Remote Sensing-Based Attribution of Crop Water Requirements Dynamics in the Tailan River Irrigation District, 2000–2024

Abstract

Assessment of crop water requirements (ETc) and their meteorological driving mechanisms are critical for irrigation management in arid inland river basins. Taking the Tailan River Irrigation District (Xinjiang, China) as a case study, temporal changes in cropping structure, crop-specific ETc, and irrigation-district–scale agricultural water demand, as well as the meteorological controls on ETc, were quantified for the period 2000–2024 using Google Earth Engine-based crop mapping, the CROPWAT model, and path analysis. The results demonstrated that the 2024 random forest classification model achieved high accuracy (overall accuracy = 0.902; Kappa = 0.876), and validation against statistical yearbook data confirmed the reliability of crop-area estimation. Cotton dominated the cropping structure (228.6–426.0 km²), while the orchard area expanded markedly from 206.5 km² in 2000 to 393.2 km² in 2024; wheat exhibited strong interannual variability, and maize occupied a relatively small area. Crop-specific ETc differed markedly among crop types, following the order orchard > cotton > maize > wheat, with orchards maintaining the highest water requirement across all growth stages. Total agricultural water demand, estimated by integrating crop-specific ETc with remotely sensed planting areas, increased from approximately 260 million m³ to over 500 million m³ after 2010, mainly due to orchard expansion and cotton cultivation. Path analysis indicated that interannual ETc variability exhibited a stronger statistical association with wind speed than with other meteorological variables. These results provide a quantitative basis for cropping-structure optimization and water-saving irrigation management under changing climatic conditions.

1. Introduction

Under climate change and increasing agricultural intensification [1], irrigated agriculture in arid and semi-arid regions has become one of the most water-consuming and climate-sensitive land-use systems. In inland river basins of arid Northwest China, where precipitation is scarce and evapotranspiration dominates the water balance [2,3], irrigation districts represent critical management units in which crop composition, cultivated area, and climatic variability jointly determine agricultural water demand [4,5]. Quantifying crop water requirements (ETc) and clarifying their climatic driving mechanisms are therefore critical for irrigation optimization and sustainable water-resource management [6].

Conventional crop monitoring and water-demand assessments rely mainly on field surveys, agricultural statistics [7], and in situ observations such as lysimeters and eddy-covariance measurements [8]. Although accurate at local scales, these approaches are difficult to upscale to heterogeneous irrigation districts where cropping patterns vary spatially and temporally. Spatially explicit information on crop distribution has therefore become a prerequisite for regional ETc estimation, motivating the increasing use of remote sensing in agricultural water studies [9]. However, single-sensor observations are constrained by inherent trade-offs between spatial and temporal resolution [10,11]. MODIS provides high temporal frequency but coarse spatial detail [12,13,14], whereas Landsat and Sentinel-2 offer finer spatial resolution but are limited by revisit cycles and cloud contamination [15,16]. These constraints directly affect the reliability of crop-area mapping, and consequently the accuracy of irrigation water-demand assessment.

To overcome such limitations, multi-sensor fusion and harmonized products have been developed to support consistent crop classification and phenological analysis at the regional scale [17,18]. Crop classification accuracy, however, directly influences ETc estimation by determining the area-weighted contribution of different crops to total irrigation demand, and thus depends on feature construction, algorithm selection, and data-processing efficiency [19]. Spectral bands, texture metrics, and vegetation indices such as NDVI and EVI are commonly used to characterize crop growth dynamics [20,21,22,23]. Among machine-learning methods, Random Forest has been widely applied due to its robustness and efficiency [24,25,26], while cloud-computing platforms such as Google Earth Engine enable efficient processing of long-term, multi-source satellite data for irrigation districts [27,28,29,30,31].

Crop water requirement (ETc) is a key indicator linking crop distribution to agricultural water use [32,33,34]. Existing ETc estimation approaches include empirical methods, physically based energy-balance models, and process-oriented crop models such as the FAO-56 Penman–Monteith framework and CROPWAT [35,36]. Despite these advances, two gaps remain at the irrigation-district scale. First, crop classification and ETc estimation are often conducted independently, limiting understanding of how spatiotemporal changes in cropping structure translate into variations in irrigation water demand. Second, although the influence of meteorological conditions on ETc is well recognized, quantitative attribution of interrelated climatic variables remains limited, particularly in terms of separating their contributions under a predefined causal structure [37]. Simple correlation analysis cannot address this problem.

Path analysis is a multivariate statistical method that partitions associations among variables into direct and indirect effects based on an a priori causal model [38]. When applied with clearly defined assumptions, it allows for quantitative evaluation of the relative contributions of correlated meteorological variables to ETc, without implying the discovery of causal mechanisms. This makes it suitable for irrigation-district–scale analyses where climatic variables are strongly interdependent.

Accordingly, this study was conducted in the Tailan River Irrigation District, Xinjiang, a typical inland arid irrigation system characterized by scarce precipitation and high evaporative demand. Based on the identified gaps, we propose the following testable hypotheses: (i) At the irrigation-district scale, spatiotemporal changes in crop planting structure affect total ETc primarily through area-weighted differences in crop-specific water requirements under given climatic conditions; (ii) The influences of meteorological variables (e.g., temperature, radiation, humidity, and wind speed) on ETc can be quantitatively partitioned into direct and indirect effects within a predefined causal framework using path analysis. Specifically, long—term Landsat and MODIS imagery were processed on the Google Earth Engine platform to derive crop—distribution dynamics. ETc was estimated using the CROPWAT model, and the predefined relationships among meteorological variables and ETc were evaluated using path analysis. The objectives are to characterize spatiotemporal crop—pattern dynamics, quantify crop-specific ETc variations, and assess the relative contributions of climatic drivers within an explicit analytical framework. This approach supports irrigation optimization and climate-adaptive water management in arid irrigation districts.

2. Materials and Methods

2.1. Overview of the Study Area

The Tailan River Irrigation District is located in Wensu County, Aksu Prefecture, Xinjiang, China (80°21′–81°10′ E, 40°42′–42°15′ N, Figure 1) [39]. The irrigation district comprises two townships and three towns (Yixilaimuqi Township, Guleawati Township, Jiamu Town, Gongqingtuan Town, and Kezile Town), with a total area of 5145.93 km². The region has a temperate continental arid climate, with a mean annual temperature of 11.5 °C, mean annual precipitation of 73–110 mm, and potential evapotranspiration exceeding 1600 mm. Elevation ranges from 1050 m to 1300 m. Irrigation water is mainly supplied by the Tailan River originating from the southern slope of the Tianshan Mountains, supplemented by canal networks and reservoirs. The dominant crops include wheat, maize, cotton, and orchard crops, forming a mixed agricultural system typical of oasis irrigation regions in northwest China.

2.2. Data Sources and Preprocessing

2.2.1. Remote Sensing Image Data

This study used Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI/TIRS (https://earthexplorer.usgs.gov/) surface reflectance products from their respective operational periods between 2000 and 2024, complemented by MODIS (https://ladsweb.modaps.eosdis.nasa.gov/) MOD13Q1 NDVI (250 m, 16-day composite) to enhance temporal continuity. Specifically, Landsat 5 TM data were used for the period 2000–2011, Landsat 7 ETM+ for 2000–2024 (excluding scan-line corrector gaps after 2003 where applicable), and Landsat 8 OLI/TIRS for 2013–2024. In years with overlapping observations from multiple sensors, Landsat 8 OLI/TIRS was prioritized, followed by Landsat 7 ETM+ and Landsat 5 TM, to ensure radiometric consistency.

Landsat imagery was used for annual crop mapping due to its 30 m spatial detail, while MOD13Q1 was used only as a temporal reference to constrain and stabilize NDVI temporal profiles during periods affected by cloud contamination and Landsat 7 SLC failure. MODIS data were not directly used as classification inputs, nor were 250 m pixels resampled to 30 m for crop labeling. Their role was limited to guiding temporal smoothing and gap-filling of Landsat-derived NDVI curves, thereby minimizing mixed-pixel contamination in the classification stage. Although Landsat surface reflectance products are commonly referred to as 30 m data, it should be noted that thermal bands have coarser native resolutions (60–120 m depending on the sensor) and were not used in NDVI or crop classification features. All spectral features used for classification were derived from bands with 30 m spatial resolution. The satellite sensors and spectral band characteristics used in this study are summarized in

Table 1. Remote sensing image bands and their resolutions.

Satellite & Sensor	Band Number	Band Name	Wavelength Range (μm)	Spatial Resolution (m)
Landsat 5 (TM)	3	Red	0.63–0.69	30
	4	NIR	0.76–0.90	30
	5	SWIR1	1.55–1.75	30
	7	SWIR2	2.08–2.35	30
	6	Thermal	10.40–12.50	120
Landsat 7 (ETM+)	3	Red	0.63–0.69	30
	4	NIR	0.77–0.90	30
	5	SWIR1	1.55–1.75	30
	7	SWIR2	2.09–2.35	30
	6	Thermal	10.40–12.50	60
Landsat 8 (OLI/TIRS)	4	Red	0.64–0.67	30
	5	NIR	0.85–0.88	30
	6	SWIR1	1.57–1.65	30
	7	SWIR2	2.11–2.29	30
	10	Thermal	10.60–11.19	100
Terra/Aqua (MODIS)	1	Visible Red	0.620–0.670	250
	2	Near Infrared	0.841–0.876	250

Note: NDVI = (NIR − Red)/(NIR + Red). SWIR/Thermal bands were documented to ensure spectral completeness.

.

All image processing was implemented in Google Earth Engine (GEE). Preprocessing included atmospheric correction (LEDAPS/LaSRC for Landsat SR), cloud and shadow masking (CFMask/QA-based masking), BRDF normalization, mosaicking, and annual compositing.

2.2.2. Meteorological Data

Meteorological variables used in this study include temperature (T), precipitation (Pre), relative humidity (RH), wind speed (V), and sunshine hours (S). All meteorological variables were obtained from the National Earth System Science Data Center (NESSDC), China (https://www.geodata.cn/) [40]. These datasets are provided as spatially continuous gridded products generated through the integration of station observations and interpolation algorithms, ensuring spatial completeness across the study region. For this study, gridded meteorological data covering the period 2000–2024 were extracted and spatially averaged over the Tailan River Irrigation District to derive area-representative climatic inputs for crop water requirement (ETc) estimation using the CROPWAT model.

2.2.3. Sample Data

Crop sample points were interpreted from sub-meter historical imagery in Google Earth and supplemented by field surveys conducted in September 2024 and May 2025. A total of 1,350 samples were collected, including wheat (271), maize (103), cotton (363), orchard (388), and other land-cover types (225). Samples were randomly split into training (70%) and validation (30%) subsets. To reduce spatial autocorrelation and spectral redundancy, sample distribution was examined to minimize spatial clustering. Representative field photographs of the major crop types used for sample interpretation are shown in Figure 2.

Although field surveys were conducted in 2024–2025, the collected samples were used to establish stable spectral–phenological signatures of major crop types. These signatures were assumed to be temporally transferable and were applied consistently across annual classifications from 2000 to 2024, rather than representing year-specific crop distributions.

2.3. Model and Methods

2.3.1. NDVI Time-Series Reconstruction

NDVI is a widely used indicator for characterizing vegetation growth conditions and for extracting vegetation information from remote sensing data [41]. NDVI is calculated as:

$$NDVI={\displaystyle \frac{{\rho }_{NIR}-{\rho }_{Red}}{{\rho }_{NIR}+{\rho }_{Red}}}$$

(1)

where ${\rho }_{NIR}$ denotes the reflectance in the near-infrared band; ${\rho }_{Red}$ denotes the reflectance in the red band.

To reconstruct smooth NDVI temporal profiles and suppress residual noise after standard preprocessing, the time series was reconstructed using the Savitzky–Golay (S–G) filter [42,43]. The reconstructed NDVI temporal profiles have an effective temporal resolution of approximately 16–32 days, depending on Landsat observation availability, and cover the full crop growing season from April to October for each year. These profiles were generated independently for each year from 2000 to 2024. The S–G reconstruction is expressed as

$$Y_j^{\ast }={\displaystyle \frac{{\displaystyle \sum _{i=-m}^{i=m}{C}_i\times Y_{j+1}}}{N}}$$

(2)

where $Y_{j+1}$ and $Y_j^{\ast }$ represent the original and reconstructed data; C_i denotes the coefficient derived from S–G polynomial fitting, indicating the weight assigned to the i-th NDVI value in the filter; and m defines the half-width of the smoothing window, m = 3 and 2m + 1 = 7.

2.3.2. Classification and Validation

A Random Forest (RF) classifier with 300 trees [44] was trained using multi-temporal NDVI features and spectral reflectance bands (Red, NIR, SWIR1, and SWIR2). Multi-temporal NDVI metrics included mean, maximum, minimum, and key phenological-phase averages extracted from the reconstructed NDVI time series. Spectral features were derived from annual growing-season composites.

The classification was conducted annually from 2000 to 2024 within Google Earth Engine (GEE) to generate annual crop-type maps used to identify the crop planting structure. Accuracy was evaluated using producer accuracy (PA), user accuracy (UA), overall accuracy (OA), and the Kappa coefficient (Table 2). Confusion matrices were generated for each year, and temporal consistency was validated using statistical yearbook data (Figure 3).

Table 2. Formulas for accuracy assessment.

Type	Formula
Producer’s Accuracy (PA)	$PA={\displaystyle \frac{X_{ii}}{X_j}}\times 100\%$	(3)
User’s Accuracy (UA)	$UA={\displaystyle \frac{X_{ii}}{X_i}}\times 100\%$	(4)
Overall Accuracy (OA)	$OA={\displaystyle \sum _{i=1}^k\frac{X_{ii}}{N}}\times 100\%$	(5)
Kappa Coefficient (K)	$K={\displaystyle \frac{N{\displaystyle {\sum }_{i=1}^k{X}_{ii}}-{\displaystyle {\sum }_{i=1}^k{X}_i{X}_j}}{N^2-{\displaystyle {\sum }_{i=1}^k{X}_i{X}_j}}}$	(6)

Table 2. Formulas for accuracy assessment.

Type	Formula
Producer’s Accuracy (PA)	$PA={\displaystyle \frac{X_{ii}}{X_j}}\times 100\%$	(3)
User’s Accuracy (UA)	$UA={\displaystyle \frac{X_{ii}}{X_i}}\times 100\%$	(4)
Overall Accuracy (OA)	$OA={\displaystyle \sum _{i=1}^k\frac{X_{ii}}{N}}\times 100\%$	(5)
Kappa Coefficient (K)	$K={\displaystyle \frac{N{\displaystyle {\sum }_{i=1}^k{X}_{ii}}-{\displaystyle {\sum }_{i=1}^k{X}_i{X}_j}}{N^2-{\displaystyle {\sum }_{i=1}^k{X}_i{X}_j}}}$	(6)

2.3.3. Trend Analysis of Meteorological Elements

To analyze long-term trends in meteorological variables, the Mann–Kendall (MK) non-parametric trend test combined with Sen’s slope estimator (Sen+MK method) was applied [45,46]. This method does not require the data to follow a specific distribution and is robust to outliers, making it suitable for climatic time-series analysis.

$$Z=\left\{\begin{array}{cc}{\displaystyle \frac{S-1}{\sqrt{Var\left(S\right)}}},& S>0\\ 0,& S=0\\ {\displaystyle \frac{S+1}{\sqrt{Var\left(S\right)}}},& S<0\end{array}\right.$$

(7)

$$S={\displaystyle \sum _{i=1}^{n-1}{\displaystyle \sum _{j=i+1}^n\mathrm{sgn}(x_j-x_i}})$$

(8)

$$\mathrm{s}\mathrm{g}\mathrm{n}(x_j-x_i)=\left\{\begin{array}{cc}+1,& \mathrm{i}\mathrm{f} x_j-x_i>0,\\ 0,& \mathrm{i}\mathrm{f} x_j-x_i=0,\\ -1,& \mathrm{i}\mathrm{f} x_j-x_i<0.\end{array}\right.$$

(9)

$$Var(S)={\displaystyle \frac{n(n-1)(2n+5)-{\displaystyle {\sum }_{p=1}^g{t}_p(t_p-1)(2t_p+5)}}{18}}$$

(10)

where Z is the standardized test statistic, S is the Mann–Kendall statistic, x_i and x_j represent meteorological observations in years i and j, and β denotes Sen’s slope. A positive β indicates an increasing trend, whereas a negative β indicates a decreasing trend.

2.3.4. Estimation of Crop Water Requirement (ET_c)

Crop water requirement (ETc) was estimated using the CROPWAT 8.0 model based on the FAO-56 Penman–Monteith framework [47]:

$$ETc=Kc\times E{T}_0$$

(11)

$$E{T}_0={\displaystyle \frac{0.408\mathrm{\Delta }(R\mathrm{s}-G)+\gamma \frac{900U_2(e_s-e_a)}{T+273}}{\mathrm{\Delta }+\gamma (1+0.34U_2)}}$$

(12)

where ET₀ is the reference evapotranspiration (mm d⁻¹); R_n is the net radiation at the crop surface (MJ m⁻² d⁻¹); G is the soil heat flux (MJ m⁻² d⁻¹); T is the mean daily air temperature at 2 m height (°C); U₂ is the wind speed at 2 m height (m s⁻¹); e_s and e_a are the saturation and actual vapor pressures (kPa); Δ is the slope of the saturation vapor pressure curve (kPa °C⁻¹); and γ is the psychrometric constant (kPa °C⁻¹).

Daily reference evapotranspiration (ET₀) was calculated using station-based meteorological inputs. Crop coefficients (Kc) were adopted from FAO-56 and adjusted using regional agronomic studies in Xinjiang. Uniform soil parameters representative of the irrigation district were applied due to the lack of spatially distributed soil data. Crop growth-stage lengths and Kc values are listed in

Table 3. Crop coefficients (Kc).

Crop Type	Crop Coefficients			References
	Initial	Mid-Season	Maturity
Wheat	0.30	1.15	0.40	[37,47]
Maize	0.30	1.20	0.35	[37,47]
Cotton	0.35	1.20	0.60	[37,47]
Orchard	0.86	1.45	0.81	[48,49]

.

2.3.5. Path Analysis

The Path Analysis Method [50] was used to quantify the direct and indirect effects of meteorological factors on annual ETc. Path analysis was conducted in MATLAB (R2023b). All variables (temperature, precipitation, relative humidity, wind speed, sunshine duration, and annual ETc) were standardized prior to analysis. Standardized path coefficients were obtained using multiple linear regression, and direct/indirect effects were calculated from the regression coefficients and the correlation matrix. Model fit was evaluated using R² and χ²/df.

$$y=b_0+b_1{x}_1+b_2{x}_2+\cdots +b_n{x}_n$$

(13)

Equation (9) can be mathematically changed to the following positive equation of moments by substituting the actual observations into it and solving the system of equations using least squares.

$$\begin{array}{c}\left[\begin{array}{cccc}1& {\gamma }_{x_1{x}_2}& \dots & {\gamma }_{x_1{x}_n}\\ {\gamma }_{x_2{x}_1}& 1& \dots & {\gamma }_{x_2{x}_n}\\ ⋮& ⋮& \ddots & ⋮\\ {\gamma }_{x_n{x}_1}& {\gamma }_{x_n{x}_2}& \dots & 1\end{array}\right]\left[\begin{array}{c}{P}_{y_{x_1}}\\ P_{y_{x_2}}\\ ⋮\\ P_{y_{x_n}}\end{array}\right]=\left[\begin{array}{c}{\gamma }_{x_{1}y}\\ {\gamma }_{x_{2}y}\\ ⋮\\ {\gamma }_{x_{n}y}\end{array}\right]\end{array}$$

(14)

where γ_xixj (i, j ≤ n) represents the simple correlation coefficient between the n independent variables; γ_xiy (I ≤ n) denotes the simple correlation coefficient between the n independent variables and the dependent variable; P_xny is the direct path coefficient; γ_xixj P_xny represents the indirect path coefficient.

3. Results

3.1. Spatiotemporal Dynamics of Crop Planting Structure in the Tailan River Irrigation District

3.1.1. Results of Crop Planting Structure Identification and Accuracy Assessment

As shown in

Table 4. Classification confusion matrix and accuracy metrics for major crops.

Crop Type	Wheat	Maize	Cotton	Orchard	Other	PA	UA	OA	Kappa Coefficient
Wheat	245	6	8	5	7	0.904	0.906	0.902	0.876
Maize	4	92	2	1	4	0.893	0.894
Cotton	7	5	338	9	4	0.931	0.928
Orchard	3	2	11	362	10	0.933	0.867
Other	12	4	7	11	234	0.873	0.913

, the confusion matrix and accuracy metrics for the 2024 training year indicate strong classification performance. The overall accuracy (OA) was 0.902, with a Kappa coefficient of 0.876, suggesting high agreement between classification results and reference samples. Producer’s accuracy (PA) ranged from 0.873 to 0.933, and user’s accuracy (UA) ranged from 0.867 to 0.928 across the five classes. Notably, UA values for orchards (0.867) and maize (0.894) were lower than those for wheat and cotton, indicating comparatively higher commission uncertainty for these categories.

Regression-based validation indicates strong agreement between remotely extracted crop areas and statistical yearbook records across all major crops (Figure 4). Wheat exhibited high consistency (R² = 0.896, RMSE = 11.420 km²) followed by maize (R² = 0.857, RMSE = 5.322 km²). Cotton showed comparatively larger dispersion (R² = 0.881, RMSE = 25.546 km²) while orchards maintained a high level of agreement (R² = 0.885, RMSE = 23.727 km²). Mean absolute percentage errors (MAPE) for all crops were below 10% (Wheat: 9.63%, Maize: 9.27%, Cotton: 6.67%, Orchard: 6.56%), indicating stable area estimation performance at the irrigation-district scale.

3.1.2. Spatiotemporal Dynamics of Cropping Patterns

From 2000 to 2024, the cropping structure of the Tailan River Irrigation District exhibited clear long-term changes (Figure 5). Cotton remained the largest crop, ranging from 228.6 to 426.0 km² and consistently accounting for more than 40% of the total planting area. Orchards expanded from 206.5 km² in 2000 to 393.2 km² in 2024. Wheat showed pronounced interannual variability, fluctuating between 22.9 and 160.5 km², with higher values observed in 2005 and 2016. Maize occupied a relatively small area, generally remaining below 100 km², with a maximum of 98.0 km² in 2005.

Spatially, crop distributions showed clear differentiation (Figure 6). Wheat and maize were mainly concentrated in the mid-to-upper reaches of the northern irrigation district. Cotton maintained a relatively stable core distribution in the central region and exhibited spatial expansion toward downstream areas over time. Orchards, initially concentrated in central areas with partial overlap with cotton, expanded outward and became the second-largest crop by 2024.

3.2. Characteristics of Climate Change in the Irrigation District

Meteorological variables are key inputs for estimating crop water requirements. Using observations from the Wensu County meteorological station, this study summarized interannual changes in temperature, precipitation, relative humidity, wind speed, and sunshine duration during 2000–2024 (Figure 7). The Mann–Kendall test combined with Sen’s slope estimator indicated a significant increasing trend in annual mean temperature (Sen = 0.0296 °C·yr⁻¹, p = 0.017, Z = 2.38) and wind speed (Sen = 0.0090 m·s⁻¹·yr⁻¹, p = 0.031, Z = 2.15), and a significant decreasing trend in relative humidity (Sen = −0.3135%·yr⁻¹, p = 0.001, Z = −3.25). Precipitation showed a non-significant decreasing tendency (Sen = −0.1875 mm·yr⁻¹, p = 0.709), and sunshine duration showed no significant trend (Sen = 0.0006 h·yr⁻¹, p = 0.907).

3.3. Dynamics of Crop Water Requirements

3.3.1. Variation Characteristics of the Average Crop Water Requirement per Unit Area over Multiple Years

Unit-area crop water requirements (ETc) from 2000 to 2024 were estimated using the CROPWAT model based on crop coefficients (Kc) and reference evapotranspiration (ET₀) (Figure 8a). Wheat exhibited unit-area water requirements ranging from 400.20 to 486.12 mm, while maize varied between 426.18 and 517.98 mm. Cotton showed a broader range, with values spanning 443.75 to 537.22 mm. Orchards consistently exhibited the highest water requirements, fluctuating between 584.98 and 708.57 mm. Across all crops and years, the mean unit-area water requirement was 515.63 mm, with orchards (650.34 mm) markedly exceeding those of cotton (493.51 mm), maize (474.09 mm), and wheat (444.59 mm).

Model estimates were compared with officially reported irrigation water-use statistics for 2019–2023 obtained from the Tailan River Irrigation District Water Use Statistical Survey System (Figure 8b). These data represent aggregated annual irrigation water-use records rather than plot-scale direct measurements. Validation results for 2019–2023 indicate a high level of consistency between CROPWAT estimates and field observations. The coefficient of determination (R²) reached 0.873, accompanied by low absolute errors (MAE = 2.85 mm; RMSE = 3.70 mm) and a MAPE of only 0.55%. Given the limited temporal coverage of available validation data (five years), this comparison is intended to provide a preliminary consistency check rather than a comprehensive long-term validation of model performance.

3.3.2. Multi-Year Average Crop Water Requirements at Different Growth Stages

Crop water requirements at different growth stages were calculated according to the FAO—56 approach (ETc = Kc × ET₀), using prescribed stage—specific Kc values and growth durations (Figure 9). Distinct crop-specific differences were observed across growth stages. Wheat reached its maximum water requirement during the mid-season stage (193.08 mm), whereas maize and cotton peaked during the rapid growth stage, with values of 171.75 mm and 209.14 mm, respectively. Orchards consistently exhibited the highest water demand across all stages, particularly during rapid growth (305.59 mm) and maturity (120.88 mm). During the maturity stage, water requirements for wheat, maize, and cotton converged to similar levels (64.03–72.57 mm), while orchards maintained substantially higher water demand.

3.3.3. Interannual Variation in Total Crop Water Requirements

Total crop water requirements were derived by combining unit-area ETc estimates with crop planting areas (Figure 10). From 2000 to 2024, annual total water demand showed marked interannual variability, ranging from 264.01 million m³ to 528.49 million m³. Orchards and cotton consistently represented the dominant contributors to total demand, with orchard water requirements varying between 126.82 and 259.91 million m³ and cotton ranging from 105.45 to 224.45 million m³. In contrast, wheat and maize contributed relatively smaller shares, with annual water demands of 9.44–76.33 million m³ and 5.02–44.80 million m³, respectively. Variations in total water demand largely reflect changes in cropping structure and modeled ETc rather than independent hydrological observations.

3.4. Analysis of Meteorological Driving Paths

3.4.1. Correlation Between Crop Water Requirement and Meteorological Factors

Correlation analysis was conducted to describe statistical associations between ETc and meteorological variables used in the ET₀ calculation (Figure 11). It should be noted that these correlations partly reflect the mathematical structure of the FAO-56 Penman–Monteith equation rather than independent empirical relationships. Crop water requirement (ETc) was positively associated with temperature (r = 0.37, p = 0.067) and wind speed (r = 0.78, p < 0.001), and negatively associated with relative humidity (r = −0.57, p = 0.003). Correlations with precipitation (r = −0.14, p = 0.508) and sunshine duration (r = 0.41, p = 0.039) were weaker. Inter-correlations among meteorological variables were also observed, including temperature versus relative humidity (r = −0.42, p = 0.036) and temperature versus wind speed (r = 0.44, p = 0.029).

3.4.2. Path Analysis

To further characterize the relative contributions of meteorological variables to modeled crop water requirement (ETc), a path analysis was applied to the major crops in the Tailan River Irrigation District (Figure 12 and Figure 13).

Wind speed (V) exhibited the largest positive direct path coefficients for ETc across all crops, with values of 0.565 for wheat, 0.574 for maize, 0.639 for cotton, and 0.662 for orchards. The corresponding decision coefficients (0.509–0.653) indicate that interannual variability in modeled ETc is statistically most sensitive to wind speed within the applied framework. Sunshine duration (S) showed a moderate positive direct association with ETc, declining from wheat (0.370) and maize (0.365) to cotton (0.311) and orchards (0.295). Relative humidity (RH) consistently exerted a negative direct effect on ETc (−0.347 to −0.378), while temperature (T) contributed only weak direct effects (−0.066 to −0.080).

Indirect pathways revealed consistent interaction patterns among meteorological variables across crops. Temperature showed positive indirect contributions to ETc, mainly mediated through RH (0.146–0.159) and sunshine duration (0.019–0.237), partially offsetting its weak negative direct effect. Relative humidity displayed compensatory indirect pathways—positive via wind speed (0.135–0.147) and negative via sunshine duration (−0.220 to −0.258)—resulting in a near-neutral net effect (|0.011–0.014|). Precipitation was characterized by a predominantly negative indirect influence mediated through sunshine duration (−0.123 to −0.144), exceeding its minimal direct contribution. Indirect effects of wind speed and sunshine duration themselves were comparatively small, indicating that their influence on ETc is primarily expressed through direct pathways.

Overall, the path analysis highlights wind speed as the meteorological variable with the strongest direct statistical association with modeled ETc across crops, followed by relative humidity and sunshine duration.

4. Discussion

4.1. The Essential Contributions and Scientific Issues of the Integration Method

The “Classification-Simulation-Analysis” framework developed in this study transcends the conventional integration of “crop mapping,” “water demand estimation,” and “statistical attribution” as mere sequential steps. Its core scientific challenge lies in effectively separating and quantifying the distinct contributions of planting structure changes and climate change to long-term agricultural water demand trends at the irrigation district scale. Previous studies either focused on single components (e.g., crop mapping alone [7,10] or evapotranspiration simulation [6,8]) or, despite multi-step analyses, failed to systematically examine their interactions and relative importance within a unified framework [13,20]. This framework ensures consistency in tracking water demand effects from land use changes through the use of a unified spatiotemporal basis (high-resolution annual crop maps based on Landsat-MODIS fusion data [51,52]). This “consistency” is critical: it means that interannual variations in ETc can be attributed not only to climate fluctuations but also clearly separated into “structural” water demand changes caused by crop type transitions (e.g., orchard expansion) through precise annual crop area weighting. Our analysis demonstrates that in the Taran River Irrigation District, the intensification of planting structures over the past two decades (particularly the expansion of water-intensive crops) has driven irrigation water demand increases comparable to or even more significantly than climate change. This finding surpasses traditional water demand modeling approaches based solely on climate data or static crop distribution.

4.2. Regional Specificity of Crop Water Demand Driven by Climate

This study conducted a climate-driven attribution analysis of crop evapotranspiration (ETc) to not only validate the inherent logic of the FAO-56 Penman-Monteith formula [53] but also to reveal the relative importance, pathways, and interactions of various climatic factors [54] affecting ETc in the Taran River Irrigation District under specific arid conditions. The findings indicate that wind speed is the primary direct climatic driver of interannual ETc variability [37,39,55]. This region-specific pattern has a physical basis: frequent strong winds in the study area significantly reduce aerodynamic resistance, thereby continuously promoting canopy evapotranspiration, particularly during the mid-growth season when energy is abundant. In contrast, precipitation, due to its scarcity and variability, exerts a weak direct limiting effect on ETc. Relative humidity, on the other hand, primarily exerts an indirect negative influence by modulating vapor pressure difference (VPD) [56,57,58]. Pathway analysis further demonstrates that factors such as temperature and sunshine duration interact through multiple direct and indirect pathways (e.g., by affecting VPD and net radiation), indicating that ETc changes result from complex feedback mechanisms among multiple correlated climatic factors. Therefore, this study emphasizes that key conclusions (e.g., the dominant role of wind speed) should be interpreted as statistical attribution results based on specific methodologies and data, revealing strong correlations rather than absolute causality. This understanding is crucial for prudent interpretation of results and subsequent management applications.

4.3. Irrigation Management Enlightenment Based on Research Findings

Management insights must be rigorously derived from and closely aligned with the aforementioned research findings, with clearly defined boundaries for their application. Building on the core empirical conclusion that “wind speed is a key climatic driver of ETc variability,” irrigation management can more effectively address wind impacts [59]. For instance, Irrigation management strategies must incorporate spatiotemporal variations in wind-induced evapotranspiration (ETc), particularly in regions or seasons characterized by elevated wind speeds [60]. Dynamic adjustment of irrigation schedules during such periods enables more precise water allocation, reducing both under- and over-irrigation risks. From a long-term adaptation perspective, the empirical link between crop structural shifts and rising agricultural water demand highlights the need to reassess land-use strategies. Specifically, the expansion of high water-consuming perennial crops (e.g., orchards) should be critically evaluated for sustainability under increasingly constrained water resources. Instead, adaptive measures—including the adoption of water-saving crop varieties, rotation systems with lower water footprints, and precision irrigation technologies—may serve as viable demand-side mitigation pathways. However, the operationalization of these strategies necessitates a context-specific evaluation framework, integrating local socio-economic dynamics, existing water infrastructure capacity, institutional constraints, and regional policy objectives. Without such integration, the effectiveness and equity of water management reforms may be undermined.

4.4. Limitations and Robustness of Results

We fully recognize the following uncertainties and demonstrate why they do not affect the core conclusions of this study regarding long-term trends and relative importance: (1) Crop classification uncertainty: Classification errors for orchards and corn primarily impact the absolute accuracy of their area measurements, but have limited influence on the temporal trends of crop area changes due to relatively stable errors over time. Therefore, the long-term upward trend in total water demand weighted by area remains robust. (2) Model parameter representativeness: Using representative soil parameters and standard crop coefficients for irrigation districts may lead to deviations in individual pixel ETc estimates. However, since this study focuses on total volume and trends at the irrigation district scale, spatial aggregation effects largely smooth out differences in local parameter sensitivity, making regional total water demand estimation and comparison reasonable. (3) Meteorological data representativeness: Single-station meteorological data has spatial limitations and may not fully capture microclimatic variations within irrigation districts. However, this mainly affects the spatial detail characterization of ETc, while having minimal impact on water demand time series integrated at the full irrigation district level and their statistical relationship patterns with climate factors, as dominant climate factors (e.g., wind speed) exhibit good spatial consistency within the region. Future research should focus on reducing these uncertainties, such as: improving crop classification and validation through multi-source remote sensing products and field observations; enhancing model accuracy by integrating spatially explicit soil attributes and irrigation management information; and utilizing regional climate reanalysis data or dense meteorological station networks to improve the representativeness of climate forcing fields. In addition, more complex attribution methods such as structural equation modeling can be used to further analyze the interaction between climate factors and changes in planting structure.

5. Conclusions

This study proposes and applies an “integrated framework of classification-simulation-analysis” to quantify the long-term dynamic effects of crop structure changes on irrigation water demand in the Taran River Irrigation District from 2000 to 2024, while identifying key meteorological drivers of crop water demand variations. Results demonstrate that intensified crop structure—particularly the sustained dominance of cotton cultivation alongside significant orchard expansion—has increased agricultural water resource pressure, which is further amplified by climate change. Attribution analysis critically reveals that wind speed serves as the primary climatic driver of water demand, while relative humidity exerts a counter-regulatory effect; temperature and precipitation mainly influence water demand through indirect pathways. These findings systematically elucidate the combined effects of crop structure changes and climate change on regional water demand, highlighting the methodological value of the proposed integrated framework in addressing water cycle challenges in complex human-nature coupled systems.

It should be noted that this study involves several uncertainties, including classification errors in orchard and maize categorization during crop mapping, representative soil parameters used in CROPWAT for irrigation districts, and the representativeness of station meteorological data. These factors may influence the absolute values of water demand estimates. However, the regional synthesis and area-weighted evaluation method adopted in this study provides a robust basis for interpreting long-term trends at the irrigation district scale. Consequently, the primary conclusion—planting structure changes and wind speed-dominated climatic impacts—has a reliable inferential foundation.

From a management perspective, wind speed’s dominant influence provides clear guidance for adaptive water conservation. For instance, during peak water demand periods, priority should be given to reinforcing windbreak forest systems, optimizing irrigation schedules, and enhancing irrigation efficiency. Under water resource constraints, prudent adjustments to planting structures—such as moderating orchard expansion, promoting water-saving crop varieties, or adopting agronomic practices—should be considered long-term. Future research should utilize irrigation records and independent evapotranspiration data to strengthen cross-validation, incorporate spatially explicit soil and management information, and explore methods like nonlinear or structural equation models to more accurately characterize the complex crop-climate-water feedback relationships. This will improve the framework’s transferability to other arid and semi-arid irrigation regions.