Utilizing Multi-Source Remote Sensing Data and the CGAN to Identify Key Drought Factors Influencing Maize Across Distinct Phenological Stages

Highlights

What are the main findings?

• The improved CGAN model achieved high-precision fitting of maize drought severity (R² > 0.96) under limited sample conditions in Northwest China’s rain-fed agricultural region, demonstrating superior capability in capturing complex nonlinear relationships among multi-source remote sensing factors.
• The dominant drivers of maize drought exhibit a clear dynamic evolution across phenological stages: ET dominates at the seedling stage, ET and SIF_min co-dominate at the jointing–tasseling stage, and temperature (T_mean) becomes absolutely dominant at the maturity stage.

What are the implications of the main findings?

• The integrated application of CGAN and SHAP provides a feasible framework for deep learning modeling and interpretable attribution using small-sample remote sensing data, expanding the methodological options for quantitative agricultural drought research.
• Solar-induced chlorophyll fluorescence (SIF), particularly its minimum value, emerges as a superior indicator of crop physiological drought stress during the mid-to-late growth stages. This provides empirical evidence for integrating crop physiological signals into drought monitoring frameworks and has direct implications for formulating stage-specific drought management strategies.

Abstract

Drought is one of the major disasters constraining crop production. The accurate identification of the dominant environmental factors that drive drought stress at different growth stages of maize is essential for developing stage-specific and precise water management strategies, enhancing drought resistance, and ensuring food security. However, a key challenge is quantifying the nonlinear interactions among multiple environmental factors. This study focuses on the rain-fed agricultural region of Northwest China. To address the limited availability of drought event samples in this region and the inadequacy of traditional statistical methods in capturing complex inter-factor relationships, we integrate a small-sample modeling framework based on an improved Conditional Generative Adversarial Network (CGAN) with an attribution framework that employs SHapley Additive exPlanations (SHAP) for interpretability analysis. We incorporate ten environmental factors derived from multi-source remote sensing: temperature (T_max, T_min, T_mean), precipitation (P), evapotranspiration (ET), soil moisture at 0–10 cm (SM_0–10) and at 10–40 cm (SM_10–40), and solar-induced chlorophyll fluorescence (SIF_max, SIF_min, SIF_mean). Sample sets were established for different maize phenological stages. The CGAN model was employed to achieve high-precision estimation of maize drought severity levels, while the SHAP method was used to quantitatively analyze the dominant factors and their contributions at each phenological stage. The results show that the CGAN model achieved coefficients of determination (R²) of 0.963, 0.972, and 0.979 for the seedling, jointing–tasseling, and maturity stages, respectively, demonstrating excellent nonlinear modeling capability under small samples. SHAP analysis reveals a clear dynamic evolution of dominant factors across phenological stages. Evapotranspiration (ET) dominated in the seedling stage, reflecting the primary role of surface water–heat balance, while the jointing–tasseling stage transitioned to a co-dominance of ET, topsoil moisture (SM_0–10), and minimum SIF, indicating intensified crop transpiration and physiological stress under the meteorological drought framework, and the maturity stage shifted to an absolute dominance centered on mean temperature (T_mean), highlighting the critical impact of heat stress. This study provides a data-driven quantitative perspective for understanding maize drought mechanisms and offers a scientific basis for formulating differentiated drought management strategies for different growth stages. Furthermore, it demonstrates the potential of integrating CGAN with SHAP for agricultural remote sensing and drought attribution research in data-scarce regions.

1. Introduction

Drought is a complex environmental abiotic stressor formed by the interaction of climatic, hydrological, soil, and anthropogenic factors, posing a persistent threat to agricultural production and socio-economic stability [1,2]. Northwest China is located in the deep inland, with water scarcity and low and highly variable precipitation, making it a typical arid and semi-arid region [3,4]. In this region, rain-fed agriculture is the predominant practice. Maize, as one of the primary cultivated crops, has yield stability closely tied to regional food security and farmers’ livelihoods.

Maize exhibits significant differences in water demand and response mechanisms across its growth stages [5,6]. During the seedling stage, maize has a small leaf area and low transpiration rate, with soil evaporation dominating water loss. Drought stress primarily affects plant establishment and uniformity. In the jointing–tasseling stage, rapid vegetative growth coincides with reproductive organ differentiation, making it the most water-sensitive period; drought at this stage can severely impact kernel number and yield potential. During the maturity stage, water demand gradually decreases, and high temperature stress during grain filling can accelerate leaf senescence and reduce kernel weight [7,8,9,10]. Therefore, elucidating the key driving factors and their dynamic weights in drought formation during each phenological stage is essential for achieving precise water management. However, current research primarily focuses on the impacts of a single or limited number of factors, often lacking quantitative attribution regarding the nonlinear interactions of multiple factors [11,12]. This challenge is especially pronounced in regions such as the Northwest rain-fed area, where drought event samples are relatively scarce and spatial heterogeneity is significant. Traditional statistical models often struggle to fully capture these complex relationships.

In recent years, machine learning methods have provided new tools for simulating and predicting complex agricultural systems [13,14,15]. Among these, tree-based ensemble models such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost) have been widely applied in drought monitoring and crop yield prediction due to their robustness and ability to handle nonlinear relationships [16,17,18,19]. Deep learning architectures such as Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) have also demonstrated advantages in capturing temporal and spatial dependencies in remote sensing data [20,21]. However, these models typically require large training samples and often struggle with generalization in data-scarce regions. Furthermore, their “black-box” nature limits mechanistic interpretation, hindering the identification of dominant drought drivers.

Generative Adversarial Networks (GANs) offer a unique advantage in capturing high-dimensional nonlinear relationships through adversarial training [22,23]. By incorporating conditional variables, Conditional Generative Adversarial Networks (CGANs) are particularly well-suited for supervised prediction tasks [24,25,26,27]. The study area faces the challenge of limited drought records, representing a typical small-sample learning scenario. CGAN can enhance the model’s ability to learn data distributions through adversarial training, making it well-suited for learning complex mapping relationships from limited observations. However, deep models like CGANs are often considered “black boxes” with limited interpretability, restricting their application in mechanistic research [28].

To address this interpretability bottleneck, SHapley Additive exPlanations (SHAP), a unified interpretability framework based on cooperative game theory, quantifies the marginal contribution of each feature to the model output [29,30,31]. SHAP provides consistent and locally accurate feature attributions, enabling transparent interpretation of complex model predictions. The combination of predictive models with SHAP has been successfully applied in various environmental and agricultural studies for identifying key drivers [32]. Nevertheless, few studies have integrated CGAN with SHAP for drought factor attribution, particularly in small-sample agricultural settings.

Furthermore, advances in remote sensing technology provide multi-dimensional and continuous data support for agricultural drought monitoring [33,34,35]. The temperature, precipitation, and evapotranspiration data used in this study have kilometer-level spatial resolution, adequately reflecting regional environmental background. Soil moisture data characterize soil water status. Additionally, solar-induced chlorophyll fluorescence (SIF) data are directly linked to crop photosynthetic activity and respond to water stress more directly and sensitively than traditional vegetation indices [36,37,38]. Recent studies have demonstrated the potential of SIF for detecting crop physiological stress [39,40], highlighting its value in advancing drought monitoring capabilities [41,42].

Despite these advances, several research gaps remain: (1) Most existing studies focus on single-factor analysis or fail to capture nonlinear interactions among multiple environmental factors; (2) In data-scarce regions such as Northwest China, conventional machine learning and deep learning models often suffer from overfitting or poor generalization; (3) The dynamic evolution of dominant drought drivers across maize phenological stages remains poorly quantified; (4) The integration of advanced generative models with interpretability frameworks for agricultural drought attribution has yet to be systematically explored.

Accordingly, this study focuses on the rain-fed agricultural region of Northwest China and constructs an integrated analysis framework of “Improved CGAN + SHAP” (Figure 1). This study integrates multi-source remote sensing data along with historical drought records, with the following objectives: (1) Validate the high-precision fitting capability of the Conditional Generative Adversarial Network (CGAN) for assessing maize drought severity under limited sample conditions. (2) Systematically identify and quantify the dominant environmental factors affecting maize drought across different phenological stages using the SHAP method, and reveal their dynamic evolution patterns. (3) Propose targeted drought management recommendations for various growth stages based on the identified factor importance, providing a theoretical basis for regional grain production security.

2. Materials and Methods

2.1. Study Area Overview

The rain-fed agricultural region of Northwest China (103°47′01″–111°25′19″E, 32°59′41″–39°59′08″N) is located deep inland, covering 65 counties across the Longdong and Longzhong regions of Gansu, the Northern Shaanxi Plateau, and the mountainous area of southern Ningxia (Figure 2). As a typical drought-prone area, this region heavily depends on natural rainfall for maize cultivation, making it highly vulnerable to drought stress.

The study area features complex topography with significant elevation gradients—mountain ranges such as Liupan and Ziwuling create an interlaced landscape of mountains, hills, and tablelands. Elevation drives the redistribution of hydrothermal conditions: higher altitudes receive more precipitation but have shorter growing seasons, while low-altitude valleys experience ample heat but intense evaporation. This topography-induced hydrothermal differentiation leads to varying drought formation mechanisms across elevations, providing essential topographic context for spatial analysis of multi-source remote sensing data.

2.2. Research Data

Temperature, precipitation, and evapotranspiration data were obtained from the National Tibetan Plateau Data Center (http://www.tpdc.ac.cn (accessed on 23 June 2025)), each with a spatial resolution of 1 km [43,44,45,46]. Soil moisture data (0–10 cm and 10–40 cm layers) were acquired from NASA’s Earth Science Data website, with a spatial resolution of 10 km. Solar-induced chlorophyll fluorescence data [47] were obtained from the GOSIF (Global, OCO-2-based SIF product) dataset, with a spatial resolution of 0.05° (approximately 5 km) and a temporal coverage from 2000 to 2024. This dataset is a global high-resolution SIF product reconstructed using a data-driven approach that integrates discrete fluorescence observations from the OCO-2 satellite with MODIS remote sensing data and meteorological reanalysis data. Historical maize drought records (2001–2011) at the county scale were compiled from the National Meteorological Science Data Center. The Digital Elevation Model (DEM) was sourced from the Shuttle Radar Topography Mission (SRTM) product, with a spatial resolution of 30 m.

All datasets were resampled to a uniform 1 km resolution using bilinear interpolation for continuous variables and nearest-neighbor interpolation for categorical variables. Raster data were then clipped to the county administrative boundaries of the study area, and the mean value of each environmental factor was calculated per county to align with the county-scale drought records. Prior to aggregation, rigorous quality control was applied to all remote sensing datasets to mask out poor-quality pixels. For soil moisture data, pixels affected by radio frequency interference, frozen ground, or snow cover were excluded based on the quality assurance flags provided with the dataset. For SIF data, only high-quality retrievals were retained according to the quality control recommendations of the GOSIF product. Temperature, precipitation, and evapotranspiration datasets were already subjected to preliminary quality control by the data providers and were used directly.

Aggregating raster data to the county level means inevitably losing spatial variability information within each county, which smooths local heterogeneity from topography, land use, and soil properties, potentially masking fine-scale drought response patterns at the field level. This trade-off was necessary to align multi-source remote sensing data (with resolutions of 1–10 km) with the available county-scale drought records. Future studies could employ grid-based modeling approaches with higher-resolution data to capture sub-county spatial heterogeneity.

In this region, maize cultivation follows a single-cropping (spring maize) system. Based on long-term local agrometeorological observations and farming records, we delineated three key phenological stages: the seedling stage (sowing to pre-jointing, April–May), the jointing–tasseling stage (concurrent vegetative and reproductive growth, June–August), and the maturity stage (grain filling to maturity, September–October). All data were subsequently partitioned according to these three stages. Consequently, each sample comprised ten environmental factors (Tmax, Tmin, Tmean, P, ET, SM_0–10, SM_10–40, SIF_max, SIF_min, SIF_mean) for a given county, year, and phenological stage, along with the corresponding drought severity label. Land use data were reclassified into six major categories following the classification system of the Chinese Academy of Sciences. The Digital Elevation Model (DEM) was processed in ArcGIS 10.6 to generate a topographic map. County-level historical maize drought disaster records (2001–2011) used in this study were obtained from the National Meteorological Science Data Center, documenting the severity levels of drought events in each year (including mild, moderate, severe, and extreme drought). As these levels are categorical variables and cannot be directly used as inputs or outputs for deep learning models, numerical encoding was required. Therefore, we referenced the standardized precipitation index (SPI) classification thresholds specified in the Chinese national standard “GBT20481-2017 Meteorological Drought Grade [48]” (

Table 1. Standardized Precipitation Index (SPI) Drought Classification Table.

Grade	Drought Status	Standardized Precipitation Index (SPI)
I	No Drought	−0.5 < SPI
II	Mild Drought	−1.0 < SPI ≤ −0.5
III	Moderate Drought	−1.5 < SPI ≤ −1.0
IV	Severe Drought	−2.0 < SPI ≤ −1.5
V	Extreme Drought	SPI ≤ −2.0

) to map the drought event levels to corresponding numerical codes. The SPI is used here solely as a reference framework for drought level encoding and is not the training target variable of the model. The actual target for model fitting is the drought event levels recorded in historical documents, and the SPI itself is not used in model input or loss function computation. This encoding approach ensures the ordinality and comparability of drought levels while avoiding potential biases that might arise from subjective assignment.

2.3. Research Methods

The methodology of this study comprises three main components: data normalization, development and training of an improved Conditional Generative Adversarial Network (CGAN) model, and SHAP-based dominant factor attribution analysis.

Data normalization eliminates potential bias caused by differences in units and scales among the multi-source remote sensing data, mapping all environmental factors to a unified scale to ensure training stability. An improved CGAN model achieves high-precision fitting of maize drought severity levels. Through the adversarial training mechanism between the generator and discriminator, the model learns complex nonlinear mapping relationships between environmental factors and drought levels from limited samples, enabling accurate prediction of drought severity. SHAP interprets the trained model. Rooted in cooperative game theory, this method quantifies the marginal contribution of each environmental factor to the model’s predictions, identifying the dominant drought factors across different phenological stages and revealing their dynamic evolution patterns.

2.3.1. Data Normalization

Normalization was applied to all input environmental factors to mitigate potential bias in model training caused by significant variations in their units and scales. Each factor was scaled to the [0, 1] range. The normalization was performed using the following formula:

$$y_i=\left(x_i-x_{\mathit{min}}\right)÷\left(x_{\mathit{max}}-x_{\mathit{min}}\right)$$

(1)

where y_i is the normalized value of influencing factor i at a given location; x_i is the original (observed) value of factor i at that location; and x_max and x_min denote the maximum and minimum values of factor i, respectively.

2.3.2. Model Selection

This study proposes a drought severity level prediction model based on a Conditional Generative Adversarial Network (CGAN) [49,50,51]. The model is designed to precisely quantify the influence of environmental factors on maize drought in the rain-fed agricultural region of Northwest China. The model framework centers on a designed generator and discriminator. A suite of optimization strategies was employed during training to enhance the model’s robustness and predictive accuracy.

In terms of model selection, CGAN offers distinct advantages over other commonly used machine learning approaches. Traditional machine learning models have limited capacity to capture complex nonlinear features and typically require large training datasets, making them prone to underfitting or overfitting under small-sample conditions. In our preliminary experiments, we attempted to train other deep learning models (such as CNNs) as well as baseline tree-based models (such as random forest) as benchmarks; however, their performance was insufficient under the current small-sample condition, exhibiting severe overfitting, unstable convergence, or high variance. In contrast, CGAN leverages an adversarial training mechanism between the generator and discriminator to more effectively learn data distribution characteristics from limited samples, demonstrating superior fitting capability and robustness in small-sample scenarios [52,53], which further supports the selection of CGAN in this study. Accordingly, this study selects CGAN as the core modeling approach, and a schematic diagram of its operation workflow is shown in Figure 3.

The model consists of two sub-networks—a generator and a discriminator—which are jointly optimized through adversarial training. During the training phase, the generator receives two types of input: a random noise vector (to introduce diversity) and a condition vector composed of the 10 environmental factors, including temperature (T_max, T_min, T_mean), precipitation (P), evapotranspiration (ET), soil moisture at 0 to 10 cm (SM_0–10), soil moisture at 10 to 40 cm (SM_10–40), and solar-induced chlorophyll fluorescence (SIF_max, SIF_min, SIF_mean). Based on these inputs, the generator attempts to output predicted drought severity levels, aiming to make the distribution of these predictions closely approximate the distribution of the real drought severity records. The model was trained using the Adam optimizer for 2000 epochs. For each phenological stage, data from 2001 to 2009 were used as the training set (approximately 70%), and data from 2010 to 2011 were used as the test set (approximately 30%). A validation set was further split from the training set for early stopping to prevent overfitting, and the model checkpoint with the lowest validation loss was selected as the final model.

(1)

Generator

The generator is a core component of the CGAN model and is responsible for learning the mapping from environmental features to drought severity levels. It receives two types of input: standardized environmental feature data (10-dimensional) and a randomly generated noise vector (5-dimensional). These inputs are first passed through a hidden layer with 32 neurons, where they undergo batch normalization, LeakyReLU activation, and dropout. The output is then fed into a second hidden layer with 16 neurons, where similar processing is applied. Finally, a standardized drought severity level prediction is generated through an output layer using a Tanh activation function to ensure the output falls within the range [−1, 1]. The objective of the generator is to learn the distribution of real data to produce predictions capable of deceiving the discriminator, while simultaneously ensuring that the predicted values remain close to the true values through a regression loss.

(2)

Discriminator

The discriminator is designed to distinguish between real data pairs (environmental features + actual drought severity levels) and generated data pairs (environmental features + drought severity levels predicted by the generator). It takes a combination of feature data and target values as input, which are first concatenated and then passed through a hidden layer with 32 neurons, where batch normalization, LeakyReLU activation, and dropout are applied. The output is subsequently passed to a second hidden layer with 16 neurons, where similar operations are performed. Finally, an output layer with a Sigmoid activation function generates a probability value constrained to the range [0, 1], representing the likelihood that the input data pair is real. The discriminator aims to accurately identify real and generated data, thereby providing learning signals that drive the generator to iteratively improve its generative capability.

2.3.3. Loss Function

During the training of the Conditional Generative Adversarial Network (CGAN), three loss functions are jointly optimized to balance the authenticity, accuracy, and smoothness of the generated results. The overall objective function consists of an adversarial loss, a regression loss, and an L1 loss.

(1)

Adversarial Loss

Adversarial loss is used to constrain the game process between the generator G and the discriminator D, and is the core of the GAN model [54]. The discriminator aims to distinguish real samples from generated samples, while the generator attempts to produce results that can deceive the discriminator. The objective function can be written as:

$$L_{cGAN}(G,D)={\mathbb{E}}_{x,y}\left[\mathit{log}D\right(x,y\left)\right]+{\mathbb{E}}_{x,z}\left[\mathit{log}\right(1-D\left(G\right(x,z\left)\right)\left)\right]$$

(2)

where x denotes the conditional input, y the ground truth, and z the random noise vector. In practice, the generator is optimized by maximizing log(D(G(x,z))), while the discriminator is updated by minimizing the binary cross-entropy loss. The discriminator loss is obtained by summing the loss on real samples and the loss on generated samples, and the generator’s adversarial loss is computed based on the discriminator’s output for the generated samples.

(2)

Regression Loss

The regression loss directly constrains the numerical error between the generator’s output and the ground truth, ensuring the accuracy of the generated results [55]. In this paper, the Mean Squared Error is adopted as the regression loss, defined as:

$$L_{G_{reg}}={\mathbb{E}}_{(x,y)\sim P_{data},z\sim P_z}[\Vert G(x,z)-y{\Vert }^2]$$

(3)

By minimizing the Euclidean distance between the predicted and target values, this loss encourages the generator to learn a stable and accurate mapping from the conditional input to the desired output.

(3)

L1 Loss

The L1 loss enhances the smoothness of the generated results and reduces extreme fluctuations, making it particularly suitable for data scenarios with noise or outliers [56]. Its form is:

$$L_{G_{L1}}={\mathbb{E}}_{x,y,z}\left[\right|\left|G\right(x,z)-y{\left|\right|}_1]$$

(4)

Compared with the MSE loss, the L1 loss is more robust to outliers, as it penalizes absolute differences linearly rather than quadratically. This property helps improve the model’s generalization capability, especially when the training data contains irregularities.

(4)

Combined Loss and Weight Settings

The total loss of the generator is obtained by the weighted sum of the above three loss terms, specifically:

$$L_G=L_{cGAN}(G,D)+{\lambda }_1·L_{G_{reg}}+{\lambda }_2·L_{G_{L1}}$$

(5)

In the experiments conducted under the small-sample scenario, the weights are set to λ₁ = 10 and λ₂ = 5, with the adversarial loss weight being 1. This configuration aims to strengthen the guiding effect of the regression loss on prediction accuracy while using the L1 loss to enhance smoothness and robustness, thus balancing the generator’s adversarial capability and predictive performance.

2.3.4. Evaluation Metrics

The performance of the CGAN model was evaluated using three metrics [57,58,59,60]: the coefficient of determination (R²), root mean square error (RMSE), and bias. These metrics are defined as:

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$$

(6)

$$\mathit{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$$

(7)

$$\mathit{Bias}={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left(y_i-{\widehat{y}}_i\right)$$

(8)

where y_i is the observed drought level, ŷ_i is the predicted value, ȳ is the mean of observed values, and n is the number of samples.

2.3.5. Dominant Factor Analysis

Following the training of the optimal CGAN generator, the SHAP (SHapley Additive exPlanations) method was employed to interpret the influence of environmental factors during each phenological stage [61,62]. This step quantitatively identifies the dominant factors governing maize drought severity. The SHAP method is rooted in cooperative game theory. Its core principle involves calculating the average marginal contribution of each feature to the model’s prediction across all possible feature combinations. This approach provides a consistent and reliable metric for quantifying feature importance.

For a given sample, the SHAP values decompose the model’s prediction into the additive contributions of individual features. The SHAP value for feature i is computed as follows:

$${\varphi }_i\left(f,x\right)=\sum _{S\subseteq M\backslash \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(M-\left|S\right|-1\right)!}{M!}}\left[f_x\left(S\cup \left\{i\right\}\right)-f_x\left(S\right)\right]$$

(9)

where ϕ_i(f,x) denotes the SHAP value of feature i for an instance x, quantifying the contribution of this feature to the model prediction (f,x); S is a subset of features excluding feature i; M is the total number of features; and fx (S) = E[f(x)|xS] is the expected output of model f given the feature subset S.

To assess the global importance of each feature, the mean absolute SHAP value across all samples is computed:

$$I_j={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|{\varphi }_j^{\left(i\right)}\right|$$

(10)

where I_j is the global importance score of the j-th feature, and the summation is over all n samples indexed by i.

3. Results and Analysis

3.1. Model Training Process and Convergence Analysis

The training dynamics of a Generative Adversarial Network (CGAN) are typically captured by the evolution of its generator and discriminator loss functions. Figure 4 presents the training loss curves of the CGAN model across the three phenological stages.

During the initial training phase, the loss values of both the generator and discriminator exhibit typical adversarial oscillations, characterized by alternating rises and falls in a competitive manner, reflecting a phase of intense adversarial learning. After approximately 2000 epochs, the losses gradually converge to a dynamic equilibrium, marked by reduced oscillation amplitude and enhanced stability. This convergence indicates that the distribution of the drought severity predictions output by the generator closely approximates the distribution of the real data, demonstrating that the model has effectively learned the mapping from environmental factors to drought severity levels. It should be noted that the loss curves presented in Figure 4 have been smoothed using an exponential moving average (decay factor = 0.9) to better illustrate the overall convergence trends; the raw loss curves retain the characteristic adversarial oscillations described above.

3.2. Model Performance Evaluation

The predictive performance of the CGAN model for drought severity levels was quantitatively evaluated on the test set using three metrics: the coefficient of determination (R²), root mean square error (RMSE), and bias. The optimal performance metrics across the three phenological stages are summarized in

Table 2. Performance Metrics of the CGAN Model for Different Maize Phenological Stages.

Phenological Stage	R2	RMSE	Bias
Seedling Stage	0.9634	0.0585	−0.018
Jointing–Tasseling Stage	0.972	0.0522	0.0195
Maturity Stage	0.979	0.051	−0.0042

, and the R² value variation curve across the three phenological stages is shown in Figure 5.

R² values exceeded 0.96 for all stages, peaking at 0.979 for the maturity stages. This indicates that the model explains over 96% of the variance in drought severity, demonstrating its strong capability in capturing the complex nonlinear relationships with the input environmental factors. All RMSE values were below 0.06, indicating minimal deviation between predictions and observations and thus high predictive accuracy. The absolute bias was near zero for all stages (maximum = 0.0195), suggesting the absence of substantial systematic bias in the predictions. From the seedling to the maturity stage, R² consistently increased and RMSE decreased. This trend may indicate either a stabilization in the drought–environment relationship as maize develops, or more effective model learning during the maturity stage.

In order to further evaluate the generalization ability of the model and avoid overfitting, we used the 5-fold cross-validation to test the robustness of the model’s performance. The dataset for each phenological stage was randomly partitioned into five mutually exclusive subsets of approximately equal size. In each fold, four subsets were used for model training and the remaining subset was used for testing, with the process repeated five times such that each subset served as the test set exactly once. The reported metrics represent the mean and standard deviation computed across the five folds. The cross-validation results are shown in

Table 3. 5-fold cross-validation table.

Phenological Stage	R2 (Mean ± SD)	RMSE (Mean ± SD)	Bias (Mean ± SD)
Seedling Stage	0.9714 ± 0.0026	0.0752 ± 0.0022	0.0002 ± 0.0019
Jointing–Tasseling Stage	0.9722 ± 0.0017	0.0740 ± 0.0016	0.0002 ± 0.0013
Maturity Stage	0.9790 ± 0.0014	0.0670 ± 0.0016	0.0000 ± 0.0008

.

It can be seen from Table 3 that the R² of each phenological period is more than 0.97, which is highly consistent with the results of the original test set (0.963–0.979). The standard deviation of each index is small (R² standard deviation < 0.003, RMSE standard deviation < 0.003), indicating that the model has stable prediction performance under different data divisions. The bias values were close to 0, which further confirmed that there was no systematic deviation in the model. These results fully show that the CGAN model has good generalization ability without overfitting problems, and its high-precision fitting reflects the ability of the model to truly capture the complex nonlinear relationship between environmental factors and drought levels.

3.3. Interpretability Analysis

To deeply interpret the complex relationships learned by the CGAN model and identify the dominant factors influencing maize drought severity levels across different phenological stages, this study employed SHAP values (Figure 6) for post hoc model interpretation. SHAP values quantitatively reflect the contribution of each input feature to the model output (drought severity prediction), with the magnitude of their absolute value representing the importance of that factor.

The global importance scores for each factor were calculated and ranked according to Formula (10), resulting in

Table 4. Ranking of Drought Impact Factor Importance Based on SHAP Values for Each Phenological Stage.

Rank	Seedling Stage	Jointing–Tasseling Stage	Maturity Stage
1	ET	ET	T(mean)
2	T(min)	SIF(min)	T(min)
3	T(mean)	SM(0–10cm)	T(max)
4	T(max)	T(max)	SIF(min)
5	SIF(min)	P	ET
6	SIF(max)	T(min)	P
7	SIF(mean)	SIF(mean)	SIF(mean)
8	SM(10–40cm)	SIF(max)	SM(10–40cm)
9	P	SM(10–40cm)	SM(0–10cm)
10	SM(0–10cm)	T(mean)	SIF(max)

.

As shown in Figure 6 and Table 4, evapotranspiration (ET) is the most influential factor during the seedling stage, far exceeding all others. This suggests that surface water loss (evaporation) is the dominant process, with temperature playing a secondary role. In the jointing–tasseling stage, ET remains the primary factor. However, its dominant component likely shifts from soil evaporation (seedling stage) to plant transpiration. Concurrently, the influence of topsoil moisture and the minimum SIF value increases significantly. During the maturity stage, temperature factors occupy an absolutely dominant position. This indicates that heat stress supplants water balance as the paramount factor governing drought severity in the late growth stage.

3.4. Factor Importance and Dynamic Evolution Analysis

The importance of features influencing drought severity across different maize phenological stages is illustrated in Figure 7.

During the seedling stage, evapotranspiration (ET) emerged as an overwhelmingly dominant factor (Figure 7a). This indicates that surface water loss, driven primarily by physical evaporation from bare soil, is the core drought-driving process at this stage. The small plant size, open canopy, and large exposed soil area favor direct soil evaporation over plant transpiration, leading to substantial non-productive water loss. Temperature indicators also ranked highly, particularly minimum temperature, suggesting potential compound stress from concurrent low temperature and drought conditions.

Entering the jointing–tasseling stage, the importance of features that affect drought severity underwent significant changes, as shown in Figure 7b. While ET remained the most important factor, its primary driver likely shifted from soil evaporation to crop transpiration. Meanwhile, the importance of SM_0–10 rises sharply to the third position, reflecting that during this stage, with well-developed root systems, maize absorbs substantial water primarily from the topsoil to meet rapid growth demands. Crop dependence on direct topsoil water availability peaked. Notably, SIF_min rose to the second position, with its importance even surpassing soil water supply. SIF is tightly coupled with photosynthesis, and its minimum value may characterize the extreme state where crop photosynthetic activity is suppressed under acute water stress.

In the maturity stage, the dominant drought mechanism underwent a fundamental shift, as shown in Figure 7c. The importance of temperature indicators increased by an order of magnitude. Mean temperature (T_mean) became overwhelmingly dominant, with an importance score more than tenfold greater than in earlier stages. This signifies that in the late growth stage, heat stress supplants water balance as the most critical factor affecting drought severity (likely more associated with final yield formation). In stark contrast, the importance of ET and soil moisture declined markedly. SIF_min still maintains relatively high importance, possibly reflecting leaf senescence, premature termination of photosynthetic function, and other senescence processes related to high temperature and drought.

4. Discussion

This study developed a CGAN-SHAP integrated analytical framework at the county scale in the rain-fed agricultural region of Northwest China, based on multi-source remote sensing data and small-sample drought records. The framework achieved high-precision estimation of maize drought severity levels (R² > 0.96) and quantitative identification of dominant factors. Overall, the framework demonstrates excellent nonlinear modeling capability under small-sample conditions, effectively addresses the interpretability bottleneck of deep learning models, and reveals the stage-dependent evolution of maize drought formation mechanisms. The following discussion is structured into three parts: research findings, data limitations, and future perspectives.

4.1. Discussion of Research Findings

The CGAN model achieved high fitting accuracy (R² ranging from 0.963 to 0.979) across all phenological stages. While conventional machine learning and deep learning models—such as Linear Regression, Random Forest, XGBoost, LSTM, and CNN—are typically designed for large-sample training and may suffer from overfitting or poor generalization when data are limited, the CGAN employed in this study demonstrates a distinct advantage in handling small-sample conditions. Through its adversarial training mechanism, the model effectively learns complex nonlinear mapping relationships between environmental factors and drought severity levels from limited drought event samples, providing a viable pathway for agricultural remote sensing drought research in data-scarce regions. Five-fold cross-validation results further demonstrate that the model achieved mean R² values exceeding 0.97 across all phenological stages with standard deviations below 0.003, confirming strong generalization capability and absence of overfitting.

SHAP attribution analysis further reveals the stage-dependent evolution of maize drought formation mechanisms. The seedling stage is dominated by evapotranspiration, reflecting the core role of surface water–heat balance, with soil evaporation serving as the primary water dissipation pathway. Temperature factors, particularly minimum temperature, also exhibit high importance, potentially indicating compound stress from concurrent low temperature and drought conditions. The jointing–tasseling stage transitions to co-dominance by evapotranspiration, topsoil moisture, and minimum SIF, indicating enhanced crop transpiration and a highly coupled relationship between water stress and photosynthetic inhibition, with crop dependence on topsoil water availability reaching its peak. The maturity stage shifts to absolute dominance centered on mean temperature, with its importance score more than tenfold higher than in earlier stages, highlighting the critical impact of heat stress on drought severity during the late growth stage, while the importance of evapotranspiration and soil moisture declines markedly.

Notably, the importance of minimum SIF surpasses that of most traditional environmental factors during the jointing–tasseling stage, even exceeding soil moisture supply, indicating its unique capacity to capture crop physiological response information not fully reflected by traditional environmental indicators. This demonstrates its distinctive advantages in identifying crop stress conditions. However, it should be emphasized that this analysis was conducted within a meteorological drought framework (based on SPI encoding), and the importance of SIF reflects its correlation with meteorological drought levels rather than serving as a direct measure of crop physiological drought. Therefore, the findings of this study provide a preliminary foundation and empirical evidence for the future development of drought monitoring indicator systems that integrate environmental factors with crop physiological response signals.

4.2. Discussion of Data Limitations

The multi-source remote sensing data employed in this study exhibit several uncertainties regarding spatial and temporal resolution, data types, and product attributes, which warrant careful consideration when interpreting the results.

First, the original resolutions of the data vary considerably, ranging from 1 km to 10 km. Resampling all datasets to a uniform 1 km resolution may introduce interpolation errors, particularly for soil moisture data at 10 km resolution, where upscaling inevitably results in the loss of sub-pixel spatial variability. The SIF data, with an original resolution of 0.05° (approximately 5 km), face similar issues. Notably, this dataset is a global high-resolution SIF product reconstructed using a data-driven approach that combines discrete OCO-2 satellite fluorescence observations with MODIS remote sensing data and meteorological reanalysis data. While this fusion method enhances spatial continuity and temporal coverage, it also introduces additional uncertainties arising from the empirical relationships among heterogeneous data sources, the spatial extrapolation of sparse OCO-2 soundings, and the inherent errors in the ancillary datasets used for reconstruction. Quantifying these uncertainties is challenging but essential, as they may propagate through subsequent aggregation steps and affect the robustness of drought response analyses, particularly when downscaling or resampling to finer spatial resolutions. Aggregating raster data to the county level means further smoothing out local heterogeneity caused by differences in topography, land use, and soil properties, potentially masking fine-scale drought response patterns at the field level. This trade-off was necessitated by the need to align multi-source remote sensing data with available county-scale drought records.

Second, both the soil moisture and evapotranspiration data used in this study are model-derived products rather than direct observations. The soil moisture data were obtained from the NASA Earth Science Data website, covering the period 2001–2024 with a spatial resolution of 0.1° and units of m³·m⁻³. The evapotranspiration data were sourced from the National Tibetan Plateau Data Center, spanning 2001–2023 with a spatial resolution of 1 km and units of mm. The generation of these two datasets heavily relies on meteorological forcing fields, particularly temperature and radiation, with their underlying models typically employing temperature, net radiation, and vapor pressure deficit as core input variables. Consequently, when evapotranspiration or soil moisture is identified as the dominant factor in SHAP attribution analysis, the actual contribution of meteorological factors such as temperature may be “indirectly carried” by these derived variables, leading to a systematic underestimation of temperature’s direct importance. This phenomenon is particularly pronounced during the seedling stage (dominated by evapotranspiration) and the jointing–tasseling stage (co-dominated by evapotranspiration and topsoil moisture). Therefore, when interpreting the feature importance rankings, it is essential to recognize the indirect effects introduced by the “derived nature” of the data products and to avoid interpreting the direct contributions of each factor in isolation.

Third, this study focuses on non-irrigated areas and does not fully account for the modulating effects of anthropogenic factors such as irrigation. In actual agricultural production, irrigation can significantly alleviate water stress and alter crop response patterns to meteorological drought, which limits the framework’s applicability in regions with intensive human intervention. Additionally, the sample size becomes limited after phenological partitioning (65 counties × 11 years × 3 stages), which, while suitable for the CGAN framework’s small-sample characteristics, may affect the robustness of some statistical analyses.

4.3. Future Research Perspectives

To address the above limitations, future research can be extended in the following directions. First, data integration and scale optimization. Higher-resolution remote sensing data (e.g., Sentinel, Landsat, with spatial resolutions of 10–30 m) can be integrated with grid-based modeling approaches to validate findings at finer spatial scales and capture sub-county heterogeneity in topography and land use. Additionally, incorporating in situ observational data (such as field-scale soil moisture monitoring stations and eddy covariance flux tower observations) to calibrate model-derived products locally, or adopting multi-source data integration methods, can mitigate the impacts of uncertainties inherent in individual datasets.

Second, mechanistic deepening and indicator expansion. Incorporating irrigation indicators or soil moisture stress indices can better characterize anthropogenic interventions and enhance the framework’s applicability in irrigated agricultural regions. Furthermore, integrating physiological indicators such as crop water deficit indices and direct photosynthesis observations can help further disentangle the direct and indirect roles of various environmental factors in maize drought formation, and provide a deeper assessment of the independent value of SIF in crop physiological drought monitoring, advancing drought monitoring from an “environmentally driven” paradigm toward an integrated “environment–physiology coupled” approach.

Third, model generalization and cross-validation. The generalizability of the framework can be validated across broader geographical regions (such as the Northeast China Plain and North China Plain) and crop types (such as wheat and soybean) to test its robustness under different climatic conditions and cropping systems. Exploring coupling with process-based crop models (such as APSIM and DSSAT) can deepen the mechanistic understanding of how the identified dominant factors influence crop growth and yield formation, thereby providing theoretical support for developing more universally applicable agricultural drought monitoring and attribution systems.

In summary, the CGAN-SHAP framework developed in this study achieves high-precision fitting and interpretable attribution of maize drought in data-scarce regions, revealing the dynamic evolution of dominant drought factors across different maize phenological stages and providing a scientific basis for formulating differentiated stage-specific drought management strategies. Despite limitations related to data resolution, product attributes, and anthropogenic factors, this study offers a new methodological reference for agricultural remote sensing drought research. Through future efforts in multi-source data integration, physiological indicator expansion, and model generalization validation, agricultural drought monitoring is expected to advance from an “environmentally driven” paradigm toward an integrated “environment–physiology coupled” approach.

5. Conclusions

This study developed an attribution framework that integrates an improved Conditional Generative Adversarial Network (CGAN) with SHAP interpretability analysis. The framework was used to investigate the phenological differentiation of drought impact mechanisms on maize in the rain-fed agricultural region of Northwest China. Supported by multi-source remote sensing data, the framework achieved high-precision estimation of maize drought severity levels under small-sample conditions and quantitatively identified dominant driving factors. The main conclusions are as follows:

(1)

Methodological contribution to small-sample deep learning modeling. This study demonstrates that the adversarial training mechanism of CGAN can effectively learn complex nonlinear relationships between environmental factors and drought levels from limited samples, providing a novel technical pathway for agricultural remote sensing modeling in data-scarce regions. Combined with SHAP interpretability analysis, the framework addresses the “black box” problem of deep learning, enabling a closed-loop process from prediction to attribution and offering a reproducible methodological framework for quantitative agricultural drought attribution research.

(2)

Phenology-dependent dynamic evolution of maize drought formation mechanisms. The study reveals a stage-dependent transition in maize drought formation mechanisms: the seedling stage is dominated by surface water–heat balance (ET), which characterizes meteorological drought; the jointing–tasseling stage transitions to a coupled water-physiology dominance (ET, SM(0–10 cm), SIFmin); and the maturity stage shifts to heat stress (Tmean) as the dominant factor. These findings, derived from a data-driven perspective, underscore the complexity of maize drought formation mechanisms and provide a theoretical basis for developing stage-specific drought management strategies.

(3)

Potential of SIF for monitoring crop physiological responses. The study found that the importance of minimum SIF surpassed that of most traditional environmental factors during the mid-to-late growth stages of maize. This result indicates that SIF can capture crop physiological response information not fully reflected by traditional environmental indicators, demonstrating unique advantages for identifying crop stress conditions. It provides a preliminary foundation and empirical evidence for the future development of drought monitoring indicator systems that integrate environmental factors with crop physiological response signals. Subsequent research should further validate the independent value of SIF in crop physiological drought monitoring by incorporating physiological indicators such as crop water deficit indices and direct photosynthesis observations.

Limitations and future directions. This study primarily focused on non-irrigated areas at the county scale, did not fully account for anthropogenic factors such as irrigation, and had limited sample sizes after phenological partitioning. Future research should incorporate irrigation indicators, integrate higher-resolution remote sensing data, validate the framework’s generalizability across broader geographical spaces and crop types, and explore coupling with process-based crop models to deepen the understanding of drought formation mechanisms.