*2.6. Model Performance Metrics*

The applicability of SWAT was evaluated in terms of Nash–Sutcliffe Efficiency (NSE), percent bias (PBIAS), and squared correlation coefficient (R<sup>2</sup> ). The NSE can range between −∞ and 1. If the value of the NSE was closer to 1, it was indicative of a better simulation performance and high reliability of the SWAT model. When the NSE was closer to 0.5, the model simulation results were similar to the mean of observations, that is, the model results in general were reliable. A PBIAS between −10% and 10% indicated a good simulation performance of the model. Additionally, larger values of R<sup>2</sup> were indicative of a better simulation performance of the model. The calculation process and significance of the three metrics have been elaborated elsewhere [34]. Model performance during the simulations of the groundwater table depth was evaluated mainly in terms of the absolute error and R<sup>2</sup> in this study.

$$\text{NSE} = 1 - \left[ \frac{\sum\_{\text{i=1}}^{n} \left( \mathbf{V\_{i}^{\text{obs}}} - \mathbf{V\_{i}^{\text{sim}}} \right)^{2}}{\sum\_{\text{i=1}}^{n} \left( \mathbf{V\_{i}^{\text{obs}}} - \mathbf{V^{mean}} \right)^{2}} \right] \tag{1}$$

$$\text{PBIAS} = \left[ \frac{\sum\_{i=1}^{n} \left( \mathbf{V\_i^{obs}} - \mathbf{V\_i^{sim}} \right) \times 100}{\sum\_{i=1}^{n} \mathbf{V\_i^{obs}}} \right] \tag{2}$$

$$\mathbf{R}^2 = \frac{\left[\sum\_{\mathbf{i}=1}^{\mathrm{N}} \left(\mathbf{V\_{\bar{i}}^{\mathrm{sim}}} - \overline{\nabla}^{\mathrm{sim}}\right) \left(\mathbf{V\_{\bar{i}}^{\mathrm{obs}}} - \overline{\nabla\_{\mathrm{i}}}^{\mathrm{obs}}\right)\right]^2}{\sum\_{\mathbf{i}=1}^{\mathrm{N}} \left(\mathbf{V\_{\bar{i}}^{\mathrm{sim}}} - \overline{\nabla}^{\mathrm{sim}}\right)^2 \sum\_{\mathbf{i}=1}^{\mathrm{N}} \left(\mathbf{V\_{\bar{i}}^{\mathrm{obs}}} - \overline{\nabla}^{\mathrm{obs}}\right)^2} \tag{3}$$

#### **3. Results**

## *3.1. Regional Vegetation Change*

The main types of vegetation change in the study region from 2002 to 2018 are summarised in Table 3. Specifically, the main types of vegetation change pertained to the conversion of low-coverage grassland, farmland, and bare land to forestland, in which the converted areas amounted to 2,877,518 and 321 hm<sup>2</sup> , respectively. Correspondingly, the irrigation water volume of each revegetation plot underwent dramatic changes. In addition, the conversion of farmland to forestland mainly occurred in the irrigated area of the Gahai Lake in the south-eastern part of the basin, while the conversion of low-coverage grassland and bare land to forestland mainly occurred in the irrigated area of Delingha, which is situated in the north-western part of the basin (Figure 6).

**Table 3.** Main types of revegetation in the study region.


#### *3.2. Comparison of the Original SWAT-MODFLOW and LU-SWAT-MODFLOW*

#### 3.2.1. Difference in HRUs

Figure 9 shows the HRUs generated by the original SWAT-MODFLOW model versus those from the LU-SWAT-MODFLOW model. The original SWAT-MODFLOW model generated 1304 HRUs, and the LU-SWAT-MODFLOW model generated 2978 HRUs. The higher number of HRUs in the new model was due to the land use/cover data of LU-SWAT-MODFLOW being a superposition of years-long data and thereby covering a higher number of patches. The higher number of HRUs also implied that the operation and parameter tuning of the LU-SWAT-MODFLOW model would be more complicated [35].

**HRU**

irrigation water volume of each revegetation plot underwent dramatic changes. In addi‐ tion, the conversion of farmland to forestland mainly occurred in the irrigated area of the Gahai Lake in the south‐eastern part of the basin, while the conversion of low‐coverage grassland and bare land to forestland mainly occurred in the irrigated area of Delingha,

> **Revegetation Area (hm2)**

estland <sup>518</sup> <sup>5800</sup>→<sup>5400</sup>

estland <sup>321</sup> <sup>0</sup>→<sup>5400</sup>

Figure 9 shows the HRUs generated by the original SWAT‐MODFLOW model versus those from the LU‐SWAT‐MODFLOW model. The original SWAT‐MODFLOW model generated 1304 HRUs, and the LU‐SWAT‐MODFLOW model generated 2978 HRUs. The higher number of HRUs in the new model was due to the land use/cover data of LU‐ SWAT‐MODFLOW being a superposition of years‐long data and thereby covering a higher number of patches. The higher number of HRUs also implied that the operation and parameter tuning of the LU‐SWAT‐MODFLOW model would be more complicated

*3.2. Comparison of the Original SWAT‐MODFLOW and LU‐SWAT‐MODFLOW*

**Change in the Annual Irrigation Rate (m3/hm2∙a)**

2877 0→5400

which is situated in the north‐western part of the basin (Figure 6).

**Table 3.** Main types of revegetation in the study region.

**Main Type of Revegetation**

Conversion of low‐coverage grass‐ land to forestland

Conversion of farmland to for‐

Conversion of bare land to for‐

3.2.1. Difference in HRUs

[35].

**Figure 9.** HRUs generated by (**a**) SWAT‐MODFLOW and (**b**) LU‐SWAT‐MODFLOW. **Figure 9.** HRUs generated by (**a**) SWAT-MODFLOW and (**b**) LU-SWAT-MODFLOW.

#### 3.2.2. Comparison of the LAI Simulation Results 3.2.2. Comparison of the LAI Simulation Results

Figure 10 shows the simulated LAI from the calibrated original SWAT‐MOD‐ FLOW model and the calibrated LU‐SWAT‐MODFLOW model for July 2005, 2010, 2015, and 2018. Compared to the remote‐sensed LAI, the calibrated LU‐SWAT‐MOD‐ FLOW model has better performance in expressing the spatial variation of the LAI than SWAT‐MODFLOW since the LU‐SWAT‐MODFLOW model has more HRUs. In addition, we randomly chose 10 positions in different parts of the study area and cal‐ culated the performance metrics of LAI in corresponding HRUs of original SWAT‐ Figure 10 shows the simulated LAI from the calibrated original SWAT-MODFLOW model and the calibrated LU-SWAT-MODFLOW model for July 2005, 2010, 2015, and 2018. Compared to the remote-sensed LAI, the calibrated LU-SWAT-MODFLOW model has better performance in expressing the spatial variation of the LAI than SWAT-MODFLOW since the LU-SWAT-MODFLOW model has more HRUs. In addition, we randomly chose 10 positions in different parts of the study area and calculated the performance metrics of LAI in corresponding HRUs of original SWAT-MODFLOW and LU-SWAT-MODFLOW model, respectively. The performance metrics for the calibrated and validated original SWAT-MODFLOW model were NSE > 0.75, PBIAS of <sup>−</sup>25–25%, and R<sup>2</sup> > 0.73, and the counterparts for the calibrated and validated LU-SWAT-MODFLOW were NSE > 0.83, PBIAS of <sup>−</sup>20–20%, and R<sup>2</sup> > 0.83 (Table 4). This indicates that the LU-SWAT-MODFLOW model was more accurate than the original SWAT-MODFLOW model in simulations of the monthly LAI after both models were calibrated and validated. *Water* **2021**, *13*, x FOR PEER REVIEW 11 of 20 MODFLOW and LU‐SWAT‐MODFLOW model, respectively. The performance met‐ rics for the calibrated and validated original SWAT‐MODFLOW model were NSE > 0.75, PBIAS of −25%–25%, and R2 > 0.73, and the counterparts for the calibrated and validated LU‐SWAT‐MODFLOW were NSE > 0.83, PBIAS of −20–20%, and R2 > 0.83 (Table 4). This indicates that the LU‐SWAT‐MODFLOW model was more accurate than the origi‐ nal SWAT‐MODFLOW model in simulations of the monthly LAI after both models were calibrated and validated.

**Figure 10.** Simulated leaf area index (LAI) from the (**a**–**d**) calibrated original SWAT‐MODFLOW model, (**e**–**h**) calibrated LU‐SWAT‐MODFLOW model, and (**i**–**l**) remote‐sensed LAI. **Figure 10.** Simulated leaf area index (LAI) from the (**a**–**d**) calibrated original SWAT-MODFLOW model, (**e**–**h**) calibrated LU-SWAT-MODFLOW model, and (**i**–**l**) remote-sensed LAI.

E1 0.85 20.84 0.83 0.81 17.63 0.79 0.92 14.11 0.88 0.87 16.54 0.85 E2 0.86 23.71 0.85 0.80 21.25 0.76 0.91 11.13 0.87 0.86 14.52 0.83 W1 0.82 19.89 0.82 0.78 15.65 0.75 0.88 13.45 0.90 0.84 17.89 0.88 W2 0.83 17.41 0.81 0.75 19.32 0.73 0.89 12.65 0.89 0.83 19.21 0.84 S1 0.85 13.72 0.84 0.81 17.29 0.77 0.90 10.99 0.90 0.86 16.54 0.86 S2 0.83 19.24 0.81 0.79 14.23 0.78 0.91 11.11 0.90 0.85 17.32 0.87 N1 0.87 15.36 0.85 0.77 13.22 0.78 0.93 14.65 0.92 0.88 15.21 0.89 N2 0.86 17.52 0.84 0.76 11.21 0.77 0.90 17.53 0.89 0.84 18.56 0.90 C1 0.84 21.49 0.83 0.78 22.17 0.76 0.92 16.46 0.90 0.87 14.12 0.87 C2 0.85 20.76 0.83 0.81 19.78 0.74 0.91 13.02 0.92 0.83 11.76 0.89

3.2.3. Comparison of the ET Simulation Results

**Calibration Period Validation Period Calibration Period Validation Period NSE PBIAS R2 NSE PBIAS R2 NSE PBIAS R2 NSE PBIAS R2**


**Table 4.** Performance of the original SWAT-MODFLOW model versus the LU-SWAT-MODFLOW model in simulating LAI.

#### 3.2.3. Comparison of the ET Simulation Results

During the calibration and validation period, the performance metrics of the original SWAT-MODFLOW model were NSE > 0.65, PBIAS of <sup>−</sup>20%–20%, and R<sup>2</sup> > 0.63 in simulations of the monthly mean ET for each sub-basin, while the counterparts for the LU-SWAT-MODFLOW model were NSE > 0.72, PBIAS of <sup>−</sup>20%–20%, and R<sup>2</sup> > 0.73 (Figure 11). This indicates that the LU-SWAT-MODFLOW model was more accurate than the original SWAT-MODFLOW in simulations of the monthly mean ET for most of the sub-basins. Figure 12 shows the multi-year mean of the simulated ET from the original SWAT-MODFLOW model (Figure 12a) versus that of the LU-SWAT-MODFLOW model (Figure 12b) in comparison with the multi-year mean of remote sensing-derived ET (Figure 12c). The multi-year mean of remote sensing-derived ET exhibited a spatial distribution pattern of high values in the north-eastern mountains and low values in the southwestern plains, and such a distribution pattern existed for both calibrated models.

#### 3.2.4. Comparison of the Simulation Results for Groundwater Table Depth

Groundwater table depth data were scarce within the study region. Observation wells 1, 2, and 3 only provided monthly data for 2009–2011, and observation well 4 only provided monthly data for 2013–2015; meanwhile, observation well 5 only provided monthly data for 2014–2015. Thus, the observed groundwater table depth of wells 1, 2, and 3 were used to calibrate the two models, and the rest were used to validate the models. Linear regression results of the simulated groundwater table depth on observed groundwater table depth were compared between the original SWAT-MODFLOW model and the LU-SWAT-MODFLOW model (Figure 13). Both models performed well in simulating the changes in the groundwater table depth of the study region, with an R<sup>2</sup> > 0.95 and absolute error within 0.5 m. In addition, the simulation performance of LU-SWAT-MODFLOW was slightly better than that of SWAT-MODFLOW, which was likely attributed to the detailed consideration of the spatiotemporal changes in irrigation and land cover by the former model versus the latter model.

and such a distribution pattern existed for both calibrated models.

During the calibration and validation period, the performance metrics of the original SWAT‐MODFLOW model were NSE > 0.65, PBIAS of −20%–20%, and R2 > 0.63 in simula‐ tions of the monthly mean ET for each sub‐basin, while the counterparts for the LU‐ SWAT‐MODFLOW model were NSE > 0.72, PBIAS of −20%–20%, and R2 > 0.73 (Figure 11). This indicates that the LU‐SWAT‐MODFLOW model was more accurate than the original SWAT‐MODFLOW in simulations of the monthly mean ET for most of the sub‐ basins. Figure 12 shows the multi‐year mean of the simulated ET from the original SWAT‐ MODFLOW model (Figure 12a) versus that of the LU‐SWAT‐MODFLOW model (Figure 12b) in comparison with the multi‐year mean of remote sensing‐derived ET (Figure 12c). The multi‐year mean of remote sensing‐derived ET exhibited a spatial distribution pattern of high values in the north‐eastern mountains and low values in the southwestern plains,

**Figure 11.** NSE (**a**), R2 (**b**), PBIAS (**c**) of the calibrated original SWAT‐MODFLOW model versus the calibrated LU‐SWAT‐ MODFLOW model in simulating evapotranspiration in each sub‐basin. **Figure 11.** NSE (**a**), R<sup>2</sup> (**b**), PBIAS (**c**) of the calibrated original SWAT-MODFLOW model versus the calibrated LU-SWAT-MODFLOW model in simulating evapotranspiration in each sub-basin. *Water* **2021**, *13*, x FOR PEER REVIEW 13 of 20

**Figure 12.** Multi‐year mean ET simulated by (**a**) SWAT‐MODFLOW, (**b**) LU‐SWAT‐MODFLOW, and (**c**) retrieved from remote sensing images. **Figure 12.** Multi-year mean ET simulated by (**a**) SWAT-MODFLOW, (**b**) LU-SWAT-MODFLOW, and (**c**) retrieved from remote sensing images.

Groundwater table depth data were scarce within the study region. Observation wells 1, 2, and 3 only provided monthly data for 2009–2011, and observation well 4 only

monthly data for 2014–2015. Thus, the observed groundwater table depth of wells 1, 2, and 3 were used to calibrate the two models, and the rest were used to validate the models. Linear regression results of the simulated groundwater table depth on observed ground‐ water table depth were compared between the original SWAT‐MODFLOW model and the LU‐SWAT‐MODFLOW model (Figure 13). Both models performed well in simulating the changes in the groundwater table depth of the study region, with an R2 > 0.95 and absolute error within 0.5 m. In addition, the simulation performance of LU‐SWAT‐MODFLOW was slightly better than that of SWAT‐MODFLOW, which was likely attributed to the de‐ tailed consideration of the spatiotemporal changes in irrigation and land cover by the for‐

3.2.4. Comparison of the Simulation Results for Groundwater Table Depth

mer model versus the latter model.

*Water* **2021**, *13*, x FOR PEER REVIEW 14 of 20

**Figure 13.** Monthly groundwater table depth simulated by the (**a**–**e**) SWAT‐MODFLOW model ver‐ sus the (**f**–**j**) LU‐SWAT‐MODFLOW model. **Figure 13.** Monthly groundwater table depth simulated by the (**a**–**e**) SWAT-MODFLOW model versus the (**f**–**j**) LU-SWAT-MODFLOW model.

#### *3.3. Impacts of Vegetation Change on Hydrological Processes 3.3. Impacts of Vegetation Change on Hydrological Processes*

The case study area is located in a water consumption area of an inland river and almost never generates runoff. Thus, here we focus on the analysis of the vegetation change impacts on ET and groundwater processes. The case study area is located in a water consumption area of an inland river and almost never generates runoff. Thus, here we focus on the analysis of the vegetation change impacts on ET and groundwater processes.

#### 3.3.1. Impacts on ET 3.3.1. Impacts on ET

The LU‐SWAT‐MODFLOW model was run in the following two scenarios to accu‐ rately analyse the impacts of revegetation and the related extensive irrigation on ET: (1) revegetation was assumed absent while considering the actual changes in other types of land use/cover; and (2) the actual changes in land use/cover were considered, including those pertinent to revegetation (irrigation) and other types of land use/cover. Figure 14a The LU-SWAT-MODFLOW model was run in the following two scenarios to accurately analyse the impacts of revegetation and the related extensive irrigation on ET: (1) revegetation was assumed absent while considering the actual changes in other types of land use/cover; and (2) the actual changes in land use/cover were considered, including those pertinent to revegetation (irrigation) and other types of land use/cover. Figure 14a shows the simulated monthly ET in the revegetation-absent scenario versus the revegetation-present scenario from 2002 to 2018. The results indicate that revegetation and related irrigation did not change the trend of monthly ET in the basin, in which the monthly ET in the revegetation-present scenario was only 1.5 mm higher than that in the revegetation-absent scenario for most months. Similarly, the trend of annual ET was almost the same in both scenarios. In 2004 and later years, ET showed weakly higher values in the revegetation-present scenario than in the revegetation-absent scenario (Figure 14b). Figure 13c illustrates the difference in the multi-year mean ET between the two scenarios in each sub-basin. Such a difference was greater than 10 mm in sub-basins 4, 12, 13, 14, 26, and 33, that is, the ET increase was most obvious in these sub-basins. Comprehensive comparisons of the land use/cover map (Figure 6) with the LAI map for the study region during the study period further confirmed that relatively obvious revegetation had been achieved in these sub-basins. irrigation did not change the trend of monthly ET in the basin, in which the monthly ET in the revegetation‐present scenario was only 1.5 mm higher than that in the revegetation‐ absent scenario for most months. Similarly, the trend of annual ET was almost the same in both scenarios. In 2004 and later years, ET showed weakly higher values in the revege‐ tation‐present scenario than in the revegetation‐absent scenario (Figure 14b). Figure 13c illustrates the difference in the multi‐year mean ET between the two scenarios in each sub‐ basin. Such a difference was greater than 10 mm in sub‐basins 4, 12, 13, 14, 26, and 33, that is, the ET increase was most obvious in these sub‐basins. Comprehensive comparisons of the land use/cover map (Figure 6) with the LAI map for the study region during the study period further confirmed that relatively obvious revegetation had been achieved in these sub‐basins.

shows the simulated monthly ET in the revegetation‐absent scenario versus the revegeta‐ tion‐present scenario from 2002 to 2018. The results indicate that revegetation and related

*Water* **2021**, *13*, x FOR PEER REVIEW 15 of 20

**Figure 14.** (**a**) Monthly and (**b**) yearly ET with revegetation and without revegetation; (**c**) yearly average ET change in different sub‐basins after revegetation. **Figure 14.** (**a**) Monthly and (**b**) yearly ET with revegetation and without revegetation; (**c**) yearly average ET change in different sub-basins after revegetation.

#### 3.3.2. Impacts on Groundwater Recharge 3.3.2. Impacts on Groundwater Recharge

charge.

Groundwater is the most important water resource in the arid endorheic river water‐ shed. Changes in groundwater recharge may affect the groundwater storage and further impact the ecological environment. Figure 15 shows the monthly (Figure 15a) and yearly (Figure 15b) groundwater recharge in the entire study area. After revegetation, the groundwater recharge increased by approximately 1.27 mm on average per month and 14.02 mm on average per year. Fan et al. [36], Yang and Lu [37], and Qubaja et al. [30] showed that canopy interception and root water absorption would lead to reduction of soil water, surface runoff, and groundwater recharge in woodland. However, here, alt‐ hough considerable areas of low‐coverage grassland, farmland, and bare land were con‐ verted to forestland, the groundwater recharge with revegetation was evidently higher than that without revegetation. We reported the yearly average groundwater recharge after revegetation in the entire study area (Figure 14c). The groundwater recharge in the irrigation district where the revegetation was applied was the highest (>14.51 m3/day); that is, the irrigation for the recovered vegetation strongly affected the groundwater re‐ Groundwater is the most important water resource in the arid endorheic river watershed. Changes in groundwater recharge may affect the groundwater storage and further impact the ecological environment. Figure 15 shows the monthly (Figure 15a) and yearly (Figure 15b) groundwater recharge in the entire study area. After revegetation, the groundwater recharge increased by approximately 1.27 mm on average per month and 14.02 mm on average per year. Fan et al. [36], Yang and Lu [37], and Qubaja et al. [30] showed that canopy interception and root water absorption would lead to reduction of soil water, surface runoff, and groundwater recharge in woodland. However, here, although considerable areas of low-coverage grassland, farmland, and bare land were converted to forestland, the groundwater recharge with revegetation was evidently higher than that without revegetation. We reported the yearly average groundwater recharge after revegetation in the entire study area (Figure 14c). The groundwater recharge in the irrigation district where the revegetation was applied was the highest (>14.51 m3/day); that is, the irrigation for the recovered vegetation strongly affected the groundwater recharge.

*Water* **2021**, *13*, x FOR PEER REVIEW 16 of 20

**Figure 15.** (**a**) Monthly and (**b**) yearly groundwater recharge with and without revegetation; (**c**) yearly average ground‐ water recharge after revegetation in space. **Figure 15.** (**a**) Monthly and (**b**) yearly groundwater recharge with and without revegetation; (**c**) yearly average groundwater recharge after revegetation in space.

#### 3.3.3. Impacts on Surface Water and Groundwater Exchange There was frequent surface‐water–groundwater exchange in the study region, which 3.3.3. Impacts on Surface Water and Groundwater Exchange

dominated the regional hydrological processes. We analysed the surface water and groundwater exchange affected by revegetation. Figure 16 shows the amount of ground‐ water recharge and discharge in the revegetation‐absent scenario minus the value in the revegetation‐present scenario. Specifically, river reach I was in the upper study region, where groundwater was recharged by river water. River reach II was situated in the lower study region, where significant amounts of groundwater were discharged to the river. In river reach III, both surface water recharge to groundwater and groundwater discharge to surface water were present. River reach III was situated in the irrigated area of De‐ lingha, where it was greatly affected by agricultural, forestland, and grassland irrigation, which led to a relatively complex pattern of surface‐water–groundwater exchange. The area where river reach III was situated was also the main revegetation area of the study region. Comparisons of Figure 16 revealed that the direction of surface‐water–groundwa‐ ter exchange was reversed in six grid cells, which was attributed to the changes in the irrigation volume within these grid cells after revegetation. There was frequent surface-water–groundwater exchange in the study region, which dominated the regional hydrological processes. We analysed the surface water and groundwater exchange affected by revegetation. Figure 16 shows the amount of groundwater recharge and discharge in the revegetation-absent scenario minus the value in the revegetation-present scenario. Specifically, river reach I was in the upper study region, where groundwater was recharged by river water. River reach II was situated in the lower study region, where significant amounts of groundwater were discharged to the river. In river reach III, both surface water recharge to groundwater and groundwater discharge to surface water were present. River reach III was situated in the irrigated area of Delingha, where it was greatly affected by agricultural, forestland, and grassland irrigation, which led to a relatively complex pattern of surface-water–groundwater exchange. The area where river reach III was situated was also the main revegetation area of the study region. Comparisons of Figure 16 revealed that the direction of surface-water–groundwater exchange was reversed in six grid cells, which was attributed to the changes in the irrigation volume within these grid cells after revegetation. *Water* **2021**, *13*, x FOR PEER REVIEW 17 of 20

**Figure 16.** Revegetation impacts on groundwater recharge and discharge. dominated by the conversion of low‐coverage grassland to bare land and forestland, and **Figure 16.** Revegetation impacts on groundwater recharge and discharge.

validated. LAI plays a key role in SWAT for estimating ET, canopy interception, and biomass accumulation [35]. The enhanced modelling of LAI could improve the per‐ formance of the SWAT model in eco‐hydrological processes [26,38]. However, accu‐ rate simulation of LAI relies on many parameters which are difficult to calibrate. Gen‐ erally, parameters of SWAT‐MODFLOW are calibrated with observed data in water‐ shed outlets or sub‐basins [23,27]. Only few studies have calibrated the parameters at the HRU level because of its difficulty and complexity. In this study, the remote‐ sensed monthly LAI data were used to calibrate the SWAT‐MODFLOW and LU‐ SWAT‐MODFLOW at the HRU level using the SWAT‐CUP software (https://swat.tamu.edu/software/swat‐cup/) (accessed on 01 October 2021)with a sat‐ isfactory result. This suggests that the model calibration at the HRU level is possible

The LU‐SWAT‐MODFLOW model was more accurate than the original SWAT‐ MODFLOW in simulating the monthly mean ET for most sub‐basins. Ma et al. [26] re‐ ported that canopy interception and soil water content would be seriously affected by LAI in SWAT, which would further affect ET. Therefore, the enhancements of LU‐SWAT‐ MODFLOW in modelling the monthly ET can be attributed to the more accurate simula‐

Revegetation projects have been conducted in both the Gahai Lake irrigated area and the irrigated area of Delingha, but the revegetation had a relatively high impact on the direction and amount of surface‐water–groundwater exchange in the latter area; in the former area, there was an almost negligible impact. This discrepancy was attributed to the fact that revegetation in the Gahai Lake irrigated area was mainly characterised by the conversion of farmland to forestland, and the irrigation volume did not differ significantly between the two land cover types [38]. In contrast, the irrigated area of Delingha was

and effective if the related observation data exist.

**4. Discussion**

tion of the LAI.

#### **4. Discussion**

The LU-SWAT-MODFLOW model was more accurate than the original SWAT-MODFLOW model in simulating the monthly LAI after both models were calibrated and validated. LAI plays a key role in SWAT for estimating ET, canopy interception, and biomass accumulation [35]. The enhanced modelling of LAI could improve the performance of the SWAT model in eco-hydrological processes [26,38]. However, accurate simulation of LAI relies on many parameters which are difficult to calibrate. Generally, parameters of SWAT-MODFLOW are calibrated with observed data in watershed outlets or sub-basins [23,27]. Only few studies have calibrated the parameters at the HRU level because of its difficulty and complexity. In this study, the remote-sensed monthly LAI data were used to calibrate the SWAT-MODFLOW and LU-SWAT-MODFLOW at the HRU level using the SWAT-CUP software (https://swat.tamu.edu/software/swat-cup/) (accessed on 1 October 2021) with a satisfactory result. This suggests that the model calibration at the HRU level is possible and effective if the related observation data exist.

The LU-SWAT-MODFLOW model was more accurate than the original SWAT-MODFLOW in simulating the monthly mean ET for most sub-basins. Ma et al. [26] reported that canopy interception and soil water content would be seriously affected by LAI in SWAT, which would further affect ET. Therefore, the enhancements of LU-SWAT-MODFLOW in modelling the monthly ET can be attributed to the more accurate simulation of the LAI.

Revegetation projects have been conducted in both the Gahai Lake irrigated area and the irrigated area of Delingha, but the revegetation had a relatively high impact on the direction and amount of surface-water–groundwater exchange in the latter area; in the former area, there was an almost negligible impact. This discrepancy was attributed to the fact that revegetation in the Gahai Lake irrigated area was mainly characterised by the conversion of farmland to forestland, and the irrigation volume did not differ significantly between the two land cover types [38]. In contrast, the irrigated area of Delingha was dominated by the conversion of low-coverage grassland to bare land and forestland, and the former two land cover types required no irrigation, while the latter land cover type required a large irrigation volume.

This study is subjected to some limitations. On the one hand, we used land use/cover map to analyse the revegetation process in our study area. In fact, plant density, age, and growth status were not considered because of the limitations in the SWAT model. Moreover, these factors may affect the eco-hydrological processes in such an arid area [26]. On the other hand, meteorological data were scarce in both original SWAT-MODFLOW and LU-SWAT-MODFLOW models. This may impact the model performance in formulating the water budget [39–41]. Nonetheless, these limitations should be addressed in future studies by using and analysing different datasets.

#### **5. Conclusions**

This study was carried out in the middle and lower reaches of the Bayin River basin in the north-eastern part of the Qaidam Basin, China, where there is frequent surface-water– groundwater interaction and evident vegetation change. A LU-SWAT-MODFLOW model was developed by integrating a coupled SWAT-MODFLOW model with dynamic HRUs in view of their ability to reflect the actual land cover changes in the basin. The impacts of revegetation and related irrigation on the main hydrological processes in the basin were more accurately simulated and analysed by the LU-SWAT-MODFLOW model than by the original SWAT-MODFLOW model.

The LU-SWAT-MODFLOW model generated dynamic HRUs by pre-defining spatial units where land use/cover changes occurred during the simulated period, thereby overcoming the inability of the original SWAT model to effectively reflect the complete or partial land cover type conversion within the same HRU. This new model outperformed the original SWAT-MODFLOW model in simulating the LAI. The LAI is an important parameter of SWAT as it affects a series of processes, such as ET and infiltration; therefore, accurate simulations of the LAI are a key to accurate hydrological simulations. Moreover, the LU-SWAT-MODFLOW model outperformed the original SWAT-MODFLOW model in simulating the ET and groundwater table depth of the basin.

The LU-SWAT-MODFLOW model was run in two different scenarios, one with revegetation and the other without it, to assess the impacts of revegetation and related irrigation on the main hydrological processes in the study region. The results showed that after regional revegetation, ET in the different sub-basins increased by approximately 1.5 mm per month and by 6 mm per year. After revegetation, the groundwater recharge increased by approximately 1.27 mm on average per month and 14.02 mm on average per year. Irrigation for the recovered vegetation strongly affected the groundwater recharge. Meanwhile, the direction and amount of surface-water–groundwater exchange underwent evident changes in areas where revegetation was characterised by the conversion of low-coverage grassland and bare land to forestland. In areas where revegetation was characterised by the conversion of farmland to forestland, the irrigation volume was not greatly altered; thus, this transition had a weak impact on the direction and amount of surface-water–groundwater exchange. Changes in the direction and amount of surface-water–groundwater exchange may lead to a series of ecological and environmental issues. To avoid problems in the future, water-saving irrigation techniques should be advocated when conducting revegetation in arid inland river basins. In addition, our findings indicate that it would be advantageous to preferentially apply revegetation measures that promote the conversion of farmland to forestland/grassland provided that they do not adversely affect regional economic development.

**Author Contributions:** Conceptualization, X.J. and X.M.; methodology, X.J.; formal analysis, X.J.; investigation, X.J.; resources, Y.J.; data curation J.Z. and D.F.; writing—original draft preparation, X.J.; writing—review and editing, Y.J.; project administration, X.J.; funding acquisition, X.J. and X.M. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the National Natural Science Foundation of China, grant number 41801094 and grants from the Natural Science Foundation of Qinghai Province, China, grant number 2021-ZJ-705.

**Data Availability Statement:** The processed data required to reproduce these findings cannot be shared at this time as the data is also a part of an ongoing study.

**Conflicts of Interest:** The authors declare no conflict of interest.
