Runoff and sediment simulation of terraces and check dams based on underlying surface conditions

In the past few decades, the Loess Plateau has undergone large-scale underlying surface changes. A large number of soil and water conservation measures have been constructed, which have affected the runoff and sediment status in the region. How runoff and sediment status respond to underlying surface changes is the key to quantitatively evaluate the effect of water and sediment reduction by soil and water conservation measures in flood events. We selected check dams and terraced fields, which account for a large proportion of soil and water conservation measures as assessment objects and constructed a runoff-sediment model combining traditional physical mechanisms and deep learning to simulate and analyze flood events in a typical basin of the Loess Plateau. The results show that the simulation effect of model is good. The relative error of runoff is within 15%, average Nash–Sutcliffe efficiency coefficient is 0.86, and the relative error of soil loss is within 30%. Check dam system in the Chenggou River Basin can intercept 55.61% of the runoff and 47% of the soil loss in the basin on average, and terracing can reduce the runoff by 10.54% and the soil loss by 33.8%.


Introduction
The surface soil of Loess Plateau is significantly eroded by soil and water, and a large amount of sediment enters the Yellow River as soil erosion occurs. Since the 1970s, to solve this problem, China has implemented a series of ecological construction projects on the Loess Plateau (Huang et al. 2013;Hu et al. 2019). By 2012, 37,130 km 2 terraces had been built on the Loess Plateau (Dang et al. 2020), and by 2015, the number of check dams had reached 56,422 (Liu et al. 2017). Climate change and the implementation of ecological construction projects on the Loess Plateau have significantly decreased the runoff and sediment of the Yellow River in the past decades (Gao et al. 2016a).
The runoff of the Yellow River basin began to decrease in different degrees from the 1970s, especially in the 1990s. The change of tributaries in the upper reaches of the Yellow River is not significant, but in the middle reaches of the Yellow River, the runoff of Wuding river, Wei River and Fenhe River decreased significantly, and the runoff reduction of Fenhe River reached 56.95% in 2010 (He 2017). The sediment transport on the Loess Plateau began to decrease in the 1950s, and most significantly in the 1980s and 1990s. Compared with that before 1969, the annual sediment discharge of Helong section of the Yellow River in the 1970s, 1980s and 1990s decreased by 68.5%, 48.9% and 57.1% (Yao et al. 2013). Many studies have shown that check dam and terraced fields play an important role in reducing water and sediment in the Yellow River basin. According to the Yellow River Water Conservancy Research Institute, in the middle reaches of the Yellow River, the efficiency of check dam in reducing runoff and sediment reached 90.7% in the 1980s and 70.8% in the 1990s (Lu et al. 2006). According to the Shaanxi Provincial Bureau of Soil and Water Conservation, terraced fields can reduce 70 to 95 percent of runoff and 90 1 3 22 Page 2 of 18 to 100 percent of soil erosion (Zhang 2014). It can be found that check dam and terraced fields have a great influence on runoff and soil erosion. This phenomenon has attracted extensive attention of scholars from all over the world in the hope of quantitatively analyzing the impact of check dam and terraced fields on river basin water and sediment.
The simplest research methods are hydrology, soil, and water conservation methods. Gao et al. (2016b) used hydrological method to study the changes of water and sediment in the Yanhe River Basin from 1950 to 2012 and believed that human activities reduced the runoff by 67% and the sediment by 80%. Ran (2016) used the soil and water conservation method to analyze the amount of water and sediment reduction in the middle Reaches of the Yellow River during 1970-1996, and found that soil and water conservation measures reduced runoff by 4.6% and sediment by 22.9%. However, these two methods are of limited use in quantitative evaluation. The hydrological method cannot estimate the effects of different measures, and the soil and water conservation method cannot reflect the effects of each factor on the process of water and sediment transport.
The hydrological model is an important way for scholars to quantitatively evaluate the runoff and sediment reduction benefits of soil and water conservation measures. Bhave et al. (2014) selected several basins in different climate patterns and used MEAP model to analyze the benefits of check dam and believed that the water reduction capacity of check dam would be reduced by 40% in the next 30 years. Chen et al. (2014) used the TANK model to simulate the terrace's water storage capacity and flood control capacity in two heavy rainfalls, and the simulation results showed the excellent capacity of terraced paddy fields for flood control under flooding conditions. SWAT model (Abebe and Gebremariam 2019), the digital Yellow River integrated model (Shi et al. 2016), and various water-sediment coupling models (Xia et al. 2015) have also been studied and applied in this field. These models may use cell meshwork and computation of physical processes, but there are certain generalizations. The physical process of water and sediment is a highly nonlinear process that may cause an increase in the model's error.
It is worth noting that many studies have used different methods in the same watershed, and the results of quantitative analysis are not completely consistent or even very different. For example, taking Huangfuchuan Basin as the study area, Shi et al. (2016) simulated the runoff from 1999 to 2012 with the digital Yellow River integrated model and found that key dam reduced the runoff by 39%, while Li et al. (2017) used SWAT model to study in the same period and concluded that the key dam reduced the runoff by 65.2%. There are many reasons for the discrepancy; the most direct reason is all models have certain errors, and another important reason is that the change of underlying surface conditions caused by soil and water conservation measures leads to the change of runoff production mechanism in the corresponding region.
The runoff generation process can be divided into saturation-excess runoff generation model and infiltrationexcess runoff generation model, which depends on rainfall characteristics, soil moisture characteristics, and underlying surface characteristics of the basin . However, considering the spatial distribution of meteorological elements and the heterogeneity of underlying surface conditions, multiple runoff processes may occur simultaneously in a basin. At present, most water-sediment models have different calculation parameters under different underlying surface conditions, but there is no difference in the calculation methods of runoff mechanism. Therefore, when runoff-sediment model may be consistent with actual runoff-sediment process, obtained parameters can only generalize the overall situation of the watershed, but cannot be accurate to different underlying surface types. Moreover, basin confluence has a complicated process, which is made more complicated by the special rules of water storage and discharge of soil and water conservation measures. For example, most small check dams do not have clear drainage regulations, and amount of upstream runoff interception by terraces is difficult to calculate. The emergence of deep learning method brings a breakthrough to this problems. The data-driven models focus on the statistical relationship between input and output data and do not consider the physical mechanism of the hydrological process, but rather establish a mathematical analysis of the time series and use the given sample to discover the statistical or causal relationship between the hydrological variables (Xu et al. 2021). Although the simulation process of the data-driven model has no hydrophysical basis, it can accurately predict the numerical process of runoff. If the physical mechanism is complex and cannot be fully understood, the black box model can be used as a better breakthrough tool. Combining the data-driven model in deep learning with the traditional physical mechanism for hydrological prediction is also a research hotspot at present (Kratzert et al. 2018).
In this study, a small watershed in the Loess Plateau, the Chenggou River Basin, is considered as the study area. Based on fully discriminating check dams and terraces, the watershed is divided into HRU, and dominant runoff-producing mechanism of various underlying surface types is analyzed. Considering the complex and unique water storage and drainage mechanism of soil and water conservation measures, LSTM was used to calculate the confluence, and CSLE model was added to calculate the amount of soil loss. Traditional physical mechanism and deep learning are combined to achieve more accurate and reasonable water and sediment calculation. By setting different soil and water conservation measures scenarios, the contribution of different soil and water conservation measures to reduce runoff and sediment yield was quantitatively simulated.

Methods and material
Methods TO accurately and quantitatively analyze the effects of soil and water conservation measures on water and sediment, it is necessary to start from the runoff mechanism and reasonably identify the impacts of soil and water conservation measures on runoff mechanism caused by the transformation of underlying surface. The data-driven model is used to deal with the nonlinear calculation that is difficult to solve. The amount of soil loss caused by each rainfall-runoff process is calculated with the reasonable sediment yield formula (Fig. 1).
In this study, the hydrological response units (HRU) were divided according to the watershed's underlying surface information: slope, land use, vegetation cover, soil and water conservation measures, and prone runoff process of each HRU was determined. To calculate each HRU evaporation and soil water content, combined with the underlying surface information to identify the runoff generation process comprehensively, the HRU with different underlying surface conditions adopted different runoff generation parameters and calculation methods. Using the runoff yield of each HRU unit and the actual flow process of the basin as input, the long-short term memory (LSTM) neural network is trained to simulate the confluence process of the basin. Finally, rainfall and flow information were used as input, and the sediment modulus was calculated using the storm-based CSLE (Fig. 2).

Identification of dominant runoff process
The dominant runoff process (DRP) is a process that contributes most to runoff (Schmocker et al., 2007). Scherrer and Naef (2003) proposed a systematic classification process for the dominant runoff process in the basin, allowing for the combination of different soil, vegetation cover, topsoil, subsoil, and dividing the dominant runoff process into four types. Based on the results of other studies and the actual situation of the Loess Plateau, this paper formulated standards for HRU classification based on soil, slope, vegetation coverage, land use, and two soil and water conservation measures including terraced fields and check dams (Fig. 3).
The key of the study is to add the analysis of the dominant runoff process of check dam and terraced field. It is generally believed that the soil near the river remains moist for a long time, so the dominant runoff process is saturationexcess runoff (Zhao 2019). Because the check dam is built on the river channel, it can be considered that the dominant runoff process of the check dam reservoir area is also saturation-excess runoff. With the deposition of silt, the slope of the dam reservoir area will gradually decrease. In addition, check dam are often combined with plant measures to further enhance the effect of soil and water conservation. Terraces directly change the slope of the underlying surface, and it can store floodwater in flood season, reduce peak discharge, increase soil water content, and supplement river runoff in non-flood season. Studies by Llorens et al. (1997) showed that terraced fields promote soil saturation and runoff generation was saturation-excess runoff.
Plain areas with gentle slopes can retain rainfall, allowing rainfall to penetrate and reduce the formation of surface runoff. On steep slopes, runoff yields will be high due to the  rapid migration of surface runoff. The higher the vegetation coverage, the more obvious the interception effect of plants to water, enhance the infiltration, enhance the soil's water storage capacity, and become more prone to saturation-excess runoff. Based on past research, when the slope of forestland is less than 3 degrees (Antonetti et al. 2016), grassland and arable land is less than 5 degrees (Naef et al. 2002) can the saturation-excess runoff occur. The main purpose of this paper is to distinguish the two dominant flow generation mechanisms of over infiltration and storage. Many types of slope classification will lead to more parameters, so this paper only takes 3 degrees and 5 degrees as a boundary. In the correlation between VFC and runoff coefficient in the Loess Plateau region, points are relatively concentrated in the VFC 0-0.4 and 0.7-1 (Liu et al. 2015). Therefore, this research divides VFC with 0.4 and 0.7 as the boundary. Saturation-excess runoff may occur when the VFC of grassland, farmland and shrub is greater than 0.7. For forestland, it is considered that saturation-excess runoff may occur when VFC is greater than 0.4. The dominant runoff process of residential land is considered an infiltrationexcess runoff.
In ArcGIS software, grid layers of various underlying surface information are constructed, and the underlying surface information is superimposed according to the process shown in Fig. 3, so as to obtain watershed HRU and judge the dominant runoff process of each HRU.

Runoff calculation
After judging the dominant runoff process under different underlying surface conditions, it is necessary to adopt appropriate flow generation calculation methods for different runoff generation mechanisms.
For the HRU, the dominant runoff process is infiltrationexcess runoff generation, the depth of infiltration-excess runoff is determined by rainfall intensity and surface infiltration capacity. In this model, distribution curve of infiltration capacity Luo et al. (1992) is used to calculate the runoff depth. The calculation formula is as follows: where f c is a final or equilibrium capacity (m/h), f 0 is the initial infiltration capacity (m/h), and k is a constant representing the rate of decrease in f capacity.
As for the HRU prone to saturation-excess runoff generation in the basin, we need to determine when conditions are met and how to calculate the runoff yield in each period. In the model calculation, we set that infiltration-excess runoff generation would occur in this area at the initial rainfall stage, and saturation-excess runoff generation would not begin until the soil was full of water storage. The model uses the storage capacity curve (Hu et al. 2005) to calculate the saturation-excess runoff. The calculation formula is as follows: where W ′ and W ′ m are the point soil moisture capacity and maximum point soil moisture capacity, is the area fraction in which soil moisture capacity is less than or equal to W ′ , W 0 is soil water storage, and n is the empirical index.
These two methods of runoff generation calculation have been referred in many papers (Cao 2015;Hu et al. 2005), so this paper was not elaborate.
In the infiltration process, upper soil stores water first, and the lower soil does not begin to store water until the upper soil reaches saturation.
where W 0,t and W 0,t+1 are the soil water storage on days t and t + 1 , WU is soil moisture content of an upper layer, WL is soil moisture content of lower layer, WM is water storage capacity, WUM is water storage capacity of upper layer, and WLM is water storage capacity of lower layer.
The two-layer evaporation model is used to calculate the evaporation, and the calculation formula is as follows: where E m is the evaporation capacity, kc is the evaporation conversion coefficient, E 0 is pan evaporation, EU is upper evaporation, and EL is lower evaporation.

Sediment calculation
Universal Soil Loss Equation (USLE) is an empirical model for sediment calculation widely used globally. Liu et al.
(2010) extended the model to make the model's parameters more in line with the actual situation in China and defined it as Chinese Soil Loss Equation (CSLE). However, CSLE model is used to predict the total annual soil loss. Shi et al. (2018) revised the model again and proposed the stormbased CSLE. Considering the serious gully erosion on the Loess Plateau, the gully erosion factor G (Jiang et al. 2005) was introduced, which can be expressed as follows: where A is event soil loss ( t∕ha ), R is runoff depth (mm), E k is total kinetic energy of a rainstorm ( MJ∕ha ), I 30 is maximum 30 min intensity ( mm∕h ), K is soil erodibility factor ( t h MJ −1 mm −1 ), L is slope length factor, S is steepness factor, B is biological measure factor, E is engineering measure factor, T is tillage measure factor, G is gully erosion factor, a and b are empirical coefficients, a and b are empirical coefficients, e t , v t , and t r are the unit rainfall energy ( MJ ha −1 mm −1 ), the rainfall volume (mm), and the rainfall intensity ( mm h −1 ) during time period t, respectively.

Confluence calculation
The whole basin is composed of many irregularly shaped HRU, and the confluence of each HRU is not continuous medium movement, and the characteristics of underlying surface will also affect the confluence process, which is complicated by the special rules of water storage and discharge of soil and water conservation measures. Large-and medium-sized check dams generally have designed storage capacity and discharge water according to designed flood discharge. However, most of the small check dams have no clear discharge index in the design, which greatly affects the confluence calculation. Terraces reduce and store runoff generated by themselves and intercept upstream runoff, which is hard to calculate and often overlooked (Liu et al. 2014). So, it is difficult to use simple mathematical formulas to explain how much water can be stored and under what conditions the water can be discharged. In this case, the confluence calculation driven by physical mechanism is not completely consistent with the actual situation.
Therefore, instead of the traditional confluence method of physical mechanism, LSTM was chosen for confluence calculation in this study. In this method, the input R t of LSTM is runoff depth and actual discharge of each HRU hour by hour, and the output is simulated discharge. By training and learning a certain amount of data, the LSTM model will automatically compare the relationship between flow generated by each HRU and the flow in time series, find the mathematical relationship contained therein, and use the relationship for flow prediction. Therefore, this study believes that LSTM can capture the direct correlation between HRU flow generation and outlet section flow in the total multiple data and obtain reasonable results even though the underlying surface conditions of the basin are diverse, and the confluence process is complex. Figure 4 illustrates LSTM. The detailed algorithm can be found in Xu et al. (2020).

Parameters
There are 16 main parameters of the model. According to the different meanings of parameters and their effects on different model structures, the parameters are divided into runoff and confluence parameters and sediment yield parameters. See Table 1 for details.

Model evaluation
In this study, the performance of different models is evaluated by statistical error measures, including the Nash-Sutcliffe efficiency coefficient (NSE) and relative error (RE) (Nash and Sutcliffe 1970;Hu et al. 2018). The closer the two coefficients are to 1, the better the simulation effect of the model is and the smaller the error is. The mathematical expressions of these metrics can be described as follows: where Q 0 (m 3 /s) and Q c (m 3 /s) represent the discharge of observed and simulated hydrographs, Q c is mean value of the observed discharge, and n is data point number.
where R 0 (mm) and R c (mm) represent the runoff depth of the observation and simulation, Q P0 (m 3 /s) and Q Pc (m 3 /s) represent the peak discharge of the observation and simulation, and A 0 (t/ha) and A c (t/ha) represent the soil loss of the observation and simulation, respectively.-

Study area
The geographical coordinates of the Loess Plateau are between 100°52′-114°33′ E and 33°41′-41°16′ N, covering a total area of 642,000 km 2 . In this area, loess is deep and distributed in a large area, soil erosion resistance is poor, rainstorm intensity is high, vegetation is sparse, soil and water loss is serious, and ecological environment is bad. In 2006, Chenggou River Basin was listed as the first batch of small watershed dam system demonstration project monitoring in the Loess Plateau, and a large number of soil and water conservation measures have been built (Wang et al. 2013). This study took this basin as a typical study area to quantitatively analyze the influence of check dam and terraced fields on runoff production process. The Chenggou River Basin is located in the northwest of Anding District, Dingxi City, Gansu Province. It is 45 km from the urban area. It is located between E104°14′15″-104°28′31″ and N35°41′7″-35°35′10″. It belongs to the fifth subregion of the yellow soil in the hilly gully region. The basin covers a total area of 161.37 km 2 , and annual average temperature is 6.3 °C. The annual average precipitation is 380 mm, the maximum annual precipitation is 721.8 mm (1967), and the minimum annual precipitation is 248.7 mm (1969). The precipitation was mainly concentrated in July, August, and September, accounting for 67% of the annual precipitation. The annual potential evapotranspiration exceeded 1500 mm, which is four times that of precipitation. The soil types are mostly yellow loam soil and black mature soil, which are mainly composed of fine sand and silt, and the profile texture is uniform, whole soil is loose and soft, and liable to soil erosion (Wang et al. 2013).
There are 74 dam system projects in the Chenggou River Basin, including 20 key dams, 22 medium-sized dams, and 32 small-sized dams. All check dams were completed by the end of 2008, and no new check dams were built. At present, the built check dams are not completely filled. Check dams are mainly concentrated in the southern part of the watershed, especially in the southeastern direction of the channel is under the control of check dams. By the end of 2012, 54.68 km 2 of terraced fields had been built in the whole basin, accounting for a large proportion of the basin area and evenly distributed. Part of terraced fields had been converted to grassland. There are three rainfall stations in the basin, namely Gaojiacha, Bieduchuan, and Yangshanzui, and a runoff observation station at the outlet (Fig. 5). All water and sediment data are from local hydrographic bureaus.

Data collection and processing
The data used in the research are shown in Table 2. The total number of check dams and terraced fields as well as their approximate distribution locations were obtained from the local hydrology bureau, and the specific terraced field distribution range and the location of check dam were extracted by using remote sensing images (Fig. 6). The quantity extracted is consistent with the hydrographic bureau quantity, and the site identified is confirmed to be correct according to the field investigation. The maximum   (Fig. 7). The rainfall-runoff process was interpolated for one hour and obtained 29 flood processes.

HRU division
The soil in Chenggou River Basin is mainly yellow loam soil and black mature soil, both are loose and porous with poor water storage capacity, so there is no obvious difference in runoff process. Therefore, only land-use type, vegetation cover, and slope are considered in HRU division. The vegetation coverage was divided into low vegetation cover (VFC ≤ 40%), medium vegetation cover (40% < VFC ≤ 70%), and high vegetation cover (VFC > 70%). The slope is divided into low slope (S ≤ 3°), medium slope (3° < S ≤ 5°), and high slope (S > 5°). According to the statistics, the area of grassland and shrub with high vegetation coverage or high slope is less than 0.1km 2 , and the vast majority of forest land is Fig. 6 Examples of the check dams and terraces identified from the remote sensing images in the Chenggou River Basin Fig. 7 Maps of the slope, land use and NDVI. a Slope map is to generate the four classed; b land-use map; c NDVI map located in the low-slope area. Therefore, HRU types with very small area are neglected in the runoff calculation to reduce the workload of parameter calibration. In combination with the discussion on the characteristics of the flow generation mechanism under different underlying surface conditions in 2.3.1, the dominant runoff process of each HRU was determined and the flow generation parameters were determined. A total of 47,796 HRU were identified in the study basin. For blue HRU, the dominant runoff process is saturationexcess runoff. For yellow HRU, the dominant runoff process is infiltration-excess runoff. The green HRU is terraced. Excluding terraced fields, it can be seen that the saturationexcess runoff area is mainly distributed along the river. The dominant runoff process is that the HRU quantity of saturation-excess runoff was 13,731, accounting for 28.73%. The infiltration-excess runoff was 34,065, accounting for 71.27%. From the perspective of land-use types, the HRU number of terraced fields was 1548, and that of check dams was 258. The total HRU of soil and water conservation measures accounted for 3.78% (Fig. 8).

Parameter estimation, model calibration, and validation
This study took 22 rainfall-flood processes from 2013 to 2016 as the calibration period to calibrate the model parameters. Seven rainfalls in 2017 were used as the verification period (Table 3).
In addition, infiltration capacity curve empirical index m is 0.39, Water storage capacity curve empirical index n is 0.11, and value of coefficient was 0.251. In the training of LSTM model, the number of neurons in the hidden layer was set to 50, and the training times corresponding to the number of each neuron were set to 150. The input data are the runoff depth generated by each type of HRU hour by hour.
The spatial distribution of each factor in the sediment yield model is shown in Fig. 9. The parameter calculation method is described in other related papers (Jiang et al. 2005;Shi et al. 2018). K in the basin is between 0 and 0.039, L ranged from 0.036 to 19.678, S ranged from 0 to 3.9976, B is between 0.23 and 1, T is between 0.1 and 1, and E is 0.34. G uses the simple estimation formula, and shallow gully erosion occurs only when the slope is greater than 15 degrees. Experience parameters a and b were calculated using a logarithmic regression curve, and the calculated results were 1.03 and 0.812, respectively. Table 4 shows the runoff and sediment yield of the 29 flood processes simulated by the model. Figure 10 shows the simulation results of rainfall-runoff process in some fields. In the calibration period the RE R , RE P , RE A and NSEQ are 8.86%, 9.44%, 14.45% and 0.87, respectively. The soil loss was not measured on August 3, 2015, due to the occurrence of sediment leakage. In the validation period the RE R , RE P , RE A , and NSE Q are 8.5%, 12.56%, 11.04%, and 0.84, respectively. Figure 11 shows the runoff depth generated by each HRU in the four rainfall processes. In general, rivers and check dams had the highest runoff depth, followed by grassland and shrubs, and the runoff depth was smaller in areas with higher vegetation coverage. The runoff yield of farmland and terrace was relatively unstable in different   . 9 Spatial distribution of soil loss model parameters rainfall processes, but the runoff depth of the terrace was not large. Under the condition of higher vegetation coverage, the water storage capacity of forest land is stronger and the runoff depth is smaller.

Scene deduction
In order to quantitatively analyze the effects of check dam and terraced field on runoff and sediment yield, respectively, the HRU information of the input model was modified to simulate the sediment yield of the watershed without check dam and terraced field, so as to evaluate the amount of water and sediment reduction of check dam and terraced field, respectively.  When the amount of water in the reservoir area is less than the designed flood discharge, the water and sediment in the upstream control range will be completely intercepted by the check dam. The construction time of the check dam in Chenggou River Basin is not long, and the average remaining storage capacity is more than 80%, so all water and sediment processes are simulated by the above situation. The calibrated parameters in scenario 1 were used to simulate the runoff production results of 29 rainfall events from 2013 to 2017 in the underlying surface conditions of scenario 2, and the simulation results of scenario 1 and scenario 2 were compared (Fig. 12). It can be seen from the results that the existing check dam system in the Chenggou River Basin accounts for 54.8% of the control area of the basin and can intercept 55.61%  of the runoff and 47% of the soil loss in the basin on average, with obvious effect of reducing runoff and sediment.
Since the effect of water and sediment reduction in terraced fields is mainly played in the case of large rainfall, this study selected the rainfall process with the maximum rainfall of 3 h greater than 10 mm for simulation under scenario 3. As the calibrated parameters in scenario 1 are used, the water and sediment conditions of 18 rainfall events from 2013 to 2017 are simulated under scenario 3, and the simulation results of scenario 1 and scenario 3 are compared (Fig. 13). It can be seen from the results that terracing can reduce the runoff by 10.54% and the soil loss by 33.8%.

Model applicability
In this study, based on HRU division and LSTM and CSLE models, a runoff-sediment model was established to simulate the water-sediment process in a small basin of the Loess Plateau and analyze the effect of check dam and terrace on water and sediment in watershed. A large number of measured data, such as rainfall, discharge, sediment transport, and parameters in the model, are used for estimation and analysis. The law of reservoir and drainage of check dam and terrace is complex, and the runoff velocity is different under different dominant runoff generations (Scherrer et al. 2003). Both need many experiments to measure and summarize, and it is difficult to extend from small experimental areas to watershed scales. In this paper, the deep learning method of LSTM is considered to calculate the flow intelligently according to the flow generated by each HRU. The slope and gully erosion of Chenggou River Basin were calculated by adding gully factors based on CSLE. The model combines the advantages of physical mechanism and machine learning and can better simulate the water and sediment situation in the basin. However, when this method is extended to large basins, it will encounter the problem of too many flow-producing parameters. In the future, it is necessary to consider how to adjust the runoff-producing parameters quickly. To calculate accurately, the confluence parameters in the deep learning module also adopt a large number of training times and the number of neurons. Reducing these values to improve the calculation speed while ensuring the simulation effect is the key problem whether deep learning can be used in the confluence calculation. By observing the runoff depth of each HRU in each rainfall, the runoff producing situation of different land-use types has its characteristics. The change of underlying surface will have a great influence on runoff yield (Rogger et al. 2017). Combined with the characteristics of rainfall, it is found that when the rainfall intensity is too high, the proportion of saturation-excess runoff in surface runoff will be smaller, and infiltration-excess runoff will be larger. This should be a key factor in reducing runoff in the basin. Many studies (Hu et al. 2020) have proved that the proportion of flood events dominated by infiltration-excess runoff is gradually decreasing in loess Plateau, and climate and land use are the key factors leading to the change of runoff generation pattern. Soil loss is closely related to rainfall intensity, discharge, and underlying surface. On the whole, the soil loss in Chenggou River Basin is mainly slope erosion and gully erosion (Wang 2013). The slope is relatively gentle, and the area with a slope of more than 30 degrees is less than 2 km. In addition, the vegetation recovery of this basin is good, and more than half of the sloping land has been transformed into terraced fields, so gravity erosion is ignored. Soil and water conservation measures differ in their mechanisms for controlling soil erosion (Zhang et al. 2018). Overall, these measures reduce sediment transport at the outlet of the basin by reducing flow and trapping flow, which reduces flow velocity and ultimately removes soil. If the Chenggou River Basin did not have the check dams and terraces currently built, the overall soil erosion in the basin would be about three times greater.
At present, many models are used to simulate the water and sediment condition in loess plateau area, such as the digital Yellow River integrated model constructed by Tsinghua University (Li et al. 2009), or some models with good simulation effect introduced from abroad, such as SWAT model (Xu et al. 2013), WEEP model (Yu et al. 2009), and Tank model (Chen et al. 2014). However, no matter what model is adopted, it is necessary to pay attention to whether the calculation idea of the model conforms to the actual situation of the basin. In calculating runoff yield, it may not be possible to take into account the heterogeneity of the underlying surface of the basin precisely if the whole basin or several sub-basins are generalized into the same parameters. This problem is particularly acute in the basins where human activities have a greater impact. At present, gravity erosion is seldom considered in sediment yield models. In the past, the vegetation coverage of the Loess Plateau was low, and the proportion of gravity erosion was very large, even up to 90% (Xu et al. 2015). Although the current situation is improving, attention should be paid to this problem when introducing other models on the Loess Plateau. The simulation calculation of gravity erosion also needs to be strengthened in the future.

Effects of check dam and terrace on runoff and sediment
The effect of check dam on reducing water and sediment is very obvious. Comparing the simulation results of scenario 1 and scenario 2, it can be seen that the check dam greatly reduces the amount of water and sediment at the outlet of the basin. Check dam can reduce the flow velocity and weaken the flood peak by interrupting the flow (Li et al. 2017). The deposition of silt will further increase the infiltration and evaporation of the reservoir area, replenish groundwater and change the distribution of water in the basin (Huang et al. 2013). Surplus water from the check dam will also be used for human and livestock use in the study area. In the process of silt, the dam will change the channel topography, shorten the slope length and reduce gully erosion (Tian et al. 2019). When the reservoir area is filled with silt, it becomes fertile land for planting, further improving the soil erosion situation in the area (Wang et al. 2021). Due to the shortage of the check dam in the study area, there is no full sediment deposition in the reservoir area, but this may need to be considered when carrying out studies in other basins. Most studies (Liu et al. 2020;Shi et al. 2018) only consider large and medium dams, ignoring the role of small dams in the basin. It is necessary to consider all the check dams to analyze the basin's water and sediment conditions (Shi et al. 2016). Remote sensing recognition makes this research possible, which makes the results of this research more reliable.
However, it should be noted that the main function of check dam is interception, which has little impact on the overall water and sediment of the basin. In this paper and other studies (Liu et al. 2020) on the water and sediment reduction effects of check dam, the rate of water and sediment reduction of check dam is close to the ratio of its control to the basin area. Therefore, the spatial distribution of check dam system also has an effect on runoff and sediment situation. If the rainstorm center is in the check dam control area, especially in the southeast of the study area, the check dam's water and sediment retention rate will be higher. Comparing the surface runoff components of scenario 1 and scenario 2 in simulation calculation, the proportion of saturation-excess runoff is 40.59% and 46.43%, and the proportion of infiltration-excess runoff is 59.41% and 53.57%, respectively. It can be seen that the check dam hardly affects the process of surface runoff in the watershed, and the main reason for the slight difference in runoff composition is the underlying surface condition of uneven distribution. As time goes on, the dam's effectiveness in reducing water and sediment will gradually deteriorate (Zhang et al. 2018). Therefore, check dams can only be used as a temporary solution to trap sediment and reduce flooding. Combining other soil and water conservation measures is necessary to cure serious soil and water loss (Guo et al. 2019).
Terracing is a widely used measure in China, which can increase agricultural land and reduce soil erosion (Ran et al. 2020). As terraces change the topography of the basin and reduce the slope, the hydrological connectivity on the ridge platform is greatly reduced, thus slowing down the surface runoff velocity, prolonging the infiltration time, delaying the flood peak, and achieving the effect of reducing water and sediment. Some experiments (Nunes et al. 2018) have proved that the dominant runoff process in terraced fields is saturation-excess runoff. Comparing the surface runoff components of scenario 1 and scenario 3, saturation-excess runoff accounted for 40.59% and 9.54%, and infiltration-excess runoff accounted for 59.41% and 90.46%, respectively. It can be found that the construction of a large number of terraced fields in Chenggou River has a great impact on the hydrological process of the basin. From the model simulation results, the effect of sediment reduction is more obvious than that of water reduction. For example, Li (2015) used SWAT to simulate the water-sediment effect of terraced fields in Yanhe River Basin and found that terraced fields reduced 26.31% of water and 65.52% of soil loss in the basin. The reason may be that horizontal terraced fields could reduce their water and sediment and trap the upper sand-bearing flow (Liu et al. 2014).
The effect of reducing water and sediment in terraced fields is related to many factors. The structure of terraced fields, vegetation coverage, and vegetation types will impact water and sediment reduction in terraced fields (Duan et al. 2021). However, due to the lack of long-term rainfall and runoff observation data in the study area, we could not distinguish the effects of terraced fields and vegetation in the Chenggou River Basin. By comparing the rainfall characteristics of each flood, it can be found that the larger the maximum rainfall or cumulative rainfall in 10 days intensity is better the effect of reducing water and sediment. For example, in the two floods of 20,140,618 and 20,170,806, the maximum rainfall in one hour exceeded 20 mm, and the water reduction rate reached 30%, and the sediment reduction rate reached 48%. Other experiments ) have found similar results. However, it should be noted that under extreme rainfall conditions, gully erosion and gravity erosion of terraced fields may be greater, and even collapse may occur in severe cases (Peng et al. 2020). In this case, the actual water and sediment retention efficiency may be reduced. Terrace construction needs to be fully combined with vegetation, and we will pay attention to distinguish extreme rainfall in future runoff and sediment calculations.

Conclusion
In this study, remote sensing technology was used to accurately extract the distribution of soil and water conservation measures in watershed. Based on the runoff generation mechanism, the dominant runoff process of the underlying surface of the basin was differentiated and HRU was divided. And according to the HRU dominant runoff process to choose the appropriate runoff calculation method, the runoff generation and confluence model of LSTM were constructed, and the CSLE soil erosion model based on subrainfall was integrated. HRU division made the underlying surface runoff mechanism of the basin more objective and reasonable and made the runoff calculation more refined. A total of 47,796 HRU were divided in the study area in which the saturation-excess runoff and infiltration-excess runoff accounted for 28.73% and 71.27%, respectively. The traditional physical mechanism and deep learning were combined to achieve reasonable and accurate simulation of runoff and sediment in the watershed. In the calibration period, the flood mean relative error, peak flood error, Nash coefficient, and sediment quantity error are 8.86%, 9.44%, 0.87, and 14.45%, respectively. In the validation period, the flood mean relative error, peak flood error, Nash coefficient, and sediment quantity error are 8.5%, 12.56%, 0.84, and 11.04%, respectively. The simulation effect of the model is good, and it has a reasonable physical mechanism, which can be extended to other basins through setting different scenarios, quantitative analysis of the impact of check dams and terraces on the watershed water, and sediment. The results showed that the check dams could reduce the runoff by 55.61% and the soil loss by 47%, and the terrace could reduce the runoff by 10.54% and the soil loss by 33.8%.
This study is conducted based on investigation and analysis, which has been well-verified in the Chenggou River Basin. It is hoped that the experiment can be carried out in several basins in the future. Moreover, in this study, the model of combining traditional physical mechanism with deep learning is only a preliminary attempt. It is necessary to conduct in-depth discussion, research on the combination method, and the specific parameters of LSTM for confluence calculation was also not considered. In the future, we hope to be able to do dynamic parameters and the dynamic changes of runoff and soil erosion, strengthening the combination of hydrological calculation and deep learning, in order to better consider climate change and different underlying surface changes on the influence of the Yellow River basin water resources, for the Yellow River basin ecological protection strategy and further development of high quality to provide theoretical support.