure and limited current sediment reduction effectiveness (Fig. 11). This pattern may be associated with dam age, reduced remaining capacity, high upstream sediment supply, construction history, maintenance conditions, or spatial placement. Therefore, these sub-basins should be prioritized for further field assessment, capacity restoration, desilting, maintenance, or optimized dam placement. In contrast, the Fenchuan River basin showed relatively high SRC_dam but low SAR_dam, suggesting that check dams still have available storage capacity and maintain effective sediment reduction under current sediment supply conditions (Fig. 11). For these basins, future efforts may no longer need to focus on check dam construction. Instead, great emphasis should be placed on hillslope soil and water conservation measures, such as vegetation restoration.

Sub-basins with both high SRC_dam and high SAR_dam, such as the Jialu River basin (Fig. 11), indicate that check dams have contributed substantially to sediment reduction but have also consumed a large proportion of storage capacity, suggesting a need for maintenance and capacity restoration. In contrast, sub-basins with both low SRC_dam and low SAR_dam, such as the Fen River basin (Fig. 11), show limited sediment reduction effectiveness and low storage occupation, which may reflect low sediment supply, limited dam-controlled area, spatial mismatch between check dams and sediment source areas, or suboptimal dam placement. These patterns should therefore be used as diagnostic indicators for further field assessment and management optimization.

By integrating the RUSLE-IC-SDR with the TE of cascading check dams, this study provided a straightforward and practical method to quantify sediment delivery from hillslopes to channels and trapping by check dams in the basin. This method can be easily applied in other regions worldwide. This study focuses on the temporal dynamics and spatial patterns of sediment trapping and reduction by cascading check dam systems. Using the RUSLE-IC-SDR method, SY hotspot in the basin can be identified. By comparing the potential SY of the areas controlled by planned check dam sites with the designed storage capacity of the check dams, we can evaluate the rationality of planned check dam systems. This method enables precise assessment of sediment retention in complex check dam systems and optimal spatial configuration of the check dam system.

Several limitations related to spatial resolution and empirical parameterization should be noted. DEM resolution may introduce uncertainty into LS factor in RUSLE and SY estimates (Lu et al., 2020). Coarser DEMs can smooth steep gullies and small channels, whereas overly fine DEMs may increase data noise, storage requirements, and computational costs (Li et al., 2025). Therefore, DEM resolution should be selected according to study scale, terrain complexity, research objectives, and available computing resources (Bircher et al., 2019; Lu et al., 2020; Dile et al., 2024). Although the simulated SY was calibrated and evaluated using records from 17 hydrological stations, fine-scale erosion and deposition patterns may still be affected by DEM resolution. Future work should conduct multi-resolution sensitivity analyses for LS, IC, SDR, and SY in representative sub-basins. The coefficient D is another uncertainty source because it reflects sediment texture and may vary with rainfall and sediment-source conditions in TE calculation (Verstraeten and Poesen, 2000). Our further analysis showed that larger D values increased long-term simulated sediment accumulation (Fig. S12). Moreover, the influence of D on TE depends on the ratio of remaining storage capacity to contributing catchment area (V/Acca), with a comparatively limited effect for those with higher ratios. During extreme rainstorms, enhanced erosion may deliver more coarse particles to check dams (Zhu et al., 2024), increasing the effective D value and causing a fixed fine-sediment value to underestimate TE, particularly for check dams with low V/Acca. Extreme events may also rapidly reduce remaining storage capacity, accelerate TE decline (Chen et al., 2025), and increase overtopping or failure risks (Bai et al., 2020; Esteban Lucas-Borja et al., 2021). Future work should incorporate event scale particle size and hydrodynamic observations to develop time-varying TE parameterization.

Check dams are well known to rapidly halt channel incision and stabilize gully beds, yet these geomorphic adjustments are not directly captured by the current model structure (Piton et al., 2017). Future research should focus on incorporating gully erosion, bank erosion, and bed incision modules to better represent the full suite of geomorphic processes influenced by check dams (He et al., 2026). In addition, human water-sediment regulation may introduce uncertainty into SY simulation. Regulated runoff and sediment processes may affect observed sediment yield but are not fully represented in the current model. Future work should incorporate human water-sediment regulation to better separate natural erosion and sediment delivery dynamics from management induced sediment changes. Considering the growing interest in carbon burial in dammed sediments, coupling sediment retention modelling with carbon dynamics is another important direction (Yao et al., 2022). Despite the above limitations, the proposed framework substantially improves the ability to diagnose sediment dynamics in large, highly engineered basins and provides a foundation for more comprehensive modelling of cascading check dam networks.