High-Resolution Daily Evapotranspiration Estimation in Arid Agricultural Regions Based on Remote Sensing via an Improved PT-JPL and CUWFM Fusion Framework

Highlights

What are the main findings?

• Replacing relative humidity with LSWI significantly improved PT-JPL performance in arid and semi-arid regions.
• The proposed CUWFM generates daily 30 m ET data with high pixel-level accuracy and better spatial detail than existing fusion methods.

What are the implications of the main findings?

• Enables field-scale irrigation scheduling and precision water management in arid agricultural regions.
• Enhances the physical consistency and general applicability of ET estimation and fusion frameworks.

Abstract

Evapotranspiration (ET) plays a crucial role in the terrestrial water cycle, especially in arid and semi-arid agricultural regions where precise water management is essential. However, the limited spatial resolution and temporal frequency of existing ET products hinder their application in fine-scale agricultural monitoring. In this study, we first improved the Priestley–Taylor Jet Propulsion Laboratory (PT-JPL) model by replacing the relative humidity-based soil moisture constraint with the land surface water index (LSWI), aiming to enhance model performance in water-limited environments. Second, we developed a Crop Unmixing and Weight Fusion Model for ET (CUWFM) to generate daily ET products at a 30 m spatial resolution by integrating high-resolution but infrequent PT-JPL-ET data with coarse-resolution but frequent PML-V2-ET data. The CUWFM employs a hybrid approach combining sub-pixel crop fraction decomposition with similarity-weighted regression, allowing for more accurate ET estimation over heterogeneous agricultural landscapes. The proposed methods were evaluated in the Changji region of Xinjiang, China, using field-measured ET data from two-flux-tower sites. The results show that the improved PT-JPL model increased ET estimation accuracy compared with the original version, with higher R² and Nash–Sutcliffe efficiency (NSE), and lower root mean square error (RMSE). The CUWFM outperformed benchmark spatiotemporal fusion methods, including STARFM, ESTARFM, and Fit-FC, in both pixel- and field-scale assessments, achieving the highest overall performance scores based on the All-round Performance Assessment (APA) framework. This study demonstrates the potential of integrating vegetation indices and crop-specific spatial decomposition into ET modeling, providing a feasible pathway for producing high spatiotemporal resolution ET datasets to support precision agriculture in arid and semi-arid regions.

1. Introduction

Evapotranspiration (ET) is a comprehensive process encompassing surface evaporation, vegetation transpiration, and evaporation of intercepted water from plant canopies [1,2,3]. It serves as a key variable in the soil–crop–atmosphere continuum and represents the primary component of energy and water vapor exchange. Accurate monitoring of ET is of paramount importance for water resource management, agricultural irrigation decision-making, and ecosystem assessment [4,5].

Satellite remote sensing technology, with its unique advantages of extensive spatial coverage and continuous temporal observation, has been widely applied for ET monitoring across spatial scales, from regional to global [6,7]. Remote sensing observations can simultaneously acquire multiple surface parameters, including vegetation indices, land surface temperature, and soil moisture, comprehensively reflecting vegetation conditions and hydrothermal characteristics of the underlying surface. This capability has led to continuous improvements in the accuracy and applicability of remote sensing-based ET estimation methods [8]. Currently, remote sensing ET estimation methods primarily include the feature space approach [6,9,10,11,12], energy balance methods [13,14,15,16,17], and approaches based on the Penman–Monteith and Priestley–Taylor equations [18,19,20,21,22]. Among these, the Priestley–Taylor Jet Propulsion Laboratory model (PT-JPL), as a representative two-source physical ET model [23,24,25], has gained widespread application due to its satisfactory simulation performance and relatively low dependence on ground-based data [26,27,28,29,30,31].

However, existing remote sensing ET estimation faces a technical bottleneck in balancing spatiotemporal resolution, a contradiction that is particularly pronounced in arid and semi-arid agricultural regions [8]. On the one hand, high spatial resolution satellites (such as Landsat and Sentinel-2) can provide fine spatial information, but their long revisit cycles make it difficult to meet the agricultural management requirements for continuous time series [32,33,34]. On the other hand, high temporal resolution satellites (such as MODIS) can provide daily observations, but their relatively low spatial resolution cannot effectively capture local heterogeneity characteristics at the field scale [35,36,37]. For arid and semi-arid agricultural regions, continuous and fine spatiotemporal distribution information of ET during crop growing seasons constitutes an essential foundation for implementing precision agriculture and efficient water resource management. This technical bottleneck constrains the application of remote sensing ET data in agricultural practices [38,39].

To address these challenges, spatiotemporal data fusion techniques have emerged. Over the past two decades, researchers have proposed various spatiotemporal fusion methods [40,41,42], which can be broadly categorized into three major types: weight function-based [43], unmixing-based [42,44], and learning-based approaches [45,46,47]. The weight function-based Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) reconstructs high-resolution images by comprehensively considering spectral differences and spatial distances between pixels, utilizing temporal change information from low-resolution data [48]. Subsequently developed improved algorithms, such as STAARCH and ESTARFM, have achieved significant advances in fusion accuracy [43,49]. Unmixing-based methods rely on linear spectral mixture theory, achieving effective spatiotemporal information fusion through endmember extraction and spectral unmixing.

In recent years, specialized fusion algorithms for ET data have gradually emerged. The STAEDM model improved the STARFM algorithm by introducing the Temperature Vegetation Dryness Index [50]; the SADFAET algorithm added similarity weights related to land surface temperature based on ESTARFM [51]; and the UWET model utilized precipitation data to optimize reference date selection [52]. These methods have improved ET fusion accuracy to some extent by integrating vegetation indices and land surface temperature information. However, they have not adequately considered crop growth characteristics and the physical mechanisms of ET processes in arid and semi-arid agricultural regions.

Based on the above analysis, this study aims to develop a high spatiotemporal resolution ET estimation method suitable for arid and semi-arid agricultural regions. Specifically: (1) to address the limitations of the original PT-JPL model under arid conditions, the Land Surface Water Index (LSWI) is employed to replace relative humidity in redefining the soil moisture constraint factor; (2) a Crop Unmixing and Weight Fusion Model for ET (CUWFM) is constructed, which performs spatial unmixing operations based on ET influence factors, selects fusion reference dates in combination with crop phenological information, conducts stratified screening of similar pixels using crop distribution data, and employs a hybrid strategy combining crop fraction decomposition with similarity-based weighting to effectively fuse sparse high spatial resolution temporal data with dense high temporal resolution data; (3) systematic validation is conducted based on ground observation data from agricultural sites in arid and semi-arid regions, with comparative analysis against existing mainstream ET products and spatiotemporal fusion algorithms, providing scientific data support for agriculture and water resource management in these regions.

2. Methods

2.1. Remote Sensing ET Estimation Model

This study employs an improved PT-JPL model to estimate ET at 30 m spatial resolution (Figure 1). The PT-JPL model is a physically based two-source ET model that utilizes plant physiological and ecological constraints to limit potential plant transpiration and atmospheric constraints to limit potential soil evaporation, thereby obtaining actual ET rates [53]. This model effectively addresses the difficulty of obtaining ground resistance data in ET simulation while incorporating numerous plant physiological parameters that can effectively reflect changes in ET or latent heat (LE) caused by vegetation variations in the study area [54,55]. ET and LE serve as the link between water and energy cycles. Consequently, daily ET (mm·day⁻¹) can be converted from the latent heat flux (LE, W·m⁻²) using the equation $ET=k\cdot LE$. Here, k is the unit conversion coefficient, approximately equal to 0.0353. This value provides the necessary dimensional check: considering that 1 W·m⁻² is equivalent to 0.0864 MJ·m⁻²·day⁻¹ (based on 86,400 s in a day), dividing this energy density by the latent heat of vaporization of water ($\lambda \approx 2.45 \mathrm{M}\mathrm{J}\cdot \mathrm{k}{\mathrm{g}}^{-1}$ at 20 °C) yields the conversion factor (0.0864/2.45 ≈ 0.0353) [55]. The PT-JPL model divides actual ET into three components: vegetation transpiration (${LE}_c$), soil evaporation ($L{E}_s$), and canopy interception evaporation (${LE}_i$). The algorithm can be described as [29]

$$LE={LE}_{\mathrm{c}}+{LE}_s+{LE}_i$$

(1)

$${LE}_c=\left(1-f_{wet}\right)\times f_g\times f_t\times f_m\times \alpha {\displaystyle \frac{\Delta }{\Delta +\gamma }}\times R_{nc}$$

(2)

$${LE}_s=\left[f_{wet}+f_{sm}\times \left(1-f_{wet}\right)\right]\times \alpha {\displaystyle \frac{\Delta }{\Delta +\gamma }}\times \left(R_{ns}-G\right)$$

(3)

$${LE}_i=f_{wet}\times \alpha {\displaystyle \frac{\Delta }{\Delta +\gamma }}\times R_{nc}$$

(4)

where $f_{wet}$ is the relative surface wetness (dimensionless); $f_g$ is the green canopy fraction (dimensionless); $f_t$ is the plant temperature constraint (dimensionless); $f_m$ is the plant moisture constraint (dimensionless); $f_{sm}$ is the soil moisture constraint factor (dimensionless); $\alpha$ is the PT coefficient (1.26); $\Delta$ is the slope of the saturation vapor pressure curve (kP_a °C⁻¹); γ is the psychrometric constant (0.066 kP_a °C⁻¹); G is the soil heat flux (W·m⁻²); λ is the latent heat of vaporization (MJ/kg); $R_n$ is the net radiation (W·m⁻²); $R_{nc}$ is the canopy net radiation (W·m⁻²); and $R_{ns}$ is the soil surface net radiation. The radiation flux calculations are shown in the following equations [12]:

$$R_n=\left(1-Albedo\right)R_s+R_L\downarrow -R_L\uparrow$$

(5)

$$R_{ns}=R_n\exp \left(-k_{R_n}LAI\right)$$

(6)

$$R_{nc}=R_n-R_{ns}$$

(7)

$$G={\alpha }_g\left(1-FVC\right)R_n$$

(8)

where $R_s$ is the solar shortwave radiation; $R_L\downarrow$ is the downward longwave radiation; $R_L\uparrow$ is the upward longwave radiation; $Albedo$ is the surface albedo; $k_{R_n}$ is the extinction coefficient, set to 0.6; ${\alpha }_g$ is an empirical coefficient, set to 0.3; LAI is the leaf area index; and FVC is the fractional vegetation cover.

Figure 1. Schematic diagram of the improved PT-JPL ET model.

The biophysical constraint functions in the PT-JPL model are calculated as follows:

$$f_{wet}=R{H}^4$$

(9)

$$f_g={\displaystyle \frac{f_{APAR}}{f_{IPAR}}}$$

(10)

$$f_{\mathrm{A}\mathrm{P}\mathrm{A}\mathrm{R}}=m_1\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}+b_1$$

(11)

$$f_{\mathrm{I}\mathrm{P}\mathrm{A}\mathrm{R}}=m_2\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}+b_2$$

(12)

where $RH$ is the relative humidity; $f_{APAR}$ is the fraction of absorbed PAR by the canopy; $\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}$ is the normalized difference vegetation index; and $m_1$,$b_1$, $m_2$, $b_2$ are parameters set to 1.2 × 1.136, 1.2 × (−0.04), 1, and −0.05, respectively.

$$f_t=\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\left({\displaystyle \frac{T_a-T_{opt}}{T_{opt}}}\right)}^2\right)$$

(13)

where $T_a$ is the air temperature and $T_{opt}$ is the optimal plant growth temperature, determined according to crop types in the study area. For crops in this study region, $T_{opt}$ is set to 25 °C.

$$f_m={\displaystyle \frac{f_{APAR}}{f_{APAR\mathrm{m}\mathrm{a}\mathrm{x}}}}$$

(14)

where $f_{APAR}$ is the fraction of absorbed PAR by the canopy and $f_{APAR\mathrm{m}\mathrm{a}\mathrm{x}}$ is the maximum fraction of absorbed PAR by the canopy, representing the maximum PAR absorption ratio under optimal vegetation growth conditions, calculated as the 95th percentile of the computed $f_{APAR}$ values.

$$f_{sm}=R{H}^{{\displaystyle \frac{VPD}{\beta }}}$$

(15)

where VPD is the vapor pressure deficit and RH is the relative humidity.

In the soil evaporation module of the PT-JPL model, the estimation of the key factor $f_{sm}$ affecting soil evaporation relies on atmospheric humidity constraints (calculated through RH) [56,57]. This expression assumes an equilibrium state between the surface and atmosphere, which is rarely fully achieved in reality, particularly under drier conditions (such as those in arid regions). RH may not accurately reflect the true moisture state of the soil [58,59]. Therefore, we use the LSWI to replace the RH in the original model, directly reflecting soil moisture to redefine $f_{sm}$. The improved formulas are as follows:

$$\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}}$$

(16)

$$f_{sm}={\displaystyle \frac{1+\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}}{1+{\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}}_{max}}}$$

(17)

where NIR (near-infrared band) is sensitive to soil and vegetation moisture; SWIR (shortwave infrared band) responds more directly to soil moisture; and ${\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}}_{max}$ is the maximum land surface water index in the study area, using the value at the 95% cumulative frequency of LSWI during the growing season as ${\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}}_{max}$. To ensure physical meaningfulness, the calculated soil moisture constraint factor ($f_{sm}$) is strictly constrained within the range of [0, 1]. Values exceeding these bounds are clipped to 0 or 1, respectively.

2.2. Construction of CUWFM

Based on the high spatial resolution ET inversion results from the improved PT-JPL model, this study developed a crop ET spatiotemporal fusion model to achieve continuous and fine-scale monitoring of hydrological processes during the crop growing season (Figure 2). This model comprehensively considers crop phenological characteristics, crop type distribution, and factors highly correlated with ET. It employs a spatiotemporal fusion algorithm that combines unmixing and weight functions, utilizing dense temporal but low spatial resolution ET products (PML-V2) to perform temporal extrapolation of the high spatial resolution ET data retrieved by the improved PT-JPL model. This approach achieves effective fusion of ET data at two different scales, ultimately generating daily 30 m spatial resolution ET data.

Figure 2. Flowchart of the Crop Unmixing and Weight Fusion Model for ET (CUWFM).

The model comprises four main steps: (1) spatial downscaling based on unmixing operations; (2) selection of spatiotemporal fusion base dates; (3) hierarchical screening strategy for similar pixel selection; and (4) weight-based temporal prediction.

2.2.1. Spatial Downscaling Based on Unmixing Operations

ET influence factors $E{F}_{fi}$ are used to unmix coarse pixels, initially “refining” the coarse resolution ET data from PML into high resolution data. Although coarse resolution ET products provide temporal continuity, they contain numerous mixed pixels. When directly fused with PT-JPL model-retrieved data, their low resolution limits the spatial details of the generated results. Utilizing ET influence factors to perform unmixing operations on PML ET products enables better matching with PT-JPL model-retrieved data.

The ET influence factor $E{F}_{fi}$ is calculated based on NDVI and LSWI, where NDVI represents vegetation vigor, chlorophyll content, etc., and LSWI indicates the water content of surface vegetation and soil. Daily results are obtained through Whittaker filtering. The calculation formula is as follows:

$$\begin{array}{l}E{F}_{fi}\\ =1-\mathrm{m}\mathrm{a}\mathrm{x}(min\left({\displaystyle \frac{FV{C}_{max}-FVC}{FV{C}_{max}-FV{C}_{min}}},1\right),0)\cdot \mathrm{m}\mathrm{a}\mathrm{x}(min\left({\displaystyle \frac{LSW{I}_{max}-LSWI}{LSW{I}_{max}-LSW{I}_{min}}},1\right),0)\\ \\ FVC=FV{C}_{max}\times (\mathrm{m}\mathrm{a}\mathrm{x}\left(\min \left({\displaystyle \frac{NDVI-0.1}{0.9-0.1}}\right),1\right),0)\\ \end{array}$$

(18)

The selection of this multiplicative combination ($FV{C}_{norm}\times LSW{I}_{norm}$) as the ET influence factor ($E{F}_{fi}$) is based on the physical mechanism of evapotranspiration in arid environments. Single-factor indices such as NDVI or FVC primarily reflect vegetation greenness and canopy structure (potential transpiration surface) but fail to capture short-term water stress conditions. Conversely, LSWI provides information on canopy and soil moisture content but lacks structural information. By employing a multiplicative structure, the constructed $E{F}_{fi}$ effectively represents the coupled control of vegetation quantity and water availability on ET. This logic aligns with the core mechanism of the PT-JPL model (Equation (2)), where latent heat flux is determined by the interaction of vegetation fraction ($f_g$) and moisture constraints ($f_{sm}$), making it superior to single-index approaches for downscaling in water-limited agricultural regions.

The ET influence factor can represent differences in ET, thus enabling differentiation of ET results between different land plots. The relationship formula for ET between land plots can be expressed as

$$\begin{array}{c}E{T}_{f1}:E{T}_{f2}:\cdots E{T}_{fi}={EF}_{f1}:{EF}_{f2}:\cdots {EF}_{fi}\end{array}$$

(19)

In low-resolution pixels of ET data, each pixel contains multiple high-resolution pixels, with each high-resolution pixel numbered as i. Here, $E{T}_{fi}$ refers to the ET value of the i-th high-resolution pixel, while ${EF}_{fi}$ represents the ET allocation coefficient for this pixel, indicating the contribution of each high-resolution pixel to the total ET of the entire coarse-resolution pixel. This is expressed by the following formula:

$$E{F}_{co}={\displaystyle \frac{1}{n}}\sum _{i=1}^n E{F}_{fi}$$

(20)

where ${EF}_{co}$ is the average ET influence factor of all high-resolution pixels within the coarse pixel; n is the total number of high-resolution pixels contained within one coarse-resolution pixel.

$${\displaystyle \frac{E{T}_{fi}}{E{T}_{co}}}={\displaystyle \frac{{EF}_{fi}}{{EF}_{co}}}$$

(21)

where ${ET}_{co}$ is the original coarse-resolution ET value from PML; $E{T}_{fi}$ is the result obtained through unmixing operations.

2.2.2. Selection of Spatiotemporal Fusion Base Dates

From the multiple dates of ET data retrieved by remote sensing ET models and the unmixed PML-ET data, “base date pairs” are identified to ensure that the input data for temporal prediction has sufficient physical correlation. The selection process strictly follows two key principles. First, the base dates ($t_m$, $t_n$) and the prediction date ($t_p$) must strictly fall within the same phenological period. This period is determined based on crop-specific phenological calendars derived from field surveys and local agricultural archives for the current year. To account for the spatial variability in planting and harvest dates across the region, these periods are defined as broad physiological stages (e.g., green-up, jointing, heading, and maturity) rather than precise daily timestamps. This strategy ensures that the crop physiological state remains consistent across the fusion dates, effectively accommodating slight spatial temporal offsets. Second, regarding the specific selection algorithm, when multiple candidate base dates are available within the same phenological window, the system follows the principle of temporal proximity, selecting the pair ($t_m$, $t_n$) that is closest in time to the prediction date $t_p$. This minimizes errors caused by rapid land cover changes or phenological shifts that might occur over longer intervals.

Here, base dates $t_m$, $t_n$ refer to two known time points used to construct spatiotemporal relationships, at which both high spatial resolution ET data and low spatial resolution ET data are available simultaneously; prediction date $t_p$ refers to the target date requiring high-resolution ET prediction, for which only low spatial resolution PML ET data is available. The final output of spatiotemporal fusion is the high spatial resolution ET data at prediction date $t_p$.

2.2.3. Hierarchical Screening Strategy for Similar Pixel Selection

A sliding window approach is employed, utilizing high-resolution ET data at times $t_m$ and $t_n$ to screen for similar pixels. In this study, the window side length ($w$) was set to a fixed size of 31 pixels (covering an area of approx. 1 km × 1 km). This parameter selection is based on a trade-off analysis between sample sufficiency and local spatial dependence; sensitivity tests indicated that this size effectively avoids prediction gaps in fragmented cropland boundaries while minimizing the inclusion of distant, uncorrelated pixels.

A hierarchical screening strategy is adopted, with preliminary screening of pixels within the sliding window based on agricultural crop distribution maps. The range of the sliding window can be dynamically adjusted according to the spatial distribution characteristics of different crop types, retaining only pixels that belong to the same crop cover type as the central pixel.

Based on preliminary screening, final screening of candidate pixels is performed according to ET differences. An ET difference threshold is set to eliminate pixels that do not meet the criteria, retaining only pixels that satisfy ET difference requirements as similar pixels. The final screening correlation formula is as follows:

$$\left|E{T}_F\left(i,j\right)-E{T}_F\left({\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}}\right)\right|\le \sigma$$

(22)

where $E{T}_F(i,j)$ is the ET value of the i-th high spatial resolution pixel in the window after preliminary screening; $E{T}_F(w/2,w/2)$ is the ET value of the central pixel in the sliding window; and σ is the standard deviation of ET values for all pixels after preliminary screening.

2.2.4. Weight-Based Temporal Prediction

Similarity weights are calculated based on spectral similarity, ET indicator information similarity, and geographical distance. Using the calculated normalized weights, temporal weights, and conversion coefficients, weighted prediction of ET values for target pixels is performed, completing the spatiotemporal fusion of high spatial resolution but temporally sparse ET data with the daily low spatial resolution ET product PML-V2, generating daily high spatial resolution ET data.

Spectral similarity is an important indicator for evaluating consistency between fine- and coarse-resolution images. Given the characteristics of single-band ET data, this study adopts the method of Zhu et al. [49], using a relative difference-based metric to calculate spectral similarity. The spectral similarity $S_{i,j,t_k}$ calculation formula is

$$S_{i,j,t_k}=1-{\displaystyle \frac{1}{2}}\left[\begin{array}{c}{\displaystyle \frac{|F_{i,j,t_m}-C_{i,j,t_m}|}{F_{i,j,t_m}+C_{i,j,t_m}}}+{\displaystyle \frac{|F_{i,j,t_n}-C_{i,j,t_n}|}{F_{i,j,t_n}+C_{i,j,t_n}}}\end{array}\right](k=m,n)$$

(23)

where $F_{i,j,t_k}$ and $C_{i,j,t_k}$ represent pixel values of fine-resolution and coarse-resolution images at location (i,j) and time point $t_k$, respectively, and $t_m$ and $t_n$ represent two known time points. To prevent numerical instability when ET values approach zero (e.g., in non-vegetated areas or non-growing seasons), a small constant $ϵ$ (10⁻⁶) was added to the denominator during the calculation.

In calculating ET indicator information similarity, we introduce the ET influence factor ($E{F}_{fi}$) calculated in Equation (18) (denoted as AF in Equation (24) for brevity). This factor reflects the coupled status of vegetation and soil moisture. The similarity weight ($WA{F}_{i,j,t_k}$) is determined by assessing the consistency of AF changes between the central pixel and similar pixels. Specifically, the correlation is computed using the AF values at the selected base date pair ($t_m$, $t_n$). This ensures that pixels exhibiting similar temporal evolution trends to the central pixel during the reference interval are assigned higher weights. The calculation formula is

$$WA{F}_{i,j,t_k}={\displaystyle \frac{E\left[\right(A{F}_{ij}-E\left(A{F}_{ij}\right)\left)\right(A{F}_{central}-E\left(A{F}_{central}\right)\left)\right]}{\sqrt{D\left(A{F}_{ij}\right)}\cdot \sqrt{D\left(A{F}_{central}\right)}}}$$

(24)

where central represents the central pixel (w/2, w/2), $E( )$ denotes expectation, and $D( )$ denotes variance.

The distance weight calculation formula is

$$d_{i,j,t_k}=1+{\displaystyle \frac{\sqrt{{\left(\frac{w}{2}-i\right)}^2+{\left(\frac{w}{2}-j\right)}^2}}{\left(\frac{w}{2}\right)}}$$

(25)

where $d_{i,j,t_k}$ is the distance weight, and (w/2, w/2) is the central pixel.

Next, the normalized weight of similar pixels is calculated using the formula

$$W_{i,j,t_k}={\displaystyle \frac{\frac{1}{D_{i,j,t_k}}}{\mathrm{\Gamma }\left(\frac{1}{D_{i,j,t_k}}\right)}}$$

(26)

$$D_{i,j,t_k}=\left(1-S_{i,j,t_k}\right)\left(1-WA{F}_{i,j,t_k}\right)\times d_{i,j,t_k}$$

(27)

where $W_{i,j,t_k}$ is the normalized weight; Γ represents $\sum _{i=1}^w \sum _{j=1}^w \left(X\right)$; $S_{i,j,t_k}$ is the spectral similarity between fine resolution pixels and coarse resolution pixels in similar pixels; $WA{F}_{i,j,t_k}$ represents the ET indicator information similarity weight between similar pixels and central pixels; and $d_{i,j,t_k}$ is the distance weight.

Finally, using the calculated normalized weights, temporal weights, and conversion coefficients, weighted prediction of target pixel ET values is performed. The relevant formulas are

$$E{T}_F^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_p}=T_m\times E{T}_{F,m}^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_p}+T_n\times E{T}_{F,n}^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_p}$$

(28)

$$E{T}_{F,k}^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_p}=E{T}_F^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_k}+\mathrm{\Gamma }\left(W_{i,j,t_k}\times V_{i,j,t_k}\times (E{T}_C^{i,j,t_p}-E{T}_C^{i,j,t_k})\right),(k=m,n)$$

(29)

$$T_k={\displaystyle \frac{\frac{1}{\left|\mathrm{\Gamma }\left(E{T}_C^{i,j,t_k}-E{T}_C^{i,j,t_p}\right)\right|}}{\sum _{k=m,n} \left(\frac{1}{\left|\mathrm{\Gamma }\left(E{T}_C^{i,j,t_k}E{T}_C^{i,j,t_p}\right)\right|}\right)}}$$

(30)

where $E{T}_F$ and $E{T}_C$ represent fine resolution and coarse resolution pixels, respectively, $t_p$ is the prediction date; w is the sliding window side length; (i,j) is the similar pixel location; (w/2, w/2) is the central pixel location; $E{T}_F^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_p}$ is the final high resolution ET calculated by the model at time $t_p$; $E{T}_{F,m}^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_p}$ and $E{T}_{F,n}^{{\displaystyle \frac{w}{2}},{\displaystyle \frac{w}{2}},t_p}$ are high resolution ET values at time $t_p$ predicted based on $t_m$ and $t_n$, respectively; $W_{i,j,t_k}$ is the normalized weight; $V_{i,j,t_k}$ is the conversion coefficient for similar pixels, obtained through linear regression analysis of each similar pixel at reference dates $t_m$ and $t_n$; $T_k$ represents temporal weight; and Γ represents $\sum _{i=1}^w \sum _{j=1}^w \left(X\right)$.

Repeating the above steps for all target dates and pixels generates daily 30 m resolution ET data for the study area.

3. Study Area and Data

3.1. Study Area

Changji Hui Autonomous Prefecture in Xinjiang Uygur Autonomous Region is located in northern Xinjiang, China, with geographical coordinates ranging from 43°20′N to 45°00′N and 85°17′E to 91°32′E [60]. Changji Prefecture belongs to a typical continental arid and semi-arid climate zone, characterized by an average annual temperature of approximately 10.4 °C, annual precipitation of only 150–200 mm, and annual evaporation as high as 1500–3000 mm. The region exhibits scarce precipitation, intense evaporation, arid climate, and fragile ecological conditions (Figure 3).

Figure 3. Location of the study area.

Changji Prefecture represents a typical pure irrigation agricultural region in China and is also one of the most severely water-scarce areas in the world. The cultivated land area exceeds 4 million hectares, primarily devoted to the cultivation of wheat, maize, and cotton [61]. With the continuous expansion of agricultural production scale, scientific water resource management has become the key to sustainable development in this region.

In Changji Prefecture, irrigated agriculture is dominated by three staple crops: winter wheat, maize, and cotton. To facilitate the interpretation of evapotranspiration (ET) dynamics and the selection of reference dates in the fusion framework, the typical phenological calendars of these crops are briefly described here. Maize is generally cultivated from May to September. The growth stages include the seedling stage (May–early June), jointing (mid-June–early July), tasseling and silking (mid- to late July), followed by grain filling until maturity in September; peak ET typically occurs during the tasseling–grain filling period (July–August). Winter wheat is a winter crop, sown in October and harvested by late June of the following year. Its phenology spans the overwintering stage (October–March), regreening and jointing (March–May), heading, flowering, and grain filling (mid-May–mid-June); ET reaches its maximum during late May to mid-June. Cotton is planted in late May and harvested in September, with early vegetative growth in late May–June, squaring and flowering in June–July, and boll formation and opening peaking in August; ET is highest during the boll-setting and opening stages. These phenological characteristics were explicitly used in the CUWFM to ensure that paired reference and prediction dates fall within consistent growth stages, thereby maintaining physiological coherence in the fusion process.

3.2. Input Data for PT-JPL Model

3.2.1. Satellite Data

This study utilized Landsat 8 and Sentinel-2 satellite data to retrieve surface parameters required for the improved PT-JPL model, including the normalized difference vegetation index (NDVI), land surface water index (LSWI), surface albedo, fractional vegetation cover (FVC), and leaf area index (LAI). Specifically, for the Changji Prefecture study area, the dataset covers the entire crop growing season from 30 March to 16 October 2020. By combining observations from both sensors, a total of 27 cloud-free images were acquired, achieving an average revisit interval of approximately 8 days. This high-frequency time series effectively captures the key phenological transition stages of winter wheat, maize, and cotton. All data were resampled to a 30 m spatial resolution.

For the two validation sites (Daman in Gansu and Changwu in Shaanxi), all available cloud-free Landsat 8 and Sentinel-2 scenes during the 2020–2023 growing seasons were collected. For Changji Prefecture, we used all available cloud-free images from 30 March to 16 October 2020, covering the full growing season of maize, cotton, and winter wheat. These images were employed to drive the improved PT-JPL model and to construct both the reference and prediction dates in the CUWFM framework, with reference dates selected from acquisitions falling within consistent phenological stages. In the results section, we particularly highlight the period from 23 July to 8 August 2020, which corresponds to the peak growth stages of the three major crops.

The calculation methods for NDVI and LSWI are well-established and will not be elaborated here. The fractional vegetation cover (FVC) was retrieved based on the pixel dichotomy method [62], with the calculation formula as follows:

$$FVC={\displaystyle \frac{NDVI-NDV{I}_{min}}{NDV{I}_{max}-NDV{I}_{min}}}$$

(31)

where $NDV{I}_{min}$ represents the NDVI value for bare soil or non-vegetated areas; $NDV{I}_{max}$ represents the NDVI value for completely vegetated areas; and values at 5% and 95% cumulative frequencies of NDVI were selected as $NDV{I}_{min}$ and $NDV{I}_{max}$, respectively.

The study area is primarily dominated by three major crop types: winter wheat, maize, and cotton. The leaf area index (LAI) was retrieved from FVC based on the Beer–Lambert law extinction model, which describes the attenuation of light through the canopy as follows:

$$LAI=-(1/k)\times \mathrm{l}\mathrm{n}(1-FVC)$$

(32)

where FVC is the fractional vegetation cover, and k is the light extinction coefficient. Although the actual k value depends on canopy geometry and varies across different phenological stages for wheat, maize, and cotton, a constant value of 0.5 was adopted in this study. This value assumes a spherical leaf angle distribution, which is a widely accepted standard approximation in regional remote sensing applications. Adopting this threshold effectively minimizes uncertainties when detailed field measurements of canopy structure for specific crop types are unavailable [63,64].

Surface albedo was extracted based on remote sensing data and atmospheric conditions. Liang [65] provided simple formulas for calculating surface albedo from various satellite sensors, which were adopted in this study.

$$\mathrm{A}\mathrm{l}\mathrm{b}\mathrm{e}\mathrm{d}\mathrm{o}=0.356a2+0.130a4+0.373a5+0.085a6+0.072a7-0.0018$$

(33)

where $a2$, $a4$, $a5$, $a6$, and $a7$ correspond to Landsat Band 2, Band 4, Band 5, Band 6, and Band 7, respectively; and to Sentinel-2 Band 2, Band 4, Band 8, Band 11, and Band 12, respectively.

3.2.2. Meteorological Data

The meteorological data driving the PT-JPL model were derived from the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis (ERA5), with an original spatial resolution of 0.25° × 0.25° and a daily temporal resolution. The data include relative humidity (RH), mean temperature (Ta), atmospheric pressure (Pa), solar shortwave radiation ($R_s$), downward longwave radiation ($R_L\downarrow$), and upward longwave radiation ($R_L\uparrow$). For this study, ERA5 daily meteorological fields covering 2020–2023 for Changji Prefecture were extracted and used to drive the PT-JPL model.

To maintain spatial scale consistency with satellite-retrieved surface parameters, bilinear interpolation was employed to resample the coarse-resolution meteorological data to 30 m spatial resolution. This interpolation method assumes that atmospheric variables exhibit spatially smooth variations within local regions, effectively eliminating artificial “blocky” artifacts caused by the resolution gap. Crucially, the calculation of Net Radiation ($R_n$) integrates these smooth atmospheric fluxes with high-resolution surface parameters, specifically the 30 m surface albedo retrieved from satellite data. Since the spatial heterogeneity of$R_n$at the field scale is primarily dominated by land surface characteristics (albedo) rather than the relatively homogeneous incoming atmospheric irradiance, this strategy ensures that the generated$R_n$preserves realistic 30 m spatial details without introducing artificial coarse textures.

3.3. Input Data for CUWFM

The primary input data for the spatiotemporal fusion model include improved PT-JPL-ET data with a 30 m spatial resolution but lower temporal resolution, and the daily PML-V2 ET dataset with a 500 m spatial resolution. The PML-V2 ET dataset was obtained from the National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/). For this study, we extracted the PML-V2 ET data covering 2020–2023 for the validation sites and 30 March to 16 October 2020 for Changji Prefecture.

In addition to the primary ET datasets, high-quality auxiliary data were incorporated to support the fusion framework. For the spatial domain, the 2020 crop distribution map of Changji Prefecture was derived from the “Plot-level Cropping Structure Dataset Based on Sentinel Images and Phenology Information” (available at: https://www.geodoi.ac.cn/WebEn/doi.aspx?Id=3876, accessed on 16 July 2025). Generated using high-resolution Sentinel-2 imagery and phenological features, this dataset achieves an overall classification accuracy of approximately 87%, providing a precise basis for identifying winter wheat, maize, and cotton fields during crop-specific downscaling. Regarding the temporal domain, key phenological dates for these target crops were acquired through a combination of field surveys, on-site interviews with local farmers, and consultation of local agricultural archives. These multi-source data defined precise time windows for key growth stages (e.g., green-up, heading, and maturity), which were explicitly used to constrain the selection of spatiotemporal fusion base dates.

3.4. Site Observation Data

This study collected surface observation data from two agricultural sites in arid and semi-arid regions to validate the accuracy of remote sensing-estimated ET. The validation sites are detailed in Table 1, where data from Daman station were obtained from the Heihe Watershed Allied Telemetry Experimental Research (HiWATER) network, and data from Changwu station were obtained from the Changwu National Field Scientific Observation and Research Station for Agricultural Ecosystem in Shaanxi Province.

Table 1. Validation sites used in this study.

Site	Latitude	Longitude	Land Cover Type	Years
Daman	38.86°N	100.37°E	Maize	2020–2023
Changwu	35.20°N	107.67°E	Winter wheat	2021–2023

Eddy covariance observations provide daily 30 min latent surface heat flux data, which, after rigorous quality control and temporal accumulation, yield daily ET amounts for validating remote sensing-estimated ET in this study. Large weighing lysimeter observations directly provide daily ET data, which, after rigorous quality control, are used to validate remote sensing-estimated ET in this study.

Considering the spatial representativeness differences between remote sensing pixels and ground observations, this study adopted different validation strategies. For the 30 m resolution PT-JPL-ET product, Daman station employed 3 × 3 pixel averages centered on the station to better match the source area footprint of eddy covariance observations; the lysimeter observation area at Changwu station is close to the scale of a single 30 m pixel, so station pixels were directly used for validation. Validation metrics include the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe efficiency coefficient (NSE), and percent bias (PBIAS).

4. Results

4.1. Direct Validation of Improved PT-JPL-ET

To comprehensively assess the effectiveness of the improvement strategy, this study first conducted validation of the original PT-JPL model. As shown in Figure 4a and Figure 5a, the original PT-JPL model demonstrated certain ET estimation capabilities at both validation sites, but exhibited obvious accuracy limitations. At the Daman station, the original model achieved an R² of 0.74, an NSE of 0.71, an RMSE of 1.20 mm/day, an MAE of 0.97 mm/day, and exhibited a systematic underestimation bias (PBIAS) of −7.99%. At the Changwu station, the original model showed an R² of 0.70, an NSE of 0.62, an RMSE of 0.82 mm/day, an MAE of 0.69 mm/day, and a PBIAS of 1.50%. These results indicate that the original PT-JPL model indeed suffers from accuracy deficiencies under arid and semi-arid conditions, particularly exhibiting pronounced underestimation phenomena at the maize site.

Figure 4. Comparison of daily ET estimation accuracy between the original and improved PT-JPL models at Daman station (2020–2023). (a) The original PT-JPL model results; (b) the improved PT-JPL model results.

Figure 5. Comparison of daily ET estimation accuracy between the original and improved PT-JPL models at Changwu station (2021–2023). (a) The original PT-JPL model results; (b) the improved PT-JPL model results.

Based on these findings, the model performance was significantly enhanced after implementing the improvement strategy of using LSWI to replace relative humidity. The improved PT-JPL model serves as the foundation of this entire study, with its data quality directly affecting subsequent fusion results. To assess model performance, this study conducted systematic validation based on multi-year observation data from two typical arid and semi-arid agricultural sites. The validation results demonstrate that the improved PT-JPL model exhibits excellent ET estimation capability in arid and semi-arid agricultural regions. At the Daman station with maize as the underlying surface, the model showed good daily-scale estimation accuracy, with an R² reaching 0.78 and an NSE of 0.77. At the Changwu station with winter wheat as the underlying surface, the model also demonstrated robust performance, with an R² of 0.8 and NSE reaching 0.69. This accuracy level not only validates the model’s reliability but, more importantly, confirms the effectiveness of the improvement strategy using LSWI to replace relative humidity under arid and semi-arid conditions.

In the validation at Daman station (Figure 4b), the improved PT-JPL-ET achieved an RMSE of only 1.06 mm/day, an MAE of 0.85 mm/day, and a percent bias (PBIAS) as low as 0.07%. This accuracy level demonstrates the model’s advantage in capturing the dynamic changes in maize ET. Notably, the model maintained stable estimation accuracy across different stages of the maize growing season. During the heading–grain filling period (July–August) when ET rates were high, the model could still accurately capture the high-value variations in daily ET at 6–7 mm, benefiting from the direct response characteristics of the LSWI to soil moisture conditions, effectively avoiding the applicability limitations of the relative humidity parameter in the original model under arid conditions. In the validation at Changwu station (Figure 5b), the model similarly demonstrated good adaptability. The RMSE was 0.74 mm/day, and the MAE was 0.56 mm/day, showing excellent accuracy performance. However, the PBIAS was 15.64%, slightly higher than the Daman station. This difference mainly stems from crop physiological characteristic differences, as winter wheat has more complex phenological transition processes than maize, particularly during the spring green-up period and late maturation period, where rapid changes result in higher model parameter sensitivity.

Comparative analysis demonstrates that the LSWI improvement strategy achieved substantial accuracy enhancements at both sites. Compared to the original version, the improved PT-JPL model at Daman station showed an R² improvement of 5.4%, an NSE improvement of 8.5%, RMSE, and MAE reductions of 11.7% and 12.4%, respectively, with the most significant improvement being PBIAS changing from −7.99% to the nearly ideal 0.07%, essentially eliminating systematic bias. At Changwu station, the improvements were equally notable, with R² and NSE increasing by 14.3% and 11.3%, respectively, while RMSE and MAE decreased by 9.8% and 18.8% respectively. These improvements fully demonstrate that using LSWI to directly reflect soil moisture status has stronger applicability in arid and semi-arid agricultural regions compared to traditional relative humidity parameters.

4.2. Temporal Variations in PT-JPL-ET During Crop Growing Seasons

PT-JPL-ET effectively captures the temporal variation characteristics of ET during different crop growing seasons. Figure 6 shows the ET time series for Daman agricultural station from 2020 to 2023 and Changwu station from 2021 to 2023, respectively. The PT-JPL-ET estimation results exhibited high consistency with ground observation data at both sites.

Figure 6. Temporal variations in PT-JPL-ET during crop growing seasons: (a) time series comparison between PT-JPL-ET and ground observations at Daman station; (b) time series comparison between PT-JPL-ET and ground observations at Changwu station.

4.2.1. ET Variations During Maize Growing Season at Daman Station

Analysis of maize phenological periods reveals that ET exhibited variation patterns highly coupled with crop growth and development stages (Figure 6a). During the seedling stage (May to early June), ET was approximately 1–3 mm/day, primarily dominated by soil evaporation. During the jointing stage (mid-June to early July), ET rapidly increased to 4–5 mm/day, with vegetation transpiration gradually strengthening. During the heading stage (mid-to-late July), ET rapidly climbed to 6–7 mm/day. During the grain filling stage (August), ET reached the peak value of the entire growing season at 7–8 mm/day, when both leaf area index and vegetation transpiration reached their maximum. During the maturation stage (September), ET rapidly declined to 1–3 mm/day as leaves senesced.

The ET peaks for all years occurred during maize heading and grain filling periods, highly consistent with the critical growth periods when maize water demand was greatest. During 2020–2023, although there was some interannual variation, the overall change trends remained consistent, indicating that the model possesses good interannual stability.

4.2.2. ET Variations During Winter Wheat Growing Season at Changwu Station

During the winter wheat growing season (October to June of the following year (Figure 6b)), in the green-up period (March to early April), ET began to rise with temperature recovery and vegetation growth resumption, gradually increasing to approximately 2–3 mm/day. During the jointing stage (mid-April to early May), ET continued to strengthen, reaching 3–4 mm/day, with vegetation transpiration significantly enhanced. During the heading to maturation period (mid-to-late May to mid-June), ET reached the peak value of the entire growing season at 6–7 mm/day, when the leaf area index was at its maximum and vegetation transpiration reached its strongest. During the harvest period (late June), ET rapidly declined after maintaining high levels in the early period, dropping from 6 mm/day to 2–3 mm/day, with vegetation transpiration rapidly weakening as leaves senesced and grains matured.

Compared to Daman station, with maize as the underlying surface type, winter wheat at Changwu station exhibited ET variations with longer growing cycles and more pronounced seasonal characteristics. Winter wheat showed steeper ET rise periods in spring (March–May), reflecting its sensitive response to changes in spring hydrothermal conditions. PT-JPL-ET effectively captured ET variation patterns across all stages of the winter wheat growing season, particularly demonstrating good estimation accuracy in maintaining low values during the overwintering period and rapid increases during spring.

4.3. Validation of CUWFM Fusion Results Against Ground Observations

To further verify the physical accuracy of the generated daily high-resolution ET, we conducted a continuous validation experiment at the Daman station throughout July 2020. This period corresponds to the peak growing season of maize, characterized by vigorous water consumption and rapid daily fluctuations in evapotranspiration due to weather variability.

Figure 7 illustrates the time series comparison between the CUWFM-generated daily 30 m ET and the ground-measured eddy covariance (EC) data. Visual analysis reveals that the fused ET (red dashed line) exhibits high consistency with the ground observations (black solid line). Crucially, the CUWFM successfully captured the high-frequency temporal dynamics, including sharp declines in ET on specific days (e.g., 3 July, 13 July, and 24 July), which likely correspond to cloudy or rainy weather conditions. This demonstrates that the fusion algorithm effectively integrates the temporal information from the coarse-resolution data to reconstruct daily variations between satellite overpasses.

Figure 7. Time series validation of daily ET estimated by CUWFM against ground observations at Daman station during July 2020.

Quantitative assessment confirms the reliability of the fusion results. Over the month-long validation period, the CUWFM achieved a coefficient of determination (R²) of 0.770 and a root mean square error (RMSE) of 0.716 mm/day. These metrics indicate that the proposed framework can produce reliable daily ET estimates at the field scale, satisfying the accuracy requirements for precision agricultural monitoring even on dates without high-resolution satellite acquisitions.

4.4. Assessment of Spatial Patterns in CUWFM Fusion Results

To validate the performance of CUWFM, this study used the 2020 crop growing season in Changji Prefecture, Xinjiang, as an example, selecting typical dates from different crop phenological periods (2 June, 4 July, 5 August, and 6 September) to conduct comparative analysis between the CUWFM-ET fusion results and the original PT-JPL-ET. The fusion process employed one high-resolution ET scene before and after each target date as base dates, combined with concurrent PML-V2 data to achieve spatiotemporal fusion. Specifically, the 2 June fusion result was generated through joint prediction using PT-JPL ET data from 25 May and 10 June, along with PML-V2 ET data from 25 May, 2 June, and 10 June (five scenes total); the 4 July fusion result was generated using PT-JPL ET from 26 June and 12 July, combined with PML-V2 ET from 26 June, 4 July, and 12 July (five scenes total); the 5 August fusion result was generated using PT-JPL ET from 28 July and 13 August, combined with PML-V2 ET from 28 July, 5 August, and 13 August (five scenes total); the 6 September fusion result was generated using PT-JPL ET from 29 August and 14 September, combined with PML-V2 ET from 29 August, 6 September, and 14 September (five scenes total).

The spatial pattern assessment results for winter wheat cultivation areas demonstrate that the CUWFM successfully maintained the fine spatial characteristics of the original PT-JPL-ET data (Figure 8). In terms of spatial continuity, the fused ET data completely preserved the clear contours of field boundaries, effectively avoiding the boundary blurring issues commonly encountered in traditional fusion methods. Regarding spatial heterogeneity, the CUWFM-ET accurately captured the spatial variations in ET within individual fields caused by irrigation uniformity differences, demonstrating the model’s advantage in preserving spatial details. From the perspective of crop phenological response characteristics, the fusion results accurately reflected the seasonal variation patterns of winter wheat ET. The dates 2 June and 4 July corresponded to the heading and maturation periods of winter wheat, respectively, when ET maintained relatively high levels (4–6 mm/day), highly consistent with the physiological characteristics during periods of vigorous water demand for this crop. Quantitative analysis results showed (Figure 9) that CUWFM-ET maintained high numerical consistency with PT-JPL-ET at the pixel level. The MAE ranged from 0.05 to 0.21 mm/day, which, considering that daily ET typically ranges from 0 to 8 mm/day, indicates that the bias introduced by the fusion process was minimal. The root mean square error (RMSE) ranged from 0.07 to 0.25 mm/day, further confirming the high accuracy of the fusion results. The mean relative error (MRE) was controlled within 1.3–7.4%, demonstrating satisfactory overall relative accuracy.

Figure 8. Spatial distribution of PT-JPL-ET and CUWFM-ET in winter wheat cultivation areas. The black boxes indicate the locations of the zoomed-in sub-regions, and (a–c) represent the detailed views of these corresponding areas.

Figure 9. Comparison between CUWFM-ET and PT-JPL-ET in winter wheat cultivation areas.

The validation results for cotton cultivation areas further confirmed the universality and robustness of the CUWFM (Figure 10). The results showed that the CUWFM also performed excellently when processing cotton cultivation areas, successfully maintaining fine spatial differences within farmlands and gradient changes in ET between fields. In terms of phenological ET response, the fusion results accurately reflected the water demand characteristics of cotton: 4 July and 5 August were during the flowering and boll formation period, when ET reached the peak values of the entire growing season; and 2 June and 6 September corresponded to the bud formation and boll opening periods, respectively, when ET was relatively low, consistent with cotton’s “low at both ends, high in the middle” water demand pattern. Accuracy validation results showed (Figure 11) that the fusion accuracy for cotton cultivation areas was essentially equivalent to that of winter wheat: an MAE of 0.06–0.29 mm/day, an RMSE of 0.06–0.3 mm/day, and an MRE controlled within 1.1–8.6%.

Figure 10. Spatial distribution of PT-JPL-ET and CUWFM-ET in cotton cultivation areas. The black boxes indicate the locations of the zoomed-in sub-regions, and (a–c) represent the detailed views of these corresponding areas.

Figure 11. Comparison between CUWFM-ET and PT-JPL-ET in cotton cultivation areas.

4.5. Comparison of Spatial Pattern Recognition Performance Between CUWFM and Mainstream Fusion Models

To comprehensively evaluate the performance advantages of the CUWFM, this study selected three representative mainstream spatiotemporal fusion algorithms for comparative analysis: the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM), and the function fitting-based spatiotemporal fusion method (Fit-Fc). The comparative experiment selected 4 July 2020 as the representative date as it falls during the crop growing season when ET rates are relatively high, making it an ideal time point for validating the spatial pattern recognition capabilities of different fusion algorithms.

Figure 12 intuitively demonstrates the significant differences among different fusion methods in spatial pattern recognition. From a visual perspective, the CUWFM-ET fusion results performed best in maintaining spatial details, successfully preserving the fine spatial structural characteristics of the original PT-JPL-ET data. The fused ET distribution clearly showed the boundary contours between different crop fields, effectively captured spatial heterogeneity variations within fields, and exhibited high consistency with the reference PT-JPL-ET data in spatial patterns. In comparison, while ESTARFM-ET could maintain spatial structure to some extent, it exhibited spatial smoothing effects when processing complex agricultural landscapes, with some reduction in the sharpness of field boundaries. The STARFM-ET fusion results showed more severe spatial detail loss, with ET differences between fields being smoothed, potentially making it difficult to accurately distinguish spatial variations in farmland ET under different management practices. The Fit-Fc method fusion results exhibited significant spatial discontinuity, with obvious fusion artifacts appearing in certain areas, significantly affecting the authenticity of spatial patterns.

Figure 12. Comparison of PML-V2-ET, PT-JPL-ET (reference ET), CUWFM-ET, Fit-Fc-ET, ESTARFM-ET, and STARFM-ET on 4 July 2020.

Quantitative analysis was conducted using the All-round Performance Assessment (APA) accuracy evaluation system proposed by Zhu et al. [66] to comprehensively evaluate the fusion accuracy of different methods (Table 2 and Table 3). Since ET is a geophysical parameter rather than a surface reflectance band, we adapted the interpretation of the APA indicators to focus on the accuracy of the physical quantity: root mean square error (RMSE) represents the accuracy of ET magnitude estimation, average difference (AD) indicates the systematic bias of the physical quantity, edge similarity reflects boundary preservation, and Local Binary Pattern (LBP) represents texture similarity. This comprehensive evaluation system enables systematic comparison of fusion algorithms across multiple performance dimensions.

Table 2. Accuracy evaluation of ET fusion results from different spatiotemporal fusion methods in partial agricultural crop cultivation areas of Changji Prefecture.

Method	AD	RMSE	Edge	LBP
ESTARFM	−0.15986	0.237008	−0.07719	−0.00799
STARFM	−0.31261	0.381533	−0.18695	0.005895
Fit-Fc	−0.48312	1.172498	−0.09189	−0.05638
CUWFM	−0.05064	0.174117	−0.07156	−0.00603

Bold values indicate the best performance among the compared methods.

Table 3. APA accuracy evaluation system.

Indicator	Measure	Range	Optimal Value	Positive Value	Negative Value
RMSE	Pixel spectral accuracy	[0, 1]	0
AD	Global spectral accuracy	[−1, 1]	0	Overestimation	Underestimation
Edge	Edge similarity	[−1, 1]	0	Over-sharpening	Over-smoothing
LBP	Texture similarity	[−1, 1]	0	Over-sharpening	Over-smoothing

The results further confirmed the CUWFM’s significant advantages in reconstructing accurate ET physical quantities across all four APA indicators. In terms of ET magnitude accuracy, the CUWFM-ET achieved an RMSE value of 0.174 mm/day, significantly outperforming ESTARFM-ET (0.237 mm/day), STARFM-ET (0.382 mm/day), and Fit-Fc-ET (1.172 mm/day). This result indicates that the CUWFM achieved the highest precision in reproducing the daily water consumption values, with minimal numerical distortion introduced by the fusion process. The systematic bias indicator AD reflects the overall bias level of fusion results relative to reference data. The CUWFM-ET achieved an AD value of −0.051 mm/day, closest to the ideal value of 0, indicating almost no systematic bias at the global scale. In contrast, ESTARFM-ET (−0.160 mm/day), STARFM-ET (−0.313 mm/day), and Fit-Fc-ET (−0.483 mm/day) all exhibited varying degrees of underestimation trends, with the Fit-Fc method showing the most severe global bias. The edge similarity indicator Edge evaluates the fusion results’ capability in maintaining spatial boundary details. The CUWFM-ET achieved an Edge value of −0.072, closest to the optimal value of 0, indicating that the model was the best performer in maintaining field boundary sharpness and avoiding both excessive smoothing and over-sharpening. ESTARFM-ET (−0.077) was close to the CUWFM, but STARFM-ET (−0.187) showed obvious edge over-smoothing phenomena. The texture similarity indicator LBP assesses fusion result performance in maintaining spatial texture characteristics. The CUWFM-ET achieved an LBP value of −0.006, almost reaching the ideal state, indicating that the spatial texture features after fusion were highly consistent with the original high-resolution data. The LBP values of the other three methods all deviated significantly from the ideal value, particularly the Fit-Fc method (−0.056), which showed obvious deficiencies in texture preservation.

In conclusion, regarding the comparison of fused physical quantities, the CUWFM demonstrated superior performance with an accuracy improvement of 15–65% compared to mainstream methods. The low RMSE (0.17 mm/day) and minimal bias suggest that the CUWFM can provide reliable quantitative data for precise irrigation management, whereas other methods (especially STARFM and Fit-Fc) may introduce significant errors in water flux estimation due to smoothing effects or artifacts. This advantage is mainly attributed to (1) the similar pixel screening strategy based on crop ET characteristics better aligns with the physical mechanisms of ET in agricultural areas; (2) the spatial downscaling through unmixing operations effectively reduces uncertainties from mixed pixels; and (3) the temporal prediction considering crop phenology enhances the model’s adaptability to dynamic changes in agricultural landscapes.

4.6. Validation of Continuous Monitoring Capability During Crop Growing Seasons

To further validate the practical application value of the CUWFM-ET in agricultural monitoring, this study extracted daily ET time series for winter wheat, maize, and cotton cultivation areas based on the 2020 Changji Prefecture crop growing season fusion results (Figure 13, Figure 14 and Figure 15), analyzing the seasonal variation characteristics of ET for different crop types and their coupling relationships with key phenological periods.

Figure 13. Daily ET time series for winter wheat cultivation areas.

Figure 14. Daily ET time series for maize cultivation areas.

Figure 15. Daily ET time series for cotton cultivation areas.

The CUWFM-ET successfully captured the dynamic variation patterns of ET throughout the winter wheat growing season (Figure 13). The time series clearly identifies several key stages of winter wheat ET: during the green-up period (late March to early April), ET began to slowly rise to 2–3 mm/day with temperature recovery and vegetation growth resumption; during the rapid growth period (mid-April to mid-May), ET rapidly increased from 3 mm/day to 4 mm/day, corresponding to the vigorous growth stage from the jointing to the booting of winter wheat; and during the peak period (late May to early June), ET reached the season’s maximum at approximately 5 mm/day, which highly consistent with the maximum water demand during the heading and flowering period. Notably, winter wheat ET experienced a sharp decline in early June. This phenomenon was caused by the use of average ET values from winter wheat cultivation areas across the entire Changji Prefecture, where different regions had varying harvest times, and early harvesting activities in early-maturing areas led to a temporary decrease in regional average ET. Subsequently, ET showed recovery until early July when most areas began harvesting, and ET finally exhibited a stable declining trend.

The ET time series for maize cultivation areas exhibited typical “single-peak” distribution characteristics (Figure 14). During the seedling stage (late May to June), ET was relatively low at approximately 2–3 mm/day; during the jointing stage (mid-June to early July), it began to rise rapidly, reaching peak values of 5–6 mm/day; during the heading and grain filling stage (late July to August), ET also maintained high values; and during the maturation stage (September), ET rapidly declined to 1–2 mm/day.

The ET time series for cotton cultivation areas revealed the water consumption dynamics during cotton growing seasons in this region (Figure 15). The time series indicated that during the early growth period (late May to early June), ET was relatively low at approximately 2–3 mm/day, mainly corresponding to the cotton seedling stage when plants were small and water demand was primarily from soil evaporation; during the first high-value period (mid-June to early July), ET rose to 5–6 mm/day, coinciding with cotton budding and flowering when vegetation coverage increased and transpiration enhanced; during the relative decline period (mid-July), ET experienced a brief decrease; during the second high-value period (August), ET reached 4–5 mm/day, corresponding to the cotton boll period; and after mid- to late September, ET began declining as cotton gradually matured.

The ET time series for these three major crops fully demonstrate the CUWFM-ET’s excellent performance in agricultural continuous monitoring. The model not only accurately captured ET magnitude changes for different crops but, more importantly, precisely identified the unique seasonal variation patterns and key phenological transition nodes for each crop, providing data support for precision crop water management.

5. Discussion

5.1. Improved PT-JPL-ET Outperforms Other ET Products

This study conducted comparative analysis between the improved PT-JPL inversion results and two widely used ET products, PML-V2 and MOD16, at the Daman station (Figure 16). Taylor diagram analysis results demonstrate that the improved PT-JPL-ET performed optimally across three dimensions: correlation coefficient (0.88), standard deviation (1.72), and RMSE (approximately 1.0), significantly outperforming both PML-V2 (correlation coefficient 0.83, RMSE approximately 1.49) and MOD16 (correlation coefficient 0.78, RMSE approximately 1.73).

Figure 16. Comparison of PT-JPL-ET with other ET products (PML-V2 and MOD16). The radial blue dotted lines represent the correlation coefficient; the concentric green dashed arcs represent the root mean square error (RMSE); and the black dash-dot arcs represent the standard deviation.

The superior performance of the Improved PT-JPL-ET can be attributed to two distinct factors: the algorithmic improvement in physical mechanisms and the advantage of finer spatial resolution.

First, regarding the algorithmic factor, the direct comparison in Section 4.1 (between the original and improved PT-JPL models at the same 30 m resolution) has already isolated and confirmed the contribution of the LSWI strategy. By replacing RH with LSWI to redefine $f_{sm}$, the model successfully corrected the underestimation bias. This proves that the accuracy gain is physically rooted in the better representation of soil moisture constraints and the full utilization of near-infrared and shortwave infrared band sensitivities, independent of spatial scale.

Second, regarding the resolution factor, the spatial mismatch between coarse-resolution products and ground observations significantly impacts validation accuracy. The PML-V2 (500 m) and MOD16 (500 m/1 km) products inevitably suffer from the “mixed pixel effect,” where a single pixel represents a mixture of crops, bare soil, and roads. In contrast, the 30 m resolution of our product matches the flux tower footprint much better, allowing it to capture the specific evapotranspiration characteristics of the target crop fields more purely. Therefore, the higher accuracy observed in this study is a synergistic result of both the improved physical parameterization and the reduction in spatial heterogeneity bias enabled by high-resolution data.

In comparison, although PML-V2 adopts a relatively advanced two-leaf model structure, it still relies on traditional resistance parameterization schemes in soil evaporation estimation for arid regions, which may have certain limitations when dealing with water stress conditions. Similarly, the MOD16 product is constrained not only by its simplified parameterization schemes but also heavily by its coarse spatial resolution in heterogeneous agricultural landscapes. Furthermore, MOD16 products commonly exhibit underestimation phenomena in China’s arid and semi-arid regions, which is also related to uncertainties in meteorological driving data and insufficient regional adaptability of vegetation parameters.

5.2. Advantages of CUWFM-ET Daily 30 m Spatial Resolution ET Data

The CUWFM employs a hybrid fusion strategy combining crop fraction decomposition with similarity-based weighting to generate daily 30 m spatial resolution ET data. This approach addresses key limitations of existing methods through three main improvements: (1) spatial downscaling of PML-V2 data using ET influence factors reduces mixed pixel uncertainties; (2) crop phenological constraints ensure physically meaningful base date selection; and (3) hierarchical screening based on crop types improves similar pixel identification accuracy.

Unlike traditional algorithms (STARFM or ESTARFM) that rely primarily on spectral similarity and spatial distance, the CUWFM integrates ET-specific characteristics, including NIR and SWIR spectral indices that directly relate to vegetation and soil moisture conditions. This design enhances the model’s responsiveness to agricultural ET processes and improves adaptability for complex agricultural landscapes.

The resulting daily 30 m resolution data meets field-scale irrigation requirements, providing sufficient spatial detail to distinguish management-induced variations while maintaining temporal continuity for monitoring crop water dynamics. This spatiotemporal resolution supports precision agriculture applications and refined water resource management in arid and semi-arid regions.

5.3. Uncertainties and Future Improvements

(1)

Model parameterization aspects: While the strategy of using LSWI to replace RH in the improved PT-JPL model performed excellently under arid conditions, this empirical relationship may exhibit saturation effects under extremely arid or waterlogged conditions, affecting the model’s universality. Additionally, the ET influence factor in the CUWFM, based on a linear combination of NDVI and LSWI, may not comprehensively reflect complex physical ET processes.

(2)

Input data quality: The 0.25° spatial resolution of ERA5 reanalysis data has limited accuracy at local scales; although processed through resampling, it cannot increase original spatial information. Uncertainties in PML-V2 products are directly propagated to fusion results, potentially causing biases in certain regions and seasons.

(3)

Validation sample representativeness: Current validation mainly relies on observations from two sites with relatively limited spatial coverage; energy closure issues in eddy covariance observations and scale matching differences may affect the reliability of validation results.

Future improvement directions: (1) Deepening physical mechanisms—further refine the physical parameterization schemes of the PT-JPL model, particularly dynamic optimization of plant physiological parameters, improving model adaptability under different environmental conditions; (2) multi-source data fusion—integrate optical, microwave, thermal infrared, and other multi-source remote sensing data, utilizing complementary advantages of different sensors to enhance ET estimation accuracy; and (3) validation system improvement—expand the number and spatial distribution of validation sites, adding validation cases for different crop types.

6. Conclusions

This study develops and validates a practical workflow for generating daily 30 m evapotranspiration (ET) data products in arid and semi-arid agricultural regions. Two key innovations underpin the framework: (1) an improved PT-JPL model in which the soil evaporation constraint is redefined using the Land Surface Water Index (LSWI) instead of relative humidity, and (2) a Crop Unmixing and Weight Fusion Model (CUWFM) that integrates crop fraction decomposition, phenology-aware base date selection, hierarchical similar-pixel screening, and similarity-weighted temporal prediction.

Validation against flux tower and lysimeter observations shows that the improved PT-JPL model achieves R² values of ~0.78–0.80, NSE of ~0.69–0.77, and an RMSE of 0.74–1.06 mm day⁻¹, outperforming both PML-V2 and MOD16. Crucially, the CUWFM approach was further validated against independent ground observations, confirming its capability to generate reliable daily ET time series with an R² of 0.770 and an RMSE of 0.716 mm day⁻¹. In spatiotemporal reconstruction assessments, the model demonstrates superior performance across accuracy, spatial detail, and texture preservation (APA metrics: RMSE = 0.174, AD = −0.05, Edge = −0.07, LBP = −0.01) compared to STARFM, ESTARFM, and Fit-Fc.

At the crop level, the fused results capture distinct ET dynamics of the major crops in Changji: winter wheat—regional ET rises sharply in spring, peaks at ~5 mm day⁻¹ during heading–flowering (late May to early June), and declines rapidly after harvest; maize—ET exhibits a single broad summer peak of ~5–6 mm day⁻¹ during the tasseling–grain filling stage (July–August), with stable interannual patterns across 2020–2023; and cotton—ET shows two pronounced high-value periods, in early and mid-late summer, with peak values of ~5–6 mm day⁻¹ corresponding to flowering and boll development stages, consistent with crop water-use physiology.

These findings demonstrate that the combined model–fusion framework not only improves ET estimation accuracy but also preserves field-scale crop water-use signals, enabling more precise irrigation scheduling and water-resource management in arid farmland.