Integrating GLASS LAI into the SWAT Model for Improved Hydrological Simulation in Semi-Arid Regions

Abstract

The Soil and Water Assessment Tool (SWAT) model has been widely used to simulate ecohydrological processes in watersheds. However, the SWAT model uses a simplified Environmental Policy Impact Climate (EPIC) model to simulate the leaf area index (LAI), creating a critical gap in accurately simulating evapotranspiration (ET) and runoff in semi-arid regions. This work aims to fill this gap by modifying the SWAT source code to integrate high-resolution Global Land Surface Satellite (GLASS) leaf area index (LAI) data. The modified version was applied to the semi-arid Wuding River Basin and calibrated using a Fortran-based dynamic dimension search (DDS) algorithm. The results show a relatively significant improvement in the accuracy of the daily-scale runoff simulation (R² from 0.52 to 0.71 and NSE from 0.52 to 0.7 for the calibration period, and R² from 0.21 to 0.58 and NSE from 0.2 to 0.51 for the validation period). The improved version also corrects the unrealistic default LAI peak (from >5.0 to 1.5–3.0), correcting the multi-year average ET from 251.7 mm to 341.8 mm. The improved vegetation growth module of the SWAT model effectively improved the accuracy of hydrologic simulation in the semi-arid region and enhanced the structural robustness of SWAT for water management.

1. Introduction

Water scarcity and ecological fragility are the two core bottlenecks constraining the sustainable socio-economic development of semi-arid regions around the world [1,2]. In these water-scarce environments, the interactions between vegetation dynamics and the water cycle are extremely sensitive and complex, showing strong nonlinear characteristics. As a key link between climate change and water cycle processes, vegetation directly influences the partitioning of precipitation between evapotranspiration and runoff [3]. However, precise quantification of these processes remains difficult due to the dramatic spatial and temporal variability in precipitation and subsurface conditions. Scholars around the world have conducted numerous studies in different regions, but semi-arid watersheds with complex geomorphology and climate still provide an ideal window to explore the response of the water cycle to environmental change [4,5,6,7].

Large-scale ecological restoration programs have been pursued globally over the past few decades in response to land degradation and desertification. These measures have significantly improved underlying surface conditions, reduced soil erosion, and substantially increased vegetation cover [8]. However, the restoration of vegetation has also brought some new changes [9]. Although vegetation has played a positive role in trapping precipitation and reducing soil erosion, its evapotranspiration water consumption has also increased significantly, which has led to a wide-ranging scientific controversy about the “vegetation–water” relationship [7]. Studies have shown that large-scale vegetation restoration may lead to soil water deficit (i.e., the phenomenon of soil drying) and a reduction in the flow-producing capacity of watersheds, thus threatening regional water security. Therefore, accurate quantification of the impacts of vegetation dynamics on watershed hydrological processes (e.g., runoff, evapotranspiration, and soil water content) is essential for the development of scientific water management strategies [10,11].

Distributed hydrologic models are powerful and promising tools for understanding complex watershed hydrologic processes and assessing the impacts of environmental changes. Among them, SWAT [12] has been widely used globally due to its advantages of having well-defined physical mechanisms and being able to handle long series simulations of large-scale watersheds, etc. [13]. SWAT is able to finely characterize soils by discretizing watersheds into sub-watersheds and hydrologic response units (HRUs), and it can finely characterize the effects of spatial heterogeneity in soil, land use and topography on the hydrological cycle by discretizing the watershed into sub-watersheds and HRUs [14].

Although the SWAT model performs well in humid regions, its built-in vegetation growth module reveals significant structural deficiencies when applied to semi-arid regions characterized by moisture limitation dynamics [15]. The core logic of the module is based on the heat unit theory, which assumes that temperature is the only dominant factor controlling the development of the plant phenology (from germination to maturity). In this theoretical framework, as long as the accumulated heat reaches a specific threshold, the model will simulate the ideal leaf area index (LAI) growth curve, often ignoring the decisive limiting effect of water stress on vegetation growth [16].

In semi-arid regions, such as the Wuding River Basin, moisture rather than heat is usually the primary limiting factor for vegetation growth [17]. Practical observations have shown that the start of the growing season (SOS) and biomass accumulation processes of natural vegetation are closely related to precipitation events and soil moisture conditions [18], and the default module of SWAT tends to simulate a smooth, single-peaked LAI curve with low interannual variability, which is not in line with the highly fluctuating, atypical climatic characteristics of semi-arid vegetation subject to drought events (e.g., stagnant or early growth due to drought, yellowing due to drought) which are seriously inconsistent [19]. In addition, the mechanism of SWAT that triggers vegetation dormancy through day-length thresholds is also mainly applicable to temperate photoperiod-sensitive plants and cannot accurately reflect the vegetation decline induced by water depletion in late fall in semi-arid regions [20]. This distortion of vegetation simulation can produce cascading effects that seriously affect the simulation accuracy of hydrological processes. The LAI is a key parameter that connects the land surface water cycle and energy balance, which directly determines canopy interception and vegetation transpiration rate. When the model overestimates the LAI in dry years or non-growing seasons, it incorrectly calculates excessive evapotranspiration (ET) [21], which leads to underestimation of soil moisture and runoff yield. In the context of ecological restoration on the Loess Plateau, this systematic bias greatly limits the reliability of the model in assessing the hydrological effects of vegetation restoration.

With the rapid development of earth observation technology, long time series and high-resolution spatial and temporal remote sensing data provide new opportunities to correct vegetation parameters in hydrological models [22,23,24,25,26]. Remotely sensed inversion of the LAI can directly reflect the real growth status of surface vegetation [27], which contains comprehensive information on vegetation response to environmental stresses (e.g., drought, pests and diseases, and nutrient deficiencies), and thus circumvents the uncertainty brought by parameterization of empirical growth models.

Among many remote sensing LAI products, the Global Land Surface Satellite (GLASS) LAI product shows unique advantages. Compared with the traditional Moderate Resolution Imaging Spectroradiometer (MODIS) LAI products, GLASS LAI adopts more advanced inversion algorithms (e.g., general regression neural networks GRNNs), and utilizes long time series of surface reflectance data for spatial and temporal filtering and smoothing [28]. In the Loess Plateau region with complex topography and frequent atmospheric disturbances, MODIS products often suffer from missing data and drastic fluctuations due to cloud cover or sensor noise, while GLASS LAI maintains higher temporal continuity and spatial integrity and is able to more accurately capture the climatic characteristics and interannual trends of vegetation.

Recent studies have attempted to assimilate remotely sensed LAI into the SWAT model by directly replacing or modifying the LAI calculation process within the model [29,30], with the aim of improving hydrological simulation accuracy [31,32]. However, a critical knowledge gap remains in the hydrological modeling of semi-arid and arid regions. Unlike humid or tropical regions, where vegetation growth is primarily energy-limited (temperature-driven), vegetation dynamics in semi-arid zones are strongly water-limited and exhibit pulsed responses to precipitation events [33,34,35,36]. The default SWAT model, governed by the heat unit theory, often fails to capture this moisture-driven phenology, leading to systematic biases such as “false vegetation booms” during droughts and consequent overestimation of evapotranspiration. Although high-precision products like GLASS LAI offer a solution, most existing assimilation studies focus on humid regions or rely on simple forcing methods without fundamentally correcting the model’s internal physical mechanism—specifically, the decoupling of vegetation growth from heat accumulation in water-stressed environments. Therefore, there is an urgent need to reconstruct the SWAT source code to restore the physical reality of the Soil–Vegetation–Atmosphere Transport system in semi-arid regions [37].

To accurately represent vegetation dynamics and correct for the tendency to overestimate internal plant growth modules (e.g., the EPIC module in SWAT), recent research has primarily explored three different methodological pathways: dynamic ecohydrological coupling, data assimilation (DA) [32], and direct driving of observational LAI. Dynamic coupling combines hydrological models with sophisticated carbon or crop simulators to capture bidirectional feedbacks, but this technique typically requires extensive parameterization [38]. Data assimilation dynamically updates model state variables based on periodic satellite observations (e.g., using ensemble Kalman filtering); however, it is computationally intensive and remains partially dependent on the baseline accuracy of the internal growth equation. In contrast, the approach adopted in this study, which uses the observed LAI-driven model, provides a more pragmatic and efficient alternative. By completely bypassing the standard plant growth component and replacing the simulated LAI with a continuous, high-fidelity remotely sensed time series, the approach directly eliminates internal systematic biases (e.g., overestimation of LAI > 5.0). While previous studies have attempted to use raw optical satellite data (e.g., standard MODIS products) to drive SWAT [23,30], they often struggle with temporal discontinuities and irregular fluctuations caused by frequent cloud cover. Therefore, the innovation of this work lies in the utilization of high-precision GLASS LAI products as continuous daily drive inputs [28,39]. Unlike previous intermittent drivers or complex DA frameworks, this approach provides a stable, computationally efficient, spatiotemporally continuous correction to daily hydrologic simulations that directly isolates and improves ecohydrologic drivers rather than adding complexity to the internal model.

To address these structural deficiencies, this work aims to develop a physically constrained SWAT model for the Wuding River Basin, a representative semi-arid watershed, by integrating high-resolution spatial and temporal GLASS LAI data. The specific objectives are: (1) Rebuild the vegetation growth module at the source code level: We explicitly replaced the empirical growth equation based on heat units and the dormancy function based on daylength by modifying the code, and established a direct interface to assimilate the external time-varying GLASS LAI data to accurately reflect the real-time vegetation stress response. (2) Improve the efficiency of model calibration: We used an improved dynamic dimension search (DDS) algorithm to perform parameter calibration and validation of the modified model to ensure the accuracy of parameter identification for runoff simulation. (3) Revealing the ecohydrological coupling mechanism: In addition to the improvement in accuracy, we quantitatively assessed how the improved vegetation dynamics altered the tradeoff between blue water (runoff) and green water (evapotranspiration), which provided a new perspective on the mechanism of “vegetation–water–soil” interactions in ecologically fragile semi-arid regions.

2. Materials and Method

2.1. Study Area

The Wuding River Basin is located in the central part of the Loess Plateau. As shown in Figure 1, the entire basin lies between 107°49′ E–110°57′ E and 37°06′ N–39°27′ N. The main stream of the Wuding River is approximately 491.2 km long, with a drainage area of 29,234 km². The elevation ranges from 600 to 1800 m, and the region experiences a typical semi-arid continental climate. The multi-year average temperature is 7.96 °C, and the multi-year average precipitation is 401.23 mm, with precipitation from June to September accounting for 75% of the annual total. The middle and lower reaches of the basin are characterized by a typical loess hilly and gully landscape. The dominant soil type is loessial soil with weak cohesion. During the rainy season, surface runoff on slopes causes severe soil erosion. The primary vegetation in the basin consists of herbs, shrubs, and coniferous trees. The combined effects of topography, climate, soil, and vegetation contribute to the serious soil erosion problem in the Wuding River Basin.

2.2. Data Description

2.2.1. Digital Elevation Model (DEM), Land Use, and Soil Data

DEM, land use data, soil type data, and meteorological information were used as input data to construct the SWAT model. This work utilized 30 m × 30 m resolution Advanced Spaceborne Thermal Emission and Reflection (ASTER) Global Digital Elevation Model (GDEM) data, which were downloaded from the Geospatial Data Cloud platform (http://www.gscloud.cn). The soil data were obtained from the Harmonized World Soil Database (HWSD) with a spatial resolution of 1 km. The data for China within this database were sourced from the China Soil Database and the Cold and Arid Regions Science Data Center, using the FAO-90 soil classification system. Soil information from the HWSD was used to build the soil properties database for the SWAT model. Since some soil property parameters in this database could not be directly input, SPAW software (version 6.02.70) was employed to calculate additional soil properties based on the HWSD information. Land use/land cover (LULC) data were derived from the Chinese Land Cover Dataset (CLCD), a Landsat-based annual land cover dataset for China at the provincial level [40]. In this work, eight reclassified LULC types were used: AGRL (farmland), FRST (forests), PAST (pasture), WATR (water), URBN (urban land), BARR (barren), RICE (paddy rice), and CORN (maize).

2.2.2. Meteorological, Runoff, and Leaf Area Index Data

Meteorological data were sourced from the China Surface Daily Climate Dataset provided by the National Meteorological Science Data Center (http://data.cma.cn/). Daily data, including precipitation, maximum/minimum temperature, wind speed, relative humidity, and sunshine duration from 1 January 2004 to 31 December 2018, were collected from seven meteorological stations within and around the Wuding River Basin (Wushenzhao, Yulin, Yanchi, Dingbian, Jingbian, Hengshan, and Suide). The sunshine duration data, combined with station latitude, were used to calculate solar radiation. Additionally, daily streamflow observation data from 2009 to 2016 at the hydrological station (Baijiachuan) were utilized for model parameter calibration and result analysis. All the data used are shown in

Table 1. Data used in this study, data sources, and relevant information.

Data	Source	Description
DEM	ASTER GDEM (http://www.gscloud.cn (accessed on 30 October 2025))	30 m spatial resolution
Land use/cover	CLCD (https://zenodo.org/records/8176941 (accessed on 1 November 2025))	30 m spatial resolution
Soil	HWSD (https://iiasa.ac.at/models-tools-data/hwsd (accessed on 1 November 2025))	1 km spatial resolution
Observed streamflow	The Hydrological Bureau of the Ministry of Water Resources of China (HBMWRC)	Day (2009–2016, Baijiachuan Station)
Meteorological data	National Meteorological Information Center (https://data.cma.cn/ (accessed on 16 December 2025))	Maximum and minimum daily temperature, rainfall, relative humidity, wind, and solar radiation (2004–2018)
LAI	GLASS (https://glass-product.bnu.edu.cn (accessed on 16 December 2025)) [41]	Day/250 m (2009–2016)

.

2.3. SWAT Model

2.3.1. Overview of the SWAT Model

The SWAT model [12,42] is a semi-distributed, continuous-time, physically based model developed by Dr. Jeff Arnold of the Agricultural Research Center of the United States Department of Agriculture (USDA). The SWAT model is run in daily steps and combines soil, climate, and land use factors to simulate the impacts of the hydrologic cycle, soil erosion, streamflow, water quality pollution, and agricultural management practices on watershed water resources. It can be used to assess soil erosion, hydrologic processes, and water resource management [43].

The study area was the Wuding River Basin in the Loess Plateau of China, and the simulation period was 2009–2016, with 2008 selected as the warm-up period, 2009–2013 as the calibration period, and 2014–2016 as the validation period. A total of 23 subbasins and 309 HRUs were delineated. The SWAT model was run, and subsequent code improvements were based on the SWAT_rev664 version, with no other precalibration adjustments made before the model run. The high-resolution GLASS LAI data were processed into subbasin-level daily .txt files compatible with our modified SWAT source code.

2.3.2. Plant Growth Module in SWAT

The LAI, defined as the ratio of the total area covered by plant foliage to the area occupied by plant growth, serves as a viable variable reflecting plant growth status and crop phenology. It plays a significant role in processes such as precipitation interception, radiation absorption, and transpiration within the hydrological cycle of irrigation districts, exerting a substantial influence on model simulation outcomes. Within SWAT, the HRU constitutes the fundamental simulation unit for most physical processes, encompassing runoff and vegetation growth.

The seasonal variation in the leaf area index is closely related to the dynamic growth of vegetation, and the dynamic growth of vegetation plays an important role in the hydrological cycle. The vegetation growth module in the SWAT model is a simplified version of the EPIC (Environmental Policy Impact Climate) model, which mainly simulates vegetation growth based on two theories: the heat unit theory and the sunshine length threshold theory [39]. (1) Heat unit theory: This theory holds that the growth trajectory of plants (such as plant maturity) is proportional to the temperature increment. In SWAT, plants begin to grow when the average daily temperature is above the base temperature. Heat units (HUs) are calculated by directly accumulating the number of degrees that the daily average temperature is above the base temperature. However, the heat unit theory in the SWAT model does not consider the harmful effects of high temperatures on plants, but it assumes that all temperatures above the base temperature will accelerate the growth and development of crops. (2) Day length threshold theory: This theory is used to define the dormancy period of plants. When the day length reaches the calculated threshold for a location, the plant will enter dormancy. The day length threshold is calculated based on latitude and the dormancy threshold, and different latitudes have different dormancy thresholds.

Derived from temperature requirements (i.e., minimum, maximum and optimum for growth), the heat unit (HU) is an index that is applied to measure the heat accumulation of a plant and is calculated as follows [33]:

$$HU=T_{avg}-T_{base}when{T}_{avg}>T_{base}$$

(1)

where $HU$ is the accumulated heat unit on a certain day during the growing season; $T_{avg}$ is the daily average temperature, °C; and $T_{base}$ is the minimum temperature for the start of plant growth, °C. The total heat unit required for the plant from the start of growth to maturity is calculated as follows [33]:

$$PHU=\sum _{d=1}^{m}HU$$

(2)

where Potential Heat Unit $\left(PHU\right)$ represents the total heat units required for plant growth to maturity, determined by the sowing/planting date (d = 1) and the number of days (m) required for plants to attain maturity. This parameter is computationally derived through input data processing and preloaded in the database before initiating SWAT model simulations.

Vegetation dynamics play an important part in the hydrological cycle. The SWAT model simulates annual LAI dynamics of different vegetation types (warm season annual legumes, cold season annual legumes, perennial legumes, warm season annuals, cold season annuals, perennials, and trees) in each HRU to reflect the vegetation dynamics within the river basin to a certain extent. The vegetation module of the SWAT model is a simplified version of the EPIC model, which simulates vegetation growth mainly on the basis of two theories: heat unit theory and day length threshold theory [44].

$${LAI}_i={LAI}_{i-1}+{\Delta LAI}_i$$

(3)

$${\Delta LAI}_i=\left({fr}_{LAImx,i}-{fr}_{LAImx,i-1}\right)·{LAI}_{max}·\left(1-\mathit{exp}\left(5\left({LAI}_{i-1}-{LAI}_{max}\right)\right)\right)$$

(4)

$$HRU={\displaystyle \frac{{\sum }_{i=1}^n\left(gri{d}_{value}\times overla{p}_{area}\right)}{Total}}$$

(5)

where ${fr}_{LAImx}$ is the fraction of the plant’s maximum leaf area index for the plant; ${\Delta LAI}_i$ is the change in the LAI on day $i$; and ${fr}_{LAImx,i}$ and ${fr}_{LAImx,i-1}$ are the fraction of the plant’s maximum leaf area index for days $i$ and $i-1$.

Figure 2 shows the detailed steps to improve the SWAT code, where the blue parts of the figure are the modules or steps modified based on the original subroutine, and the red parts are the new subroutine modules. The improved model code consists of four parts, including: (1) data structure definition management, (2) initialization of external input data, (3) temporal simulation control, and (4) core intervention code (LAI replacement process). The core mechanism is to add a new Fortran file named “LAI.f” in the source code of the SWAT model (based on the rev_664 version). The function of this file is to read the external GLASS LAI data and modify the original “grow.f” file of the model code so that when the model runs, it reads the GLASS LAI data instead of the LAI result calculated by the EPIC model.

Table 2. Classification and functions of modified and newly added subroutines in the SWAT model.

Subroutine (.f)	Type	Function Description
Modparm.f	Modified	Declared global allocateable arrays to store external GLASS LAI data.
allocate_parms.f	Modified	Allocated memory for GLASS arrays based on the number of HRUs and simulation days.
LAI.f	New	Developed to read external time-series GLASS LAI data (.txt).
readbsn.f/readhru.f	Modified	Adjust initialization operation.
zero.f	Modified	Initialized state variables related to the external LAI data.
clicon.f	Modified	Modified to read and pass daily external LAI values along with weather data.
simulate.f	Modified	Updated the year/day loop controller to synchronize external data injection.
command.f/subbasin.f	Modified	Updated the watershed and subbasin processing loops to support data transfer to HRUs.
plantmod.f	Modified	Modified the operation of the vegetation module.
grow.f	Modified	Injected logic to force-replace the EPIC-calculated LAI with GLASS LAI data.

lists the specifically modified and added Fortran subroutines.

2.3.3. Mapping Remote Sensing LAI Data to HRUs

The SWAT model relies on empirical formulas for LULC and crop growth stages to estimate the LAI during its default modeling process [45]. This approach often fails to capture the spatial heterogeneity and temporal dynamics of the LAI within complex watersheds. To address this limitation, we integrated remotely sensed LAI data with high spatiotemporal resolution and direct observational characteristics into the SWAT model. For this purpose, a Python-based automated processing workflow was developed to handle raw remote sensing LAI data and map it to the study watershed’s HRUs, ensuring consistency with model input requirements. The entire processing workflow is divided into five steps.

(1) Data Preprocessing: From HDF to Standardized Raster.

The remote sensing raster data originate from raw leaf area index datasets (2009–2016) stored in Hierarchical Data Format (HDF). This format is specifically designed for large-volume MODIS-type datasets but is typically incompatible with standard spatial analysis tools. To ensure accurate data extraction, we employed the Geospatial Data Library (GDAL) to traverse directories for specific years and convert HDF files into the more universal GeoTIFF format. During this process, only the primary LAI bands were retained. Simultaneously, following the remote sensing dataset documentation, pixels with values greater than or equal to 101 (typically indicating cloud interference or sensor noise) were reclassified as missing values. By using GeoTIFF-formatted data to drive the model, the spatial metadata essential for subsequent SWAT model replacements was preserved, enabling analysis of the effectiveness of SWAT model improvements.

(2) Standardization of the Spatial Reference System.

Typically, remote sensing-derived LAI data and the SWAT model exhibit spatial distribution discrepancies, which often result in mismatches between remote sensing LAI and a study area’s HRUs. This discrepancy leads to errors in SWAT model simulation results. Using Arcpy, batch projection operations were performed on GeoTIFF files. The Project Raster tool was employed to uniformly convert the input layers to the EPSG 32649 projection coordinate system. This ensured consistency between the remote sensing data and the spatial coordinates of the DEM and LULC data. Furthermore, each LAI pixel accurately reflected its actual ground location prior to further cropping and resampling.

(3) Watershed-Scale Raster Clipping.

The GLASS LAI data are categorized according to MODIS regional grid cell numbers. This work selected data from the h26v05 grid cell within the MODIS China regional grid, which generated redundant data for the study area. We also developed an automatic batch clipping program based on the vector data of the subbasins in the study area. The missing data were represented by zeros.

(4) Data resampling processing.

Resampling operations were required prior to partition statistics to address the mismatch between the spatial resolution of GLASS LAI data and that of the HRUs. The original resolution of the GLASS LAI data is 250 m. However, in this work, the SWAT model’s HRUs are delineated based on high-resolution 30 m digital elevation models and land use data. Consequently, the spatial resolution of HRUs is 30 meters. An overly coarse spatial resolution for the LAI would obscure critical spatial heterogeneity, thereby compromising simulation outcomes. To ensure spatial consistency across all model inputs, the LAI raster data were resampled to 30 meters resolution. Bilinear interpolation was employed for this resampling operation, as it effectively preserves the continuous spatial gradient of vegetation canopy cover compared to nearest-neighbor interpolation. This is crucial for capturing growth processes such as vegetation evapotranspiration at the watershed scale.

Unlike the nearest neighbor interpolation method, which produces unnatural boundaries when enhancing the spatial resolution of raw data, adjacent HRUs may exhibit abrupt numerical jumps in LAI values despite their close spatial proximity. The bilinear method, however, generates a smoother, more natural, and physically consistent LAI surface by calculating the weighted average of the four nearest 250-meter pixels. Its mathematical implementation is expressed as follows [46]:

$$\mathrm{L}\mathrm{A}\mathrm{I}\left(x,y\right)={\displaystyle \frac{1}{(x_2-x_1)(y_2-y_1)}}\left[\begin{array}{cc}{x}_2-x& x-x_1\end{array}\right]\left[\begin{array}{cc}LAI\left(Q_{11}\right)& LAI\left(Q_{12}\right)\\ LAI\left(Q_{21}\right)& LAI\left(Q_{22}\right)\end{array}\right]\left[\begin{array}{c}{y}_2-y_1\\ y-y_1\end{array}\right]$$

(6)

where $\mathrm{L}\mathrm{A}\mathrm{I}\left(x,y\right)$ represents the LAI value after resampling; $Q_{\mathrm{i}\mathrm{j}}$ denotes the surrounding four original 250 m pixel centers; and ${LAI(Q}_{\mathrm{i}\mathrm{j}})$ refers to the raw observation values at those locations. By performing three linear interpolations within the study area, this method ensures that the reduced leaf area index data align with the spatial resolution of HRU boundaries while preserving the physical characteristics of the original canopy structure.

(5) Extraction of LAI from HRU-level statistical data.

Following the generation of a high-resolution (30-meter) continuous leaf area index (LAI) surface, the next step involves discretizing this spatial distribution data into the centered HRUs required by SWAT. In this paper, we employed a Python (version3.10)-based aggregation algorithm to map the reclassified LAI data onto the HRUs. The zonal statistics method based on hydrological response units was used to assign the high-resolution leaf area index grid data to each HRU. The specific steps are as follows: First, using the spatial analysis function of the ArcGIS 10.8.2 platform, the HRU vector boundary was used as the statistical area, and the “HRUGIS” field was used as the unit identifier to perform zonal statistical calculation on the LAI raster data of each time step to obtain the average LAI value in each HRU.

The spatial mapping was executed using the ZonalStatisticsAsTable function within the ArcPy Spatial Analyst module. For each simulation time step, the algorithm overlaid the HRU vector boundaries onto the corresponding LAI raster [26]. To ensure phenological accuracy, we calculated the arithmetic mean of all valid pixels falling within each HRU polygon. The mean algorithm was used in the statistical process, and the calculation formula is:

$${\overline{LAI}}_{{HRU}_i}={\displaystyle \frac{1}{n}}\sum _{j=1}^n{LAI}_{{pixel}_j}$$

(7)

where ${\overline{LAI}}_{{HRU}_i}$ denotes the average LAI for the i-th HRU unit, ${LAI}_{{pixel}_j}$ represents the LAI value of the j-th effective pixel within that HRU, and n signifies the total number of effective pixels within the HRU.

In summary, the Python-based automated processing code and methodologies bridge the gap between raw remote sensing observations and the discrete computational requirements of the SWAT model. This resolves the challenge of establishing a mapping between GLASS LAI data and the HRU scale, ensuring the preservation of the original physical fidelity of canopy structure throughout the conversion process from 250 m to 30-m HRU resolution. The transition from temperature and moisture-driven empirical formulae to observation-based high-resolution remote sensing data provides the essential spatial heterogeneity and temporal dynamics required for ecohydrological modelling in complex semi-arid catchments. Ultimately, this enhances the accuracy of SWAT model simulations for processes such as canopy interception, vegetation evapotranspiration, and runoff.

Figure 3 shows a system architecture diagram illustrating the GLASS LAI data preprocessing pipeline (left) and its explicit integration point within the SWAT daily execution loop (right). The red module highlights the exact code intervention where remote sensing observations bypass the default empirical heat unit growth equations.

2.3.4. Implementation of the Improved Fortran-DDS Algorithm

In this work, the DDS algorithm is used as the core optimization engine. DDS is a heuristic global search algorithm designed for high-dimensional watershed models [47]. The core idea is not to optimize all parameters in each iteration but to probabilistically select a portion of parameter dimensions to be perturbed according to the ratio of the current iteration number to the maximum iteration number [48].

The algorithm’s parameter update mechanism at the $i-th$ iteration is as follows: for the $j-th$ decision variable (i.e., the model parameter), the decision of whether or not to update is made based on the probability $P\left(i\right)$:

$$P\left(i\right)=1-{\displaystyle \frac{\ln \left(i\right)}{\ln \left(i_{max}\right)}}$$

(8)

where $i_{max}$ is the set maximum number of iterations. The selected parameters will generate new candidate solutions by applying random perturbations obeying the standard normal distribution on the basis of the current optimal solution (Best Solution). This mechanism ensures that the algorithm carries out a large-scale global exploration at the beginning of the search, and automatically switches to a fine search for local optimal solutions at a later stage, so that the global optimum can be quickly approximated with limited computational resources.

In order to realize the efficient integration of algorithms and models, a modular rate-setting system based on the Fortran-90 standard is developed in this work. The system mainly consists of the following three core modules [49]. (1) Parameter control module (mod_dds): The built-in logic of the improved DDS algorithm is responsible for generating parameter perturbation sequences and dynamically updating the global optimal parameter set according to the feedback of the objective function. (2) File interaction module (mod_swat): A dedicated parameter writing interface is developed for the input file structures of the SWAT model (.mgt, .hru, .sol, .gw, etc.). The interface supports “absolute value substitution” or “physical interval constraints” for 22 key hydrological parameters, which ensures that the generated parameter combinations conform to the physical reality of the watershed and avoids a model crash due to the parameter being out of bounds. (3) Simulation Driving and Evaluation Module: The system automatically drives the SWAT kernel to execute the simulation through the system call command and extracts the simulated runoff of Subbasin 23 at the outlet of the Wu Ding River based on the output river algorithm file (output.rch).

By combining the hydrological characteristics of the Wuding River Basin and the production and catchment mechanism of the SWAT model, 22 high-sensitivity parameters were selected to participate in the rate determination in this work, which covers the four physical processes of surface runoff, soil water movement, groundwater recharge, and channel modeling. The main parameters include the runoff curve number (CN2), which reflects the characteristics of the subsurface, Manning’s coefficient (OV_N), which controls the confluence of the slope, the compensation coefficient (ESCO), which affects the evapotranspiration of soil moisture, and groundwater parameters (ALPHA_BF, GW_DELAY), which control the subsidence of baseflow. All parameters are set with clear physical upper and lower bounds to ensure the reasonableness of the optimization results.

3. Results

3.1. Model Calibration and Validation

After constructing the models, calibration of runoff simulation parameters was performed for both the original and improved models. For the improved model, code-level modifications to the SWAT model were performed in order to achieve automatic calibration of the distributed basin hydrological model and to take full advantage of the sensitivity information from the optimization process. In this work, a modified version of the DDS algorithm is used for rate calibration and validation of hydrological parameters, using the same type and number of parameters as the original model. In order to maintain the consistency of the results, the original model was also calibrated using the DDS algorithm for the original model. Theoretically, the improved DDS algorithm can achieve better calibration results in a limited number of iterations [50].

3.1.1. Sensitivity Analysis and Parameter

The model’s evaluation period was from 1 January 2009 to 31 December 2013, while the verification period was from 1 January 2014 to 31 December 2016. Based on the method for assessing parameter sensitivity mentioned earlier, the following 22 calibration parameters were selected, as shown in

Table 3. List of parameters produced for SWAT calibration.

Parameter	Definition	Initial Range	Fitted Values
CN2	SCS runoff curve number for moisture condition II	35–98	54.78
SURLAG	Surface runoff lag coefficient	0.05–24	1.05
CANMAX	Maximum canopy storage	0–100	24.4
OV_N	Slope Manning’s coefficient	0.01–1	0.24
SLSUBBSN	Average slope length (m)	−0.25–0.25
HRU_SLP	Average slope steepness	0	1
ESCO	Soil evaporation compensation factor	0–1	0.98
EPCO	Plant uptake compensation factor	0–1	0.10
EVRCH	Reach evaporation coefficient	0–1	0.68
SOL_AWC	Available water capacity	0–1	0.75
SOL_K	Saturated hydraulic conductivity	−0.5–0.5	−0.39
ALPHA_BF	Threshold depth for revaporization from the shallow aquifer	0–1	0.43
GW_REVAP	Effective hydraulic conductivity of channel	0.02–0.2	0.16
GWQMN	Denitrification exponential rate coefficient	0–5000	4145
RCHRG_DP	Denitrification threshold water content	0–1	0.49
CH_N2	Nitrate percolation coefficient	−0.01–0.3	0.007
CH_K2	Phosphorus percolation coefficient	−0.01–100	67.37
CH_K1	Phosphorus soil partitioning coefficient	0–300	107
ALPHA_BNK	First-order rate constant for denitrification	0–1	0.018
TRNSRCH	Monod half-saturation term for denitrification	0–1	0.05
USLE_K	USLE equation soil erodibility factor	0–0.65	0.25
USLE_P	USLE support practice factor	0–1	0.17

. At the same time, the optimal parameter values obtained after calibrating the improved model through 500 iterations of the DDS algorithm are also listed.

3.1.2. Streamflow

Figure 4 presents the calibration and validation results for runoff at the watershed outlet. The calibration period spans from 2009 to 2013, while the validation period covers 2014 to 2016. To assess the accuracy of runoff simulation, this work employs two performance metrics: the Nash efficiency coefficient (NSE) and the coefficient of determination (R²) [14,51]. The improved model yielded good simulation results throughout the simulation period. For the monthly scale runoff simulation, the “good” ranges of R² and NSE are 0.65~0.75 [18], while the values of R² and NSE are 0.71 and 0.7, respectively, for the daily runoff rate period of the model, which proves the accuracy of the improved model in the direction of the simulation of watershed runoff. During the model validation period, the simulation results still maintained a good level, with R² and NSE of 0.58 and 0.51, respectively, although lower than that of the rate period. However, when using a monthly time step to evaluate the runoff, NSE > 0.5 can be regarded as a satisfactory evaluation criterion for the model simulation results, while the time step used in this work is a daily scale, and the criterion for the day-by-day runoff simulation can be appropriately relaxed. Overall, the improved model showed strong performance in runoff simulation, accurately capturing the time-scale dynamics of actual runoff in the study area. This demonstrates the reliability and validity of the model in simulating hydrological processes, which is crucial in providing information for water management decisions and ecohydrological studies in the region.

3.2. Comparison of SWAT Model Simulation Results Before and After Improvement

This research presents a comparative analysis of daily runoff in the Wuding River Basin from 2009 to 2016, utilizing three indicators (R², NSE, KGE) typically employed in the assessment of distributed hydrological models [42], with 2009–2013 designated as the calibration period and 2014–2016 as the validation period.

Table 4. Comparison of evaluation parameters before and after model improvement. “Original” refers to the original model without modification of the plant generation module, and “GLASS-modified” refers to the modified SWAT model with integrated GLASS LAI.

Period	Plant Growth Module	R2	NSE	KGE
Calibration	Original	0.52	0.52	0.56
	GLASS-modified	0.71	0.7	0.79
Validation	Original	0.21	0.2	0.32
	GLASS-modified	0.58	0.51	0.64

presents the simulation results before and after the enhancements. The refined model utilizing GLASS LAI data demonstrates a marked improvement in runoff simulation accuracy relative to the original model. Furthermore, the enhanced model more precisely replicates the daily runoff process during both the calibration and validation periods, indicating its superior capability in simulating the dynamic growth of vegetation. Simultaneously, the model’s simulation performance throughout the validation period aligns with expectations, achieving a good standard according to the evaluation criteria, thus demonstrating the upgraded model’s applicability in subsequent years.

The results in Figure 4 primarily reflect two improvements using GLASS LAI. The first is a correction for the underestimation of the original simulated runoff peaks. This improvement is particularly evident in the extreme precipitation events throughout the simulation period, with the greatest improvement obtained especially during the 2013–2016 flood season. In terms of specific simulation results, the measured peak runoff is about 350 m³/s (gray line) during the extreme flood event of the 2013 flood season. However, the default model only simulated a peak runoff of about 150 m³/s (blue line), severely underestimating the peak result, which matches the LAI calculation of the default model. The higher LAI results in the default model resulted in excessive canopy interception and evapotranspiration, which reduced the peak runoff. In contrast, the improved model with the introduction of GLASS LAI simulated peak runoff more accurately due to a more accurate assessment of the actual value of the LAI. By correcting the summer leaf area index to actual values, the improved model reduced canopy interception, effectively releasing surface runoff previously reduced by the model’s overestimation of the LAI. The second improvement of the improved model is the optimization effect on the simulation results during the receding water stage and the winter drought period. The improved model exhibits lower runoff values in the out-of-season simulations. The default model (blue line) tends to be higher than the observed values during this phase, maintaining high base flow rates (e.g., more than 30 m³/s during the dry season), suggesting an underestimation of soil moisture depletion by vegetation. The improved model (red line) simulated lower values during the off-season compared to the original model, reflecting higher LAI values captured by remote sensing data and enhanced soil water depletion driven by the ecological phenomenon of “early green-up” [52].

Overall, the improved model more accurately reflected the nonlinear hydrological response in the semi-arid region of the Wuding River Basin. Incorporating the remotely sensed leaf area index (GLASS-LAI) into the model not only improved the accuracy of peak flow but also corrected the systematic bias in runoff simulation.

3.3. Analysis of Spatial and Temporal Dynamic Differences Between Simulated and Remotely Sensed LAI

The significant difference between the vegetation growth simulation defaulted by the SWAT model and the remote sensing observation is the source of the bias in the hydrological simulation. By comparing the simulated LAI and GLASS LAI data (Figure 5) for the major land use types in the watershed (AGRL, CORN, RICE, PAST), the actual situation of vegetation generation for different land uses in the semi-arid watershed was revealed. At the same time, MCD15A2H LAI data were introduced as a control. The results reflect the temporal and spatial consistency between GLASS LAI and MODIS LAI, demonstrating the feasibility of using GLASS LAI to replace the original SWAT model simulation.

3.3.1. Fundamental Differences in Climatic Characteristics

The LAI curves simulated by the SWAT-Default model show a typical “single-peaked normal distribution”, which is an inherent characteristic of the heat-driven EPIC model. The model assumes that vegetation will follow an optimal trajectory to maturity as long as temperatures are favorable. However, the GLASS LAI data reveal the complexity and volatility of vegetation growth in the Wudi River Basin.

(1) Atypical growth curves.

The measured LAI curves often show multi-peaked or jagged fluctuations, which are a direct reflection of the vegetation’s pulsed response to precipitation in the semi-arid zone. For example, during the spring droughts of 2011 and 2015, GLASS LAI showed a significant delay in vegetation rejuvenation and slow growth rates, while the default model still simulated a rapid growth trend. This “false boom” caused the original model to grossly overestimate vegetation water consumption in the dry years.

(2) Difference in growing season length.

The default model strictly limits the growing season by temperature and day length. Comparisons show that GLASS LAI tends to reflect a longer growing season than model simulations. Particularly in the fall, the data in Figure 5 show that corn cropland and grassland maintained their greenness for a longer period of time even after the temperature decreased, which is important for moisture cycling in the fall and winter seasons.

3.3.2. Differences in Response of Different Land Use Types

Default models tend to over-simulate the LAI in the Wuding River Basin; however, GLASS data show that the peak LAI of each land use in the region is usually maintained between 1.5 and 2.0 due to the limitation of moisture carrying capacity. By correcting for this overestimated bias, the SWAT-LAI model effectively reduces the amount of canopy interception and evapotranspiration calculations of the vegetation, which explains the significant reduction in the peak simulation bias in the runoff simulation.

Taking the cropland land use type as an example, as shown in Figure 5b,c, the peak LAI values simulated by the default model were seriously high, exceeding 5.0 for maize and 4.5 for rice, which exaggerated the canopy density of the cultivated land in the semi-arid watershed area, and the remote sensing LAI peaks were corrected to a reasonable range of 1.5–3.0. This correction process provides a physical basis for the increase in the peak simulation value of the improved model mentioned above, i.e., the improved model reduces the loss of vegetation canopy retention during the flood season. The default model strictly limits the growing season by temperature and day length. Comparisons show that GLASS LAI tends to reflect a longer growing season than model simulations. Particularly in the fall, the data in Figure 5 show that corn cropland and grassland maintained their greenness for a longer period of time even after the temperature decreased, which is important for moisture cycling in the fall and winter seasons.

3.4. Spatial Distribution Characteristics of Vegetation Phenology Indicators

The start time (SOS) and end time (EOS) of the growing season, two meteorological parameters derived from the GLASS LAI data, demonstrated the regional heterogeneity in vegetation phenology in the Wuding River Basin (Figure 6).

3.4.1. Spatial Gradient of Phenology

The spatial distribution map indicated that the vegetation phenology in the Wuding River Basin showed a significant south–east–northwest gradient, which was highly consistent with the precipitation and temperature gradients in the basin.

(1) SOS distribution.

The southeastern part of the Wuding River Basin has better hydrothermal conditions, and the SOS is earlier, concentrating on the 100th–120th day (mid to late April); whereas, advancing northwestward to the semi-arid areas of the Loess Plateau where precipitation is scarce, the SOS is significantly delayed to the 130th–150th day (early to mid-May), due to the slow warming in the spring and the frequent droughts.

(2) EOS distribution.

The end was mainly distributed on days 285–320 (mid-October to mid-November). EOS was generally earlier in the NW than in the SE due to early frost.

3.4.2. Interannual Trends in Phenology

Analysis of the 2009–2016 time series showed a slight basin-wide trend of earlier SOS (−0.5 days/year) and later EOS (+0.8 days/year), resulting in an overall longer growing season. This climate change, driven by climate warming and humidification, implies an expansion of the vegetation transpiration window period. The SWAT-Default model is unable to dynamically adapt to this interannual climate-driven climatic drift due to its fixed parameters. The SWAT-LAI model automatically captures this climate change signal by assimilating remote sensing data, providing a more reliable basis for assessing long-term water resources evolution under climate change.

3.5. Changes in Ecohydrological Processes After Improved Modeling

3.5.1. ET

In this work, the multi-source fused ET dataset (2000–2020) of the Yellow River Basin was selected as the reference data. This dataset has a high spatial resolution of 0.1°, and by fusing the multi-source remote sensing inversion products with ground observation data, it effectively overcomes the accuracy limitation of a single remote sensing product under the complex subsurface of the Yellow River Basin. In the data processing stage, the study period of 2009–2016 was selected, and the original NetCDF format regional evapotranspiration grid point data of the Yellow River Basin were spatially cropped and masked using the vector boundary of the Wuding River Basin, ensuring that the spatial extent was completely consistent with the basin definition of the SWAT model. For multi-scale comparison with the simulation results output from the model, we extracted the corresponding time series of grid points for representative geographic locations within the watershed and treated them as daily-scale ET data for the virtual observatory sites, so as to assess the capturing ability of the simulated values at the point scale.

For validation at the overall watershed level, an area-weighted averaging method was used to aggregate the grid-point-scale ET data into surface ET volumes. In particular, all valid grid cells falling within the watershed boundary were identified, their day-by-day ET values were extracted, and the daily average ET intensity for the whole watershed was calculated using the grid area as the weight. This process is capable of smoothing the high-resolution remote sensing fusion data to the watershed scale for direct matching with the total watershed ET output from the SWAT model, reducing the uncertainty in the data application process.

Boxplots can visualize the concentration trend and discrete distribution of leaf area index (LAI) values by showing the median, interquartile spacing (IQR) and outliers of the data. By comparing the simulation results of the default model (SWAT-Default) and the modified model (SWAT-LAI/GLASS Modified) for different land use types during the period of 2009–2016 (Figure 7), as well as the monthly-scale temporal distributions of ET (Figure 8), significant distributional differences can be found.

As shown in Figure 7, the ET simulation results indicate that the average annual ET of the improved model compared with the original model has been improved in the case of different land use types. The median annual ET of the AGRL land use type is 332.8 mm, while the value of the improved model is 373.3 mm, which is equivalent to 12.2% growth. Grassland (PAST) also has a certain degree of growth, and the median ET of the improved model is 299.4 mm, and the improved model is 356.8 mm, with a growth rate of nearly 20%. The median value of ET before improvement was 299.4 mm, while the value of ET after improvement was 356.8 mm, with a growth rate close to 20%. These results are consistent with the LAI growth and decline pattern in Figure 5, i.e., for the improved model LAI, despite being smaller at the peak, the earlier SOS and later EOS periods resulted in a greater overall cumulative LAI value and a larger ET.

In terms of the month-by-month scale change in ET, the time scale distribution of the improved model and the original model showed completely different characteristics. As shown in Figure 8, taking the agricultural land use type (AGRL) as an example, the multi-year monthly average evapotranspiration of the improved model from July to September was significantly higher than that of the original model during the period of 2009–2016. The original model had an average ET of 91.91 mm in July, while the improved model had an average ET of 79.37 mm. The original model also has a certain degree of lead in August and September compared to the improved model, which leads to the problem of peak underestimation in the runoff model of the original model. In the non-flood season from January to May, the ET of the improved model was higher than that of the original model, which was attributable to the use of GLASS LAI data by the improved model, which simulated the LAI values more accurately than the original SWAT model in the dry season (Figure 5). Over the 2009–2016 period, the average daily ET values before and after the SWAT model revision reflect the same simulation trends and results. During the first 150 days of the year, the ET simulation of the revised model is higher than that of the original model, while the peak results are reversed (Figure 9).

With the addition of GLASS LAI, there is a clear shift in the evapotranspiration pattern from a single peak to a double peak (Figure 9). The first peak occurs around DOY 150 (late spring) and is due to an improved depiction of spring phenology (winter wheat regrowth and early-season forage growth), which is ignored by the default heat unit-driven model. The second peak occurs around DOY 200 and corresponds to the summer monsoon growth phase. This shift suggests that the modified model successfully corrected for the systematic underestimation of early-season ET and overestimation of summer ET, thus simulating the reality of vegetation–water interactions in the semi-arid Indus River Basin. The actual evapotranspiration data (AET) also showed the same characteristics as the ET simulated by the improved model. A comparative analysis of the multi-year average day-by-day ET and AET before and after the improvement is presented in Figure 10, where the ET simulated by the improved model had a higher R² (0.92) with the AET compared with the original model (R² = 0.75).

Figure 11 illustrates the changes in water flux simulations for the original and improved models from 2009 to 2016, providing watershed-scale validation of the effectiveness of the improved model. Overall, the improved model simulated higher ET for the watershed than the original model, with a multi-year average ET of 251.659 mm for the original model and 341.78 mm for the improved model. This corresponds to a decrease in blue water, which is defined in this work as the total water production generated in the watershed, including surface runoff, lateral flow, and groundwater recharge. It represents the amount of net water resources available for stream discharge. The multi-year average blue water in the watershed was 161.8 mm before the improvement and 84.1 mm after the improvement. The multi-year average percentage of blue water to ET decreased from 63.8% to 24.6%. The results showed that the effect of the longer growing season on annual water production was much greater than the reduction in the peak LAI. The cumulative effect of continuous transpiration during the non-flood season in the improved model offset the reduction in transpiration during the flood season. Combined with the simulation results in Figure 4, the sharp decrease in blue water/ET is due to the improved model simulation giving more priority to vegetation transpiration (green water), which depletes the source of base flow, resulting in lower dry season flows. However, the reduction in the LAI during the flood season results in higher flood peaks, which allows the runoff simulation to be closer to the observed values.

3.5.2. LAI

Figure 12 and Figure 13 show the average seasonal dynamics of the LAI and its peak distribution for different land use types (AGRL, PAST, CORN, RICE). The results show that GLASS LAI has significantly modified the vegetation climate simulation in the original model. In terms of climate characteristics, the default EPIC module shows obvious “lag” and “sudden change” characteristics: the start of the growing season (SOS) is generally delayed by about 30–45 days compared with the actual situation, and the LAI often remains high in the dry season (November–December) (e.g., the plateau period of PAST in Figure 12b), which led the model to calculate spurious transpiration water consumption in the non-growing season. In contrast, the GLASS-LAI sequence accurately captured the SOS and EOS of the vegetation in the semi-arid zone, and the recession of its growth curve was highly synchronized with the natural climatic rhythm of the basin.

At the peak magnitude, the boxplot clearly reveals the overestimation of the default model. For major crops (CORN and RICE), the median peaks of EPIC-LAI exceeded 3.0, and some HRUs even exceeded 5.0, far exceeding the actual carrying capacity of the semi-arid zone of the Loess Plateau. By integrating satellite observations, GLASS-LAI constrained the peak values of each land use type to a reasonable peak range of 1.5–3.0, which effectively alleviated the overestimation of evapotranspiration in the original model. The introduction of GLASS LAI optimized the allocation between canopy retention and infiltration production in the SWAT model, which significantly improved the simulation of runoff during flood and dry periods.

4. Discussion

4.1. Physical Mechanisms for Improving the Model to Enhance the Accuracy of Hydrological Simulations

The core finding of this work is that using GLASS remote sensing data to replace the default EPIC vegetation growth module of SWAT can significantly improve the accuracy of runoff simulation in the semi-arid Wuding River Basin (0.19 improvement in R² and 0.18 improvement in NSE for the rate period; 0.37 improvement in R² and 0.31 improvement in NSE for the validation period). This improvement is not only attributed to the correct calibration of the DDS algorithm but also to the refinement of the physical structure of the model, i.e., the restoration of the physical reality of the Soil–Vegetation–Atmosphere Transport (SVAT) system.

The significant improvement in runoff simulation, particularly during low-flow periods, can be attributed to the correction of the vegetation-driven water consumption mechanism. In the SWAT model framework, the LAI is a critical driver that governs the canopy resistance $r_c$ within the Penman–Monteith ET module. The physical relationship is defined as follows:

$$r_c={\displaystyle \frac{r_{leaf}}{0.5·LAI}}$$

(9)

where $r_{leaf}$ is the minimum effective stomatal resistance of a single leaf and the factor 0.5 represents the effective leaf area participating in active transpiration (accounting for shading effects within the canopy).

In the default SWAT model, the empirical EPIC module systematically overestimates the peak LAI in the semi-arid Wuding River Basin (often exceeding 5.0). According to the equation, this unphysical LAI inflation artificially minimizes the canopy resistance $r_c$, acting as a “vacuum” that draws excessive moisture from the soil and shallow aquifers. By integrating high-resolution GLASS-LAI, the model constrains the vegetation growth within a physically realistic range (1.5–3.0). This correction restores $r_c$ to its rational level, effectively reducing “artificial water consumption” during the growing season. Consequently, the improved model maintains a healthier water balance and preserves the natural recession dynamics of the basin. The causal chain (1) realistic LAI, (2) reasonable $r_c,$ (3) balanced ET, and (4) robust baseflow is clearly evidenced by the superior performance of the modified model in capturing the flow duration curve (FDC) during extreme dry years.

By assimilating GLASS LAI, the model is actually driven to accept the vegetation response mechanism to environmental stress. When the remotely sensed LAI is low, the model automatically adjusts canopy resistance upwards, inhibiting transpiration and reducing interception. This mechanism allows the SWAT-LAI model to simulate the process of “adaptive water conservation by vegetation”, which retains more water in the soil to maintain baseflow or to respond to stormwater runoff. This physical modification of the internal fluxes of hydrological processes is the fundamental reason for the robust performance of the improved model during the validation period in the context of climate change.

4.2. Uncertainties and Model Limitations

To assess the robustness and uncertainty of the model under extreme climate change, this study used the standardized precipitation anomaly level (SPA) to identify 2010 as an extreme drought year during the study period (SPA < −1, Figure 14a). Analysis of flow duration curves (FDCs) for this year showed (Figure 14b) significant structural differences in simulation results between the pre- and post-improvement models during dry season runoff (p > 95%). The default SWAT model seriously overestimates the base flow during the dry season, its flow quartile $Q_{95}$ value (21.4 m³/s) is about six times the measured value (3.81 m³/s), and the end of the curve shows an irrational “plateau,” which fails to capture the natural recession dynamics of the basin. With the introduction of GLASS LAI, the improved model corrects $Q_{95}$ to 3.9 m³/s, which is closer to the measured value. Meanwhile, the improved model is more accurate in the simulation of the peak value, which is closer to the observed value than the original model. This demonstrates that the correlation between vegetation phenology and GLASS LAI is still robust under extreme drought conditions.

Although the introduction of GLASS data has significantly improved the modeling accuracy, uncertainties still exist. First, the remotely sensed LAI product itself suffers from inversion errors in complex terrain (gully areas), especially in sparsely vegetated areas where soil background reflectance may interfere with the signal. Second, although the LAI was corrected, the root depth (Root Depth) parameter in the SWAT model still mainly varies statically with the crop growth stage and fails to dynamically reflect the active water foraging behavior of the plant root system (e.g., root rooting down in drought period). Future research should consider combining the dynamic root growth module with remote sensing LAI assimilation to realize the comprehensive coupling of surface–subsurface water processes.

Subsequently, deep learning algorithms such as Physics-based Neural Networks (PINNs) will be combined to narrow the gap between deterministic physical equations and random variables to enhance hydrologic model simulation [53]. Combining the interpretability of the physical mechanisms of the SWAT model with the high-dimensional solving capabilities of deep learning is also a potential approach to capture the relationship between surface runoff and vegetation.

5. Conclusions

Aiming at the core problem of inaccurate vegetation dynamics portrayal in hydrological simulation in semi-arid areas, this work proposes a SWAT model improvement scheme integrating high-resolution spatial and temporal GLASS LAI remote sensing data, taking the Wuding River Basin of the Loess Plateau as an example. The refined hydrological simulation driven by observed data was realized by reconstructing the model source code and replacing the traditional crop growth module based on heat units. The improved model corrected the peak farmland LAI to a reasonable interval of 1.5–3.0, effectively corrected the physical basis of canopy interception and transpiration calculations, and thus significantly improved the simulation accuracy of hydrological processes, which was verified by comparing the runoff data. The improvement in the runoff simulation accuracy was relatively significant (the rate period R² was improved from 0.52 to 0.71, and the NSE was improved from 0.52 to 0.7; the verification period R² improved from 0.21 to 0.58 and NSE improved from 0.2 to 0.51). In addition, the improved model produced significant changes in the elements of watershed LAI, ET, and surface runoff (blue water). The work demonstrated that improving the vegetation growth module of the SWAT model can effectively improve the accuracy of hydrological simulation in semi-arid regions, and this improved method can also be used in other watersheds to provide guidance for refined water resources management at the watershed scale.