Next Article in Journal
Correction: Yang et al. Perfluorinated and Polyfluoroalkyl Compounds in the Atmosphere: A Review. Atmosphere 2025, 16, 1070
Previous Article in Journal
Intense Rainfall in Urban Areas: Characterization of High-Intensity Storms in the Metropolitan Area of Barcelona (2014–2022)
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Atmosphere
Volume 17
Issue 1
10.3390/atmos17010042
atmosphere-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Modeling of Drought-Induced Crop Yield Loss Based on Solar-Induced Chlorophyll Fluorescence by Machine Learning Methods

by
Han Hu
1,2,
Minxue Zheng
1,2,
Yue Niu
3,
Qiu Shen
1,2,*,
Qinyao Ren
1,2 and
Yanlin You
1,2
1
School of the Environment and Safety Engineering, Jiangsu University, Zhenjiang 212013, China
2
School of the Emergency Management, Jiangsu University, Zhenjiang 212013, China
3
School of Computer, Jiangsu University of Science and Technology, Zhenjiang 212100, China
*
Author to whom correspondence should be addressed.
Atmosphere 2026, 17(1), 42; https://doi.org/10.3390/atmos17010042
Submission received: 20 November 2025 / Revised: 21 December 2025 / Accepted: 27 December 2025 / Published: 28 December 2025
(This article belongs to the Section Biometeorology and Bioclimatology)
Browse Figures
Versions Notes

Abstract

Against the accelerating backdrop of global warming, drought-induced crop yield loss not only causes direct economic losses but may also disrupt the dynamic balance of food production and consumption, ultimately threatening global food security. In order to quantify drought-induced crop yield loss for safeguarding national food security, this study developed a model for evaluating drought-induced yield reduction in winter wheat by integrating solar-induced chlorophyll fluorescence (SIF), vegetation indices (VIs), and meteorological data. The results demonstrated that the following: (1) SIF could effectively capture interannual fluctuations in winter wheat yield and serve as a reliable quantitative indicator of yield variation. (2) Utilizing vegetation data such as SIF and the near-infrared reflectance of vegetation (NIRv), the developed models could directly quantify drought-induced yield losses in winter wheat based on normalized anomalies of vegetation and meteorological variables, without the need for additional auxiliary data or complex computations. Among all variable combinations tested, SIF demonstrated superior performance, yielding the most accurate predictions. (3) Both random forest (RF) and extreme gradient boosting (XGBoost) algorithms had similar performance in evaluating drought-induced yield loss. The results highlighted the advantages of combining the normalized anomaly of multiple sources of data as inputs in stress-induced crop yield loss evaluation, which was helpful for quick monitoring and early warning of the crop yield loss in the major grain production region.

1. Introduction

Agriculture, as the foundation of human survival, has always been a focus of policymakers worldwide. With recent intensification in global warming, increasingly frequent occurrences of extreme weather events and natural disasters trigger severe agricultural losses, which may further exacerbate imbalances in food production and consumption, and ultimately threaten global food security [1]. Therefore, how to timely and accurately assess crop yield reductions caused by environmental stressors has long been a critical issue for government departments across countries. Such assessment is essential for national agricultural management, economic planning, and commodity market forecasting [2].
Drought is a key stressor inflicting substantial damage on agricultural systems in China [3,4]. Most previous studies have investigated the impacts of drought on crop yield through field experiments, crop model simulations, and statistical analyses of historical data. However, field-based approaches are often time-consuming, labor-intensive, and require specific site conditions. Crop models have the advantage of greater flexibility in simulating various drought scenarios, but they face some challenges due to the complexity of crop growth mechanisms and the uncertainties in input parameters. In contrast, statistical analysis based on historical data remains the most widely applied method [5]. This approach typically establishes quantitative relationships between meteorological drought indices (e.g., standardized precipitation index (SPI), standardized precipitation-evapotranspiration index (SPEI), etc.) [6], soil moisture (SM) indicators [7,8], and crop yield reduction rates. Additionally, from the perspective of vegetation’s response to drought, various reflectance-based vegetation indices (VIs) [9,10] and crop physiological indicators [11,12] are also commonly employed in agricultural drought monitoring. Among them, VIs are often used to predict crop yield by modeling the relationships between VIs and crop yield during different phenological stages or the whole growing season. However, VIs primarily capture canopy greenness rather than directly reflecting photosynthetic activity, and are often considered to exhibit a significant lag effect [1]. These characteristics make VIs have certain deficiencies in predicting the drought-induced reduction in crop yield.
Solar-induced chlorophyll fluorescence (SIF) refers to an electromagnetic signal produced by chlorophyll a molecules. Its emission occurs specifically when the level of solar radiation absorbed by vegetation surpasses the energy demand required for photosynthesis [13]. It offers a more direct probe into photosynthetic function and has shown promise in overcoming some limitations of traditional VIs. Owing to its strong correlation with photosynthesis, SIF is widely recognized as a direct and non-invasive proxy for assessing the functional status of the photosynthetic apparatus [14]. Based on measurements from towers and aircraft, as well as from satellite-based platforms, SIF provides new perspectives for observing terrestrial ecosystems at different spatial and temporal scales. In particular, satellite-based SIF observations are considered to be the most directly measurable signals of photosynthesis in terrestrial ecosystems [15,16]. As an indicator of photosynthetic activity, SIF contains much information on vegetation physiological, biochemical, and metabolic traits. It has been reported to respond rapidly to drought-induced alterations in vegetation function, which positions SIF as an emerging tool with considerable potential for drought monitoring [17]. By virtue of its high sensitivity to environmental stresses such as drought that cause physiological changes such as abnormal photosynthesis, and its close relationship with photosynthesis, SIF has the ability to assess stress-induced crop yield loss, which forms a bridge between drought stress and crop yield. Song et al. [18,19] and others have applied GOME-2 SIF to assess the impact of high temperatures on winter wheat yields in India in 2010. In addition, Cao et al. [20] and Zhang et al. [21] found the consistency between SIF reduction and crop yield reduction due to drought. The previous studies confirm the great potential of satellite-borne SIF observations for large-scale assessment of the impacts of environmental stress on crop yield.
Therefore, taking five provinces in the North China Plain (NCP)—the main winter wheat producing region in China—as the study area, this study intends to explore the impacts of drought on winter wheat yield by developing a model for estimating drought-induced yield loss by combining satellite-based SIF and VIs. This research aims to address the following questions:
  • Can SIF effectively indicate changes in winter wheat yield caused by drought?
  • How effective is a drought-induced yield reduction model based on SIF and VIs?
  • How do different machine learning (ML) models perform in predicting drought-induced yield loss in winter wheat?
Relevant results will be helpful for better understanding the potential of SIF in large-scale assessment of drought-induced crop yield loss, which is critical for national food security, agricultural supply chain management, and economic market forecasting [22].

2. Materials and Methods

2.1. Study Area

Five provinces in NCP include Henan, Shandong, Anhui, Hebei, and Jiangsu Provinces. The region is located in eastern China, whose latitudes span from 29°41′ N to 42°40′ N and longitudes span from 110°21′ E to 122°42′ E, covering a total area of approximately 760,000 km2 (Figure 1). The temperate continental monsoon climate provides favorable conditions for the cultivation of winter wheat.

2.2. Data

The available data could be categorized into the following six types (Table 1). These data were resampled to 0.05° to ensure the consistency of the spatial resolution of all data.
(1)
Satellite-based vegetation data
(a)
SIF data:
Most of the early satellite-based SIF products had coarse resolution and sparse coverage at the global scale, which somewhat hindered the application of SIF data in studies ranging from the ecosystem scale to the global scale. To address this issue, Li and Xiao [23] developed the Global OCO-2 Solar-Induced Fluorescence (GOSIF) dataset, which featured improved temporal and spatial resolution (8-day, 0.05°) and an extended record (from 2000 to the present). With high spatial resolution, continuous global coverage, and an extended temporal record, the GOSIF dataset provided a valuable resource for assessing terrestrial photosynthesis, monitoring ecosystem functioning, and benchmarking terrestrial biosphere and earth system models. These capabilities made it particularly suitable for crop growth monitoring.
(b)
MODIS data:
The MCD43C4 BRDF-adjusted reflectance dataset provides stable, perspective-effect-corrected data. Based on this dataset, we calculated multiple vegetation indices, including the widely used normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI), as well as the near-infrared reflectance of vegetation (NIRv) and the 2-band EVI (Text (1)–(4)) [24]. These VIs were then aggregated to monthly values. The theoretical basis for NIRv was the total near-infrared emissivity partially reflected by surface vegetation [25]. NIRv had been shown to correlate strongly with SIF; EVI2 could be used for those sensors that did not have a blue band to calculate VIs similar to the EVI [26].
N D V I = N I R − R E D N I R + R E D
  N I R v = N I R × N D V I
E V I = G N I R − R E D N I R + C 1 R E D − C 2 B L U E + L
E V I 2 = 2.5 N I R − R E D N I R + 2.4 R E D + 1
where NIR, RED, and BLUE were reflectance in the near-infrared, red, and blue bands, respectively. For EVI, L was the soil conditioning parameter; C1 and C2 were the coefficients of the aerosol resistance term (which used the blue band to correct for aerosol effects in the red band).
(2)
Climate data:
Meteorological elements significantly influenced crop growth. For this study, we obtained monthly data from the TerraClimate dataset, which provided global land surface climate and water balance information, including vapor pressure deficit (Vpd), minimum and maximum temperatures (Tmin, Tmax), solar radiation (Srad), precipitation (Ppt), evapotranspiration (Pet), and SM [27].
(3)
Soil properties data:
Given that soil properties, which governed soil hydrology, and management practices were both critical to crop yield, this study incorporated a set of key soil parameters to represent these foundational conditions. The data, comprising soil potential of hydrogen (PH), cation exchange capacity (CEC), bulk density (BD), soil organic matter (SOM), particle size distribution–sand content (PSD-SA), particle size distribution–silt content (PSD-SI), and particle size distribution–clay content (PSD-CL), were sourced from China’s Soil Particle-Size Distribution (SPD) dataset. This dataset provided representative information for the top 30 cm soil layer at a 1 km resolution [28].
(4)
SPEI data:
Xia et al. [29] calculated the SPEI index at meteorological stations, combined with the precipitation (Ppt) from global precipitation measurement (GPM), land surface temperature (LST) from Moderate Resolution Imaging Spectroradiometer (MODIS) publications database, shortwave radiation (SR) from ERA5-Land, digital elevation model (DEM) from shuttle radar topography mission (SRTM), and random forest regression (RFR) model, and developed a high spatial resolution (1 km) SPEI dataset with multiple time scales in mainland China from 2001 to 2022. Compared to other SPEI datasets, the datasets had higher spatial resolution and could effectively identify the detailed characteristics of drought in mainland China. Therefore, it was used to select the years when droughts occurred with run theory for establishing the drought-induced loss evaluation models of winter wheat, and the drought levels were divided into mild drought, moderate drought, severe drought, and extreme drought (Table 2).
(5)
Winter wheat yield and planting area
Winter wheat yield data at the municipal level for Anhui, Henan, Hebei, and Shandong provinces from 2007 to 2020 were sourced from provincial statistical yearbooks (unit: kg/ha), while data for Jiangsu province were obtained from prefecture-level city yearbooks. Planting areas for winter wheat from 2007 to 2015 were derived from the ChinaCropArea1 km annual harvested area dataset [30]. Due to the lack of updated planting area data after 2015, the planting areas of winter wheat from 2016 to 2020 were considered to be the same as the planting area in 2015 in this study by considering the Chinese government’s policy of implementing special protection for cultivated land (red line of cultivated land) [31].
For subsequent calculation and analysis, only pixels in which winter wheat accounted for more than 60% of the area were retained, based on the proportional coverage per SIF and VI pixel [32]. Given the absence of yield records in some prefecture-level cities and the relatively small planting scale in others, 58 cities across the five provinces were ultimately selected as the study area for modeling the drought-induced loss evaluation models of winter wheat.

2.3. Methods

2.3.1. Screening for Optimal Variables

This study employed a two-stage analytical approach. First, a Spearman correlation analysis was performed to assess the initial associations between winter wheat yield and a comprehensive set of variables, including vegetation variables, climatic variables, and soil properties variables. This analysis aimed to identify key factors influencing drought-induced yield loss and to reduce input variables without incurring significant information loss. Subsequently, an exploratory data analysis (EDA) was conducted to systematically filter these variables, with a focus on structuring the climatic variables into thematic groups (temperature, water availability, water demand, radiation). The selection process applied a two-tiered criterion based on the correlation results: identifying the most yield-significant variable per group (p < 0.01), then retaining others from the same group only if they were not highly correlated (r ≤ 0.5) with the primary choice.

2.3.2. Calculation of SYRS

Using crop meteorological yields, Potopová et al. [33] developed a standardized yield residuals series (SYRS) (see Equation (5) to quantify the impacts of drought on yield by establishing a statistical relationship with the SPEI.
S Y R S = X i − μ σ
where SYRS is the standardized yield residual series, Xi is the yield residual, µ is the mean of the yield residual, and σ is the standard deviation of the yield residual. The mean of SYRS was 0 and the standard deviation was 1. Therefore, to facilitate the analysis of yield variability, this same transformation method was used in this study. This method standardized the residuals obtained from the regression model. Ultimately, SYRS was calculated and used to assess the interaction between drought and yield reduction.
Moreover, thanks to its standardized nature, SYRS enabled consistent comparison between the response of crops to drought across diverse regions and growing seasons, enhancing its generalizability across space and time. This approach provided an agriculturally and economically relevant measure of drought effects by directly linking climate anomalies to final crop yield, thereby offering valuable insights for yield forecasting and agricultural risk management strategies.

2.3.3. Experiment Design

To better investigate the impacts of drought on winter wheat yield, drought years were identified based on the run theory with SPEI as the drought monitoring indicator. The selected years were 2008, 2009, 2011, 2014, 2015, 2016, 2017, and 2019. Among them, data of multiple variables and winter wheat yield in 2008, 2009, 2011, 2014, 2015, 2016, and 2017, were used for model training, while data in 2019 as an independent prediction year was reserved to evaluate model performances under drought conditions. Since the growth and development (e.g., jointing and heading stage) of winter wheat during the vigorous growth stage was the most crucial for yield formation [4,34], the mean values of predictors (including climate data, satellite-based vegetation variables data, and soil properties data) from March to May were calculated to develop the SIF-based model for assessing drought-induced yield reduction in this study, whose dataset was split into 70% for training and 30% for validation during modelling.
Before model training and testing, the correlation analysis between vegetation variables, climate variables, and winter wheat yield was carried out for screening for optimal input variables (Figure S1). With higher correlation coefficients with winter wheat yield than NDVI, EVI, EVI2, and NIRv, SIF displayed the strongest relationship with winter wheat yield, explaining approximately 47% of the yield variation. In contrast, climate variables generally exhibited weaker correlations with yield than vegetation variables. Five climate variables were retained from an initial pool of seven: Ppt, SM, Tmin, Vpd, and Srad. Similarly, BD and PSD-CL were excluded from the soil properties data because they had very little correlation with winter wheat yield. Then, all selected input variables (the mean values of predictors from March to May) and winter wheat yield were standardized using the z-score method, which prevented model performance degradation due to individual features deviating from a standard normal distribution. During the model development, four model configurations (C1: SIF + climate + soil properties; C2: VIs + climate + soil properties; C3: SIF + NIRv + climate + soil properties; C4: climate + soil properties) with seven different variable combinations were designed for modeling with two different ML models (RF, XGBoost). In this study, the 10-fold CV method was employed in 50 repetitive experiments, which provided a more stable and reliable performance evaluation and reduced over-reliance on specific datasets. Finally, R2 and RMSE were used to evaluate the predictive performance of each model.

3. Results

3.1. Inter-Annual Variation in SIF and Winter Wheat Yield

Given that crop growth during key developmental stages had the greatest impact on yield formation, we first examined the interannual variations in SIF during the critical growing season (March to May) and winter wheat yield to evaluate the capability of SIF in indicating changes in yield. As shown in Figure 2, winter wheat yield showed an overall increasing trend from 2007 to 2020. The interannual fluctuations in SIF during the critical growing season were largely synchronous with yield variations over the full period. Despite a period of divergence between 2007 and 2014, SIF demonstrated a clear capacity to track yield dynamics for winter wheat. Specifically, during the drought years of 2016 and 2018, declines in SIF accurately mirrored the drought-induced yield losses.

3.2. Performance of a SIF-Based Model for Assessing Drought-Induced Yield Reduction

In this study, four distinct model configurations—(C1:SIF: SIF + climate + soil properties; C2: VIs: EVI + climate + soil properties, EVI2 + climate + soil properties, NIRv + climate + soil properties, NDVI + climate + soil properties; C3: SIF + VIs: SIF + NIRv + climate + soil properties; C4: climate + soil properties)—were employed as inputs to ML (both RF and XGBoost) algorithms to develop models for assessing drought-induced yield reduction. The predictive performance of each model was thoroughly evaluated using the R2 and RMSE, with detailed results visualized in Figure 3 and Figure 4.
Among the RF-based models, the SIF-driven model (C1) delivered superior performance, with the highest average R2 value of 0.69 and the lowest RMSE of 0.563, indicating a strong capability for accurately estimating yield loss. The hybrid model incorporating both SIF and NIRv (C3) also demonstrated commendable results, with an R2 of 0.67, an RMSE of 0.57, and the smallest standard deviation among all model configurations, reflecting high stability and reliability. When comparing individual VIs, the prediction effect of NDVI was not much different from that of NIRv, with similar R2 and RMSE. EVI2 exceeded EVI in predictive accuracy, with higher R2 values. In contrast, the model relying solely on climate and soil properties(C4) exhibited the weakest performance, with an R2 of only 0.59 and an RMSE of 0.631, underscoring the limited utility of non-optical variables for this specific task.
A similar pattern was observed within the XGBoost models. The SIF-driven model (C1) again emerged as the top performer, with an R2 of 0.67 and an RMSE of 0.576. Followed by the model using EVI2 as the sole input vegetation variable, with an R2 of 0.64 and an RMSE of 0.580. Although the combined use of SIF and NIRv led to a minor improvement over the model using NIRv alone, it did not surpass the accuracy achieved by the model using SIF alone. Consistent with the RF results, the climate + soil combination (C4) yielded the poorest predictive performance among all variable sets under XGBoost. Furthermore, the model based exclusively on EVI produced the least accurate predictions among all vegetation-related inputs.
In summary, SIF consistently proved to be the most effective variable for predicting drought-related yield reduction in winter wheat, outperforming all other single and combined inputs across both ML architectures. The combination of meteorological and soil properties parameters consistently resulted in the lowest accuracy, highlighting the superiority of remotely sensed physiological indicators. It was noteworthy that combined SIF and NIRv models, although superior to those using only traditional VIs, did not exceed the performance of SIF-only models, suggesting that SIF already captured most of the signal related to photosynthetic efficiency and stress response. This reinforced the value of SIF as a robust data source capable of enhancing prediction accuracy beyond the limits of conventional VIs.

3.3. Assessment of Yield Losses in a Typical Drought Year

To further evaluate the predictive performance of different input variables across models, the trained models were used to predict yield reduction induced by drought for the test year of 2019, and the results were compared with actual yield loss data (Figure 5 and Figure 6). Results showed that: In the RF model, SIF achieved an R2 of 0.712, while in the XGBoost model, its performance improved with an R2 of 0.727, significantly outperforming traditional VIs such as NIRv and NDVI. In contrast, climate data performed poorly in both algorithmic frameworks, reaching only an R2 of 0.658 in the RF model and declining to 0.586 in the XGBoost model, highlighting the limitations of relying solely on climate factors for assessing drought-related yield loss.
Further analysis of the synergy between SIF and NIRv showed that combining these metrics enhanced prediction accuracy in the RF and XGBoost models. Specifically, the combined inputs achieved an R2 of 0.719 and an RMSE of 0.525 in the RF model. In the XGBoost model, the combined inputs retained strong performance, though with a slight decrease in R2 to 0.711 and an increase in RMSE to 0.533. These results indicated that a multi-parameter fusion strategy could effectively capture complex dynamic processes in crop growth, especially under extreme climate conditions. In addition, by comparing the spatial distribution of drought severity (represented by SPEI) and the SYRS-simulated yield reduction (Figure 7), strong consistency between them across most drought-prone regions indicated reliable model performance in these critical areas. Combining with the irrigation district data, regions with extensive irrigation systems experienced relatively minor yield impacts, whereas areas with limited irrigation coverage exhibited yield reductions consistent with SPEI patterns, which was in line with the actual planting situation in NCP. However, although the combination of SIF and NIRv improved prediction accuracy compared to single inputs of NIRv, it did not exceed the performance of single inputs of SIF. This limitation may be due to information redundancy, scale mismatch, difficulties in model tuning, and complexity in accounting for environmental disturbances. Although combining SIF and NIRv could integrate multidimensional information in theory, SIF’s unique advantage as a direct probe of photosynthesis could not be replicated solely by the structural information provided by NIRv. When used together, NIRv may act mainly as a redundant variable, providing limited improvement in predictive accuracy.
Therefore, the model employing SIF as the input vegetation variable demonstrated excellent predictive capability in assessing drought-induced yield losses and showed strong consistency with SYRS-simulated reductions, particularly in areas of higher drought severity. Both the RF and XGBoost algorithms exhibited robust predictive performance within the multi-source data fusion framework. However, the model incorporating both SIF and NIRv as vegetation inputs did not surpass the predictive performance achieved by using SIF alone.

4. Discussion

4.1. Relationship Between SIF and Winter Wheat Yield

As a key physiological indicator directly linked to photosynthetic activity, SIF demonstrated a distinct advantage over traditional VIs (e.g., NDVI) by capturing drought-induced changes in vegetation physiological status more sensitively and in a more timely manner. Therefore, SIF showed great potential in improving agricultural drought monitoring and impact assessment, such as crop yield loss. In previous studies, Song et al. [18] found that average SIF values during the growing season in the Indian Indo-Gangetic Plains were more capable of capturing the interannual variation of winter wheat yield than NDVI and EVI. In addition to the average SIF values during the growing season, the average SIF values during different growth stages were also used to indicate changes in crop yields. Cai et al. [34] found that the correlation between month-scale EVI and SIF and wheat yield increased until October in Australia (i.e., the growing season), and then the correlation with yield began to decline. Gao et al. [2] found that during the growth peak period, the monthly average, maximum SIF, and EVI showed similar correlations with the yields of corn and soybeans. However, the performance of SIF in indicating yield at different growth stages was still not superior to that of VIs. In the study, we intended to explore the impacts of drought on winter wheat yield by developing a model for estimating drought-induced yield loss by combining satellite-based SIF and VIs. Therefore, In the selection of input parameters, the growth characteristic during the phenological stages of winter wheat were also important consideration factors.
Winter wheat was typically sown in October and harvested in June of the following year. Its main growth stages included greening (early to mid-March), jointing (late March to mid-April), heading (late April to early May), and grain-filling (mid to late May). Among these, the stages from jointing to heading were critical for determining the number of grains per spike, while the duration and rate of grain-filling affected grain weight, which ultimately influenced yield. During the critical jointing and heading stages, photosynthetic resources were predominantly allocated to reproductive development and kernel formation rather than vegetative growth, making SIF measurements during these periods particularly indicative of final yield. Given that drought during key growth stages—particularly jointing and heading—had the greatest impacts on yield [4,35], predictors including climatic variables, VIs, and soil properties were aggregated as monthly averages from March to May for use in the study. After analyzing the relationship between SIF values and winter wheat yield (Figure 2), this paper found that the average SIF values from March to May were consistent with the interannual variation of winter wheat yield, confirming that SIF during the peak growing season could indicate the interannual variation of winter wheat to a certain extent, which was consistent with the research results of predecessors.
The observed divergence between SIF and winter wheat yield from 2007 to 2011 underscores their decoupling under specific conditions. A key manifestation of this is the decline in SIF that was not accompanied by a commensurate yield loss. This phenomenon could be largely attributed to effective agronomic interventions, particularly irrigation. During periods of water stress that depressed canopy-level photosynthesis (as detected by SIF), supplemental irrigation could alleviate the water deficit, thereby protecting grain formation and final yield. This demonstrated that modern agricultural management could effectively buffer yields against short-term photosynthetic suppression, highlighting that SIF served as a sensitive early stress indicator, while final yield was strongly modulated by human intervention.

4.2. Predictive Power of Drought-Induced Yield Loss Evaluation Models of Winter Wheat

A previous study had confirmed that SIF exhibited greater sensitivity to vegetation photosynthetic capacity than traditional VIs [17], and it showed a significant correlation with crop yield [36,37]. Therefore, SIF had strong potential to serve as a reliable indicator of crop yield [18,38], and compared with VIs, it had higher accuracy in predicting the yield loss of winter wheat caused by drought. Therefore, with a comprehensive set of data sources including SIF, VIs, meteorological records, and soil properties from March to May, this study intended to build a drought-induced loss evaluation model to analyze the response of winter wheat to drought. Additionally, rather than directly using SIF to estimate yield or drought-induced loss, it applied the SYRS as the output variable for modeling drought-induced yield reduction in winter wheat. Recent research suggested that SYRS offered a more efficient and straightforward method for evaluating drought impact on yields across large spatial scales [39].
In this study, drought-induced loss evaluation models of winter wheat based on SIF (C1) outperformed those based on VIs (C2) under drought conditions, indicating that SIF could more accurately capture the impact of drought on winter wheat yield, and estimate the magnitude of drought-induced yield reduction. This enhanced performance stemmed from the well-established linear relationship between SIF and gross primary productivity (GPP) [40,41], which enabled more sensitive detection of photosynthetic dynamics under water stress conditions. The findings substantiated that SIF provided a more physiologically grounded and accurate representation of drought impacts on winter wheat productivity compared to conventional VIs. In terms of predicting yield losses caused by drought, the combination of SIF and NIRv demonstrated a moderate level of predictive ability, consistently outperforming NIRv but not entirely better than SIF. This is because while both SIF and NIRv are linked to GPP and capture overlapping structural information [42], SIF provides a more direct and sensitive measure of the actual photosynthetic physiology impaired by drought. The marginal gain from adding NIRv implies that the physiological information intrinsic to SIF is dominant for capturing drought impact, largely subsuming the structural information contributed by NIRv.
In addition, analysis of feature importance (FI) among model predictors showed that PH was the most important variable, significantly outweighing others (Figure S2). SIF and other soil properties (e.g., PSD-SI and SOM) also contributed considerably. The high importance of static environmental factors such as soil properties (e.g., PH) aligned with the findings of Zhang et al. [43]. In the C3 models, vegetation-related variables (SIF and NIRv) were notably more important than climate factors, with SIF exhibiting stronger predictive power among vegetation indicators. Further validation analyses revealed that SIF had higher explanatory power than NIRv in both RF and XGBoost models when both were included. It was noteworthy that climate variables contributed the least across all models, which was consistent with the results from those studies estimating crop yield using climate data alone [44].
Wholly, this paper used normalized anomaly of average SIF and NIRv values during the peak growing season and SYRS as the input and output parameters to establish drought-induced yield loss evaluation models of winter wheat. This model helped to more simply and quickly evaluate the impact of drought on yield on a large scale. Compared with traditional methods, the method using the relative anomalies of SIF and VIs to directly assess crop yield losses did not require the assistance of other data or complex calculations. Therefore, it had more potential and advantages in drought monitoring and assessment, thereby offering valuable insights for yield loss forecasting and agricultural risk management strategies.

4.3. Comparison of Different ML Algorithms Used in the Modeling

In terms of model selection in this study, RF and XGBoost algorithms were chosen to construct a drought-induced loss evaluation model of winter wheat. Both of them had advantages in handling high-dimensional data and were relatively less susceptible to overfitting [45], and had been used widely in crop yield estimation. For example, Zheng et al. [37] found that RF and XGBoost algorithms were significantly better than the other ML algorithms in winter wheat yield prediction. Wang et al. [46] found that RF algorithm had a relatively high accuracy in yield prediction. Zhang et al. [47] had achieved high-precision prediction by integrating VIs and meteorological data in crop yield estimation. These studies further confirmed the applicability of tree-based methods such as RF and XGBoost algorithms for crop yield estimation. This consistency in the findings of related studies enhanced the credibility of the adopted yield estimation model (RF/XGBoost), and prompted us to choose them for the assessment of drought-induced yield loss.
Specifically, the performance of models in this paper revealed that the RF algorithm typically outperformed the XGBoost algorithm during model training; however, when evaluating the actual prediction performance on the final independent testing, the drought-induced yield loss prediction accuracies of the two algorithms showed a high degree of similarity. A key reason for the difference lied in the RF algorithm’s randomness mechanism (including random selection of samples and features), which allowed it to exhibit greater robustness during cross-validation, resulting in less fluctuation in performance and more stable and consistent results across multiple experimental replications [45]. In contrast, the XGBoost algorithm may be more sensitive to specific hyperparameter settings and data distributions, so it may perform better in a single experimental run, but its performance may exhibit greater volatility over multiple repetitions.

4.4. Shortcomings and Prospects

Most spatially downscaled SIF products relied on a variety of high spatial resolution explanatory variables, such as visible light, near-infrared reflectance, and meteorological information. Since the explanatory variables could not reflect changes in vegetation photosynthesis, the vegetation physiological information contained in the original SIF data was inevitably lost. Therefore, these products would inevitably lose the vegetation physiological information of the original SIF data in the process of spatio-temporal expansion, which was different from the real SIF products. This, to some extent, made the prediction of drought-induced yield reduction potentially error-prone. The availability of SIF data with higher spatial and temporal resolution, such as Sentinel-5 [48] and FLuorescence EXplorer (FLEX) [49], would greatly contribute to the application of SIF in large-scale crop yield estimation. The use of higher resolution remote sensing satellite data in the future could contribute to improving the predictive power of drought-induced loss evaluation models of winter wheat to some extent. In addition, due to limitations of yield reports for crops, the analyses in this study were constructed on a relatively coarse municipal level. In the future, drought-induced yield reduction could be assessed on the county level scale, which may show better performance.

5. Conclusions

With a comprehensive set of data sources, including SIF, VIs, meteorological data, and soil properties data, this study investigated the impacts of drought on winter wheat yield by developing a model for assessing drought-induced yield reduction by using statistical ML algorithms (RF, XGBoost), in five provinces of the NCP. The results demonstrated the following:
1.
Average SIF values during the peaking growing season could effectively capture interannual fluctuations in winter wheat yield and could serve as a quantitative indicator of yield variability.
2.
By integrating multiple sources of data, the models based on normalized anomalies of these variables could directly quantify drought-induced yield losses without the need for additional auxiliary data or complex computations. Among all combinations of vegetation variables, SIF demonstrated superior performance, yielding the most accurate predictions.
3.
Both RF and XGBoost algorithms exhibited similar strong performance in evaluating drought-induced yield loss.
Overall, this study underscored the significant potential of SIF-driven modeling combined with ML in advancing precision agriculture, enhancing drought resilience, and supporting food security through more accurate and timely yield loss assessments.

Supplementary Materials

The following supporting information can be downloaded at https://www.mdpi.com/article/10.3390/atmos17010042/s1, Figure S1: The correlations between winter wheat yield and the different types of variables in the North China Plain. *, ** and *** represent significance levels of p < 0.05, p < 0.01 and p < 0.001, respectively; No annotation denotes significance levels above 0.05. Figure S2 Feature importance (FI) scores of predictors in the drought-induced yield loss evaluation models of winter wheat: (a) RF: SIF (b) RF: NIRv, (c) RF: SIF + NIRv, (d) XGBoost: SIF, (e) XGBoost: NIRv, (f) XGBoost: SIF + NIRv.

Author Contributions

H.H. collected and analyzed the data and wrote the manuscript. M.Z. collected and processed the data. Y.N. analyzed the data. Q.S., as the corresponding author, edited the manuscript and provided constructive comments. Q.R. and Y.Y. assisted with data processing and analysis. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China under Grant 42071401.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data supporting the findings of this study are publicly available from the following sources: SIF data at http://globalecology.unh.edu (accessed on 30 March 2025), VIs at https://lpdaac.usgs.gov (accessed on 30 March 2025), climate data at https://doi.org/10.7923/G43J3B0R (accessed on 30 March 2025), and SPEI data at http://globalchange.bnu.edu.cn (accessed on 30 March 2025).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Shen, Q.; Lin, J.; Yang, J.; Zhao, W.; Wu, J. Exploring the Potential of Spatially Downscaled Solar-Induced Chlorophyll Fluorescence to Monitor Drought Effects on Gross Primary Production in Winter Wheat. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 2012–2022. [Google Scholar] [CrossRef]
  2. Gao, Y.; Wang, S.; Guan, K.; Wolanin, A.; You, L.; Ju, W.; Zhang, Y. The Ability of Sun-Induced Chlorophyll Fluorescence From OCO-2 and MODIS-EVI to Monitor Spatial Variations of Soybean and Maize Yields in the Midwestern USA. Remote Sens. 2020, 12, 1111. [Google Scholar] [CrossRef]
  3. Wu, J.; Cheng, G.; Wang, N.; Shen, H.; Ma, X. Spatiotemporal Patterns of Multiscale Drought and Its Impact on Winter Wheat Yield over North China Plain. Agronomy 2022, 12, 1209. [Google Scholar] [CrossRef]
  4. Zhang, L.; Chu, Q.-q.; Jiang, Y.-l.; Chen, F.; Lei, Y.-d. Impacts of climate change on drought risk of winter wheat in the North China Plain. J. Integr. Agric. 2021, 20, 2601–2612. [Google Scholar] [CrossRef]
  5. Harsányi, E. Predicting agricultural drought in central Europe by using machine learning algorithms. J. Agric. Food Res. 2025, 20, 101783. [Google Scholar] [CrossRef]
  6. Yang, J.; Wu, J.; Liu, L.; Zhou, H.; Gong, A.; Han, X.; Zhao, W. Responses of Winter Wheat Yield to Drought in the North China Plain: Spatial–Temporal Patterns and Climatic Drivers. Water 2020, 12, 3094. [Google Scholar] [CrossRef]
  7. Mu, Q.; Cai, H.; Sun, S.; Wen, S.; Xu, J.; Dong, M.; Saddique, Q. The physiological response of winter wheat under short-term drought conditions and the sensitivity of different indices to soil water changes. Agric. Water Manag. 2021, 243, 106475. [Google Scholar] [CrossRef]
  8. Zhou, H.; Geng, G.; Yang, J.; Hu, H.; Sheng, L.; Lou, W. Improving Soil Moisture Estimation via Assimilation of Remote Sensing Product into the DSSAT Crop Model and Its Effect on Agricultural Drought Monitoring. Remote Sens. 2022, 14, 3187. [Google Scholar] [CrossRef]
  9. Hou, X.; Zhang, B.; Chen, J.; Zhou, J.; He, Q.-Q.; Yu, H. Response of Vegetation Productivity to Greening and Drought in the Loess Plateau Based on VIs and SIF. Forests 2024, 15, 339. [Google Scholar] [CrossRef]
  10. Zhu, X.; Li, Q.; Guo, C. Evaluation of the monitoring capability of various vegetation indices and mainstream satellite band settings for grassland drought. Ecol. Inform. 2024, 82, 102717. [Google Scholar] [CrossRef]
  11. Gao, X.-B.; Guo, C.; Li, F.-M.; Li, M.; He, J. High Soybean Yield and Drought Adaptation Being Associated with Canopy Architecture, Water Uptake, and Root Traits. Agronomy 2020, 10, 608. [Google Scholar] [CrossRef]
  12. Thakur, V.; Rane, J.; Nankar, A.N. Comparative Analysis of Canopy Cooling in Wheat under High Temperature and Drought Stress. Agronomy 2022, 12, 978. [Google Scholar] [CrossRef]
  13. Zhao, F.; Li, R.; Verhoef, W.; Cogliati, S.; Liu, X.; Huang, Y.; Guo, Y.; Huang, J. Reconstruction of the full spectrum of solar-induced chlorophyll fluorescence: Intercomparison study for a novel method. Remote Sens. Environ. 2018, 219, 233–246. [Google Scholar] [CrossRef]
  14. Liu, L.; Zhao, W.; Shen, Q.; Wu, J.; Teng, Y.; Yang, J.; Han, X.; Tian, F. Nonlinear Relationship Between the Yield of Solar-Induced Chlorophyll Fluorescence and Photosynthetic Efficiency in Senescent Crops. Remote Sens. 2020, 12, 1518. [Google Scholar] [CrossRef]
  15. Behera, S.; Elakkiyaa, T.L.; Sarkar, A.; Dutta, D. Solar-Induced Chlorophyll Fluorescence Yield Holds the Potential for Drought Early Warning. Geophys. Res. Lett. 2025, 52, 3419. [Google Scholar] [CrossRef]
  16. Mohammadi, K.; Jiang, Y.; Wang, G. Flash drought early warning based on the trajectory of solar-induced chlorophyll fluorescence. Proc. Natl. Acad. Sci. USA 2022, 119, e2202767119. [Google Scholar] [CrossRef]
  17. Ren, H.; Du, L.; Peng, C.; Yang, J.; Gao, W. The composite drought index incorporated solar-induced chlorophyll fluorescence enhances the monitoring capability of short-term drought. J. Hydrol. 2024, 637, 131361. [Google Scholar] [CrossRef]
  18. Song, L.; Guanter, L.; Guan, K.; You, L.; Huete, A.; Ju, W.; Zhang, Y. Satellite sun-induced chlorophyll fluorescence detects early response of winter wheat to heat stress in the Indian Indo-Gangetic Plains. Glob. Change Biol. 2018, 24, 4023–4037. [Google Scholar] [CrossRef] [PubMed]
  19. Song, Y.; Wang, J.; Wang, L. Satellite Solar-Induced Chlorophyll Fluorescence Reveals Heat Stress Impacts on Wheat Yield in India. Remote Sens. 2020, 12, 3277. [Google Scholar] [CrossRef]
  20. Cao, J.; An, Q.; Zhang, X.; Xu, S.; Si, T.; Niyogi, D. Is satellite Sun-Induced Chlorophyll Fluorescence more indicative than vegetation indices under drought condition? Sci. Total Environ. 2021, 792, 148396. [Google Scholar] [CrossRef] [PubMed]
  21. Zhang, Z.; Li, X.; Qiu, Y.; Shi, Z.; Gao, Z.; Jia, Y. A Spatial Downscaling Method for Solar-Induced Chlorophyll Fluorescence Product Using Random Forest Regression and Drought Monitoring in Henan Province. Remote Sens. 2024, 16, 963. [Google Scholar] [CrossRef]
  22. Zhang, Z.; Xu, W.; Qin, Q.; Long, Z. Downscaling Solar-Induced Chlorophyll Fluorescence Based on Convolutional Neural Network Method to Monitor Agricultural Drought. IEEE Trans. Geosci. Remote Sens. 2021, 59, 1012–1028. [Google Scholar] [CrossRef]
  23. Li, X.; Xiao, J. A Global, 0.05-Degree Product of Solar-Induced Chlorophyll Fluorescence Derived from OCO-2, MODIS, and Reanalysis Data. Remote Sens. 2019, 11, 517. [Google Scholar] [CrossRef]
  24. Holben, B.N. Characteristics of maximum-value composite images from temporal AVHRR data. Int. J. Remote Sens. 1986, 7, 1417–1434. [Google Scholar] [CrossRef]
  25. Zeng, Y.; Badgley, G.; Dechant, B.; Ryu, Y.; Chen, M.; Berry, J.A. A practical approach for estimating the escape ratio of near-infrared solar-induced chlorophyll fluorescence. Remote Sens. Environ. 2019, 232, 111209. [Google Scholar] [CrossRef]
  26. Jiang, Z.; Huete, A.; Didan, K.; Miura, T. Development of a two-band enhanced vegetation index without a blue band. Remote Sens. Environ. 2008, 112, 3833–3845. [Google Scholar] [CrossRef]
  27. Abatzoglou, J.T.; Dobrowski, S.Z.; Parks, S.A.; Hegewisch, K.C. TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015. Sci. Data 2018, 5, 170191. [Google Scholar] [CrossRef]
  28. Shangguan, W.; Dai, Y.; Liu, B.; Ye, A.; Yuan, H. A soil particle-size distribution dataset for regional land and climate modelling in China. Geoderma 2012, 171–172, 85–91. [Google Scholar] [CrossRef]
  29. Xia, H.; Sha, Y.; Zhao, X.; Jiao, W.; Song, H.; Yang, J.; Zhao, W.; Qin, Y. HSPEI: A 1-km spatial resolution SPEI dataset across the Chinese mainland from 2001 to 2022. Geosci. Data J. 2024. [Google Scholar] [CrossRef]
  30. Luo, Y.; Zhang, Z.; Li, Z.; Chen, Y.; Zhang, L.; Cao, J.; Tao, F. Identifying the spatiotemporal changes of annual harvesting areas for three staple crops in China by integrating multi-data sources. Environ. Res. Lett. 2020, 15, 074003. [Google Scholar] [CrossRef]
  31. Yang, T.; Qiu, J.; Xiao, G. Agricultural drought monitoring and winter wheat yield estimation in North China. Acta Ecol. Sin. 2023, 43, 1936–1947. [Google Scholar] [CrossRef]
  32. Wood, J.D.; Griffis, T.J.; Baker, J.M.; Frankenberg, C.; Verma, M.; Yuen, K. Multiscale analyses of solar-induced florescence and gross primary production. Geophys. Res. Lett. 2017, 44, 533–541. [Google Scholar] [CrossRef]
  33. Potopová, V.; Zahradníček, P.; Štěpánek, P.; Türkott, L.; Farda, A.; Soukup, J. The impacts of key adverse weather events on the field-grown vegetable yield variability in the Czech Republic from 1961 to 2014. Int. J. Climatol. 2016, 37, 1648–1664. [Google Scholar] [CrossRef]
  34. Yao, J.; Wu, Y.; Liu, J.; Wang, H. Multimodal deep learning-based drought monitoring research for winter wheat during critical growth stages. PLoS ONE 2024, 19, e0300746. [Google Scholar] [CrossRef]
  35. Cai, Y.; Guan, K.; Lobell, D.; Potgieter, A.B.; Wang, S.; Peng, J.; Xu, T.; Asseng, S.; Zhang, Y.; You, L.; et al. Integrating satellite and climate data to predict wheat yield in Australia using machine learning approaches. Agric. For. Meteorol. 2019, 274, 144–159. [Google Scholar] [CrossRef]
  36. Wang, H.; Ding, Y.; Yao, Q.; Ma, L.; Ma, Y.; Yang, M.; Qin, S.; Xu, F.; Zhang, Z.; Gao, Z. Modeling of Cotton Yield Estimation Based on Canopy Sun-Induced Chlorophyll Fluorescence. Agronomy 2024, 14, 364. [Google Scholar] [CrossRef]
  37. Zheng, M.; Hu, H.; Niu, Y.; Shen, Q.; Jia, F.; Geng, X. Modeling of winter wheat yield prediction based on solar-induced chlorophyll fluorescence by machine learning methods. Eur. J. Remote Sens. 2025, 58, 2455940. [Google Scholar] [CrossRef]
  38. Guanter, L.; Zhang, Y.; Jung, M.; Joiner, J.; Voigt, M.; Berry, J.A.; Frankenberg, C.; Huete, A.R.; Zarco-Tejada, P.; Lee, J.-E.; et al. Global and time-resolved monitoring of crop photosynthesis with chlorophyll fluorescence. Proc. Natl. Acad. Sci. USA 2014, 111, E1327–E1333. [Google Scholar] [CrossRef]
  39. Kheyruri, Y.; Asadollah, S.B.H.S.; Sharafati, A.; Rahman, A.; Mohajeri, S.H. Examining the effects of meteorological drought variability on rainfed cultivation yields in Iran. Theor. Appl. Climatol. 2024, 155, 6427–6444. [Google Scholar] [CrossRef]
  40. Bai, J.; Zhang, H.; Sun, R.; Li, X.; Xiao, J.; Wang, Y. Estimation of global GPP from GOME-2 and OCO-2 SIF by considering the dynamic variations of GPP-SIF relationship. Agric. For. Meteorol. 2022, 326, 109180. [Google Scholar] [CrossRef]
  41. Chen, R.; Liu, L.; Liu, Z.; Liu, X.; Kim, J.; Kim, H.S.; Lee, H.; Wu, G.; Guo, C.; Gu, L. SIF-based GPP modeling for evergreen forests considering the seasonal variation in maximum photochemical efficiency. Agric. For. Meteorol. 2024, 344, 109814. [Google Scholar] [CrossRef]
  42. Dechant, B.; Ryu, Y.; Badgley, G.; Zeng, Y.; Berry, J.A.; Zhang, Y.; Goulas, Y.; Li, Z.; Zhang, Q.; Kang, M.; et al. Canopy structure explains the relationship between photosynthesis and sun-induced chlorophyll fluorescence in crops. Remote Sens. Environ. 2020, 241, 111733. [Google Scholar] [CrossRef]
  43. Zhang, L.; Zhang, Z.; Luo, Y.; Cao, J.; Tao, F. Combining Optical, Fluorescence, Thermal Satellite, and Environmental Data to Predict County-Level Maize Yield in China Using Machine Learning Approaches. Remote Sens. 2019, 12, 21. [Google Scholar] [CrossRef]
  44. Li, Z.; Ding, L.; Xu, D. Exploring the potential role of environmental and multi-source satellite data in crop yield prediction across Northeast China. Sci. Total Environ. 2022, 815, 152880. [Google Scholar] [CrossRef]
  45. Ma, J.; Ding, Y.; Cheng, J.C.P.; Jiang, F.; Tan, Y.; Gan, V.J.L.; Wan, Z. Identification of high impact factors of air quality on a national scale using big data and machine learning techniques. J. Clean. Prod. 2020, 244, 118955. [Google Scholar] [CrossRef]
  46. Wang, Y.-Q.; Leng, P.; Shang, G.-F.; Zhang, X.; Li, Z.-L. Sun-induced chlorophyll fluorescence is superior to satellite vegetation indices for predicting summer maize yield under drought conditions. Comput. Electron. Agric. 2023, 205, 107615. [Google Scholar] [CrossRef]
  47. Zhang, G.; Roslan, S.; Shafri, H.Z.M.; Zhao, Y.; Wang, C.; Quan, L. Predicting wheat yield from 2001 to 2020 in Hebei Province at county and pixel levels based on synthesized time series images of Landsat and MODIS. Sci. Rep. 2024, 14, 16212. [Google Scholar] [CrossRef]
  48. Guanter, L.; Aben, I.; Tol, P.; Krijger, J.M.; Hollstein, A.; Köhler, P.; Damm, A.; Joiner, J.; Frankenberg, C.; Landgraf, J. Potential of the TROPOspheric Monitoring Instrument (TROPOMI) onboard the Sentinel-5 Precursor for the monitoring of terrestrial chlorophyll fluorescence. Atmos. Meas. Tech. 2015, 8, 1337–1352. [Google Scholar] [CrossRef]
  49. Drusch, M.; Moreno, J.; Bello, U.D.; Franco, R.; Goulas, Y.; Huth, A.; Kraft, S.; Middleton, E.M.; Miglietta, F.; Mohammed, G.; et al. The FLuorescence EXplorer Mission Concept—ESA’s Earth Explorer 8. IEEE Trans. Geosci. Remote Sens. 2017, 55, 1273–1284. [Google Scholar] [CrossRef]
Atmosphere 17 00042 g001
Figure 1. Winter wheat planting areas in the five provinces in the North China Plain.
Figure 1. Winter wheat planting areas in the five provinces in the North China Plain.
Atmosphere 17 00042 g001
Atmosphere 17 00042 g002
Figure 2. Interannual variation of yield and SIF average during the peak growing season of winter wheat from 2007 to 2020 in the five provinces in North China Plain.
Figure 2. Interannual variation of yield and SIF average during the peak growing season of winter wheat from 2007 to 2020 in the five provinces in North China Plain.
Atmosphere 17 00042 g002
Atmosphere 17 00042 g003
Figure 3. Model performance for predicting drought-induced yield reduction under different data combinations for the selected drought years (2008, 2009, 2011, 2014, 2015, 2016, and 2017). Each model underwent 50 repeated runs. The colored bar represents the mean R2, and the error bar represents the standard error across the 50 runs. The R2 values were calculated based on data standardized using the z-score method.
Figure 3. Model performance for predicting drought-induced yield reduction under different data combinations for the selected drought years (2008, 2009, 2011, 2014, 2015, 2016, and 2017). Each model underwent 50 repeated runs. The colored bar represents the mean R2, and the error bar represents the standard error across the 50 runs. The R2 values were calculated based on data standardized using the z-score method.
Atmosphere 17 00042 g003
Atmosphere 17 00042 g004
Figure 4. Model performance for predicting drought-induced yield reduction under different data combinations for the selected drought years (2008, 2009, 2011, 2014, 2015, 2016, and 2017). Each model underwent 50 repeated runs. The colored bar repre-sents the mean RMSE, and the error bar represents the standard error across the 50 runs. The RMSE values were calculated based on data standardized using the z-score method.
Figure 4. Model performance for predicting drought-induced yield reduction under different data combinations for the selected drought years (2008, 2009, 2011, 2014, 2015, 2016, and 2017). Each model underwent 50 repeated runs. The colored bar repre-sents the mean RMSE, and the error bar represents the standard error across the 50 runs. The RMSE values were calculated based on data standardized using the z-score method.
Atmosphere 17 00042 g004
Atmosphere 17 00042 g005
Figure 5. Comparison between the spatial patterns of observed and predicted drought-induced yield loss in winter wheat in 2019 using RF model with different vegetation variables as inputs.
Figure 5. Comparison between the spatial patterns of observed and predicted drought-induced yield loss in winter wheat in 2019 using RF model with different vegetation variables as inputs.
Atmosphere 17 00042 g005
Atmosphere 17 00042 g006
Figure 6. Comparison between the spatial patterns of observed and predicted drought-induced yield loss in winter wheat in 2019 using XGBoost model with different vegetation variables as inputs.
Figure 6. Comparison between the spatial patterns of observed and predicted drought-induced yield loss in winter wheat in 2019 using XGBoost model with different vegetation variables as inputs.
Atmosphere 17 00042 g006
Atmosphere 17 00042 g007
Figure 7. Spatial distribution of severity of drought occurred in 2019 (left) and irrigation districts (right) in the five provinces of North China Plain.
Figure 7. Spatial distribution of severity of drought occurred in 2019 (left) and irrigation districts (right) in the five provinces of North China Plain.
Atmosphere 17 00042 g007
Table 1. Data used for yield prediction studies.
Table 1. Data used for yield prediction studies.
TypeVariableSourceSpatialTemporalTime
ResolutionResolutionAcquired
Crop yield Recorded yield of winter wheatStatistical yearbookPrefecture-Yearly2007–2020
level city
Planting areaPlanting area of winter wheatChinaCropArea 1 km1 kmYearly2007–2015
(https://doi.org/10.17632/jbs44b2hrk.2)
(accessed on 30 March 2025)
Vegetation variableSIFGOSIF0.05°Monthly2007–2020
(http://globalecology.unh.edu)
(accessed on 30 March 2025)
VIsMCD43C40.05°Daily2007–2020
(http://ladsweb.nascom.nasa.gov/)
(accessed on 30 March 2025)
Climate variableTmin, Tmax, Ppt, Pet, Vpd, SM, SradTerraClimate datasets1/24°Monthly2007–2020
(https://doi.org/10.7923/G43J3B0R)
(accessed on 30 March 2025)
Soil properties dataPSD-SA, PSD-SI, PSD-CL,Soil particle-size distribution dataset1 kmYearly2007–2020
SOM, CEC, PH, BD(http://globalchange.bnu.edu.cn)
(accessed on 30 March 2025)
SPEISPEI-031-km spatial resolution SPEI dataset across Chinese Mainland from 2001 to 20221 kmMonthly2007–2020
Descriptions: SIF: solar-induced chlorophyll fluorescence; SPEI: standardized precipitation evapotranspiration index; SM: soil moisture; Srad: solar radiation; Tmax: maximum temperature; Tmin: minimum temperature; Vpd: vapor pressure deficit; Pet: evapotranspiration; Ppt: precipitation; PSD-SI: particle size distribution–silt content; PSD-CL: particle size distribution–clay content; PSD-SA: particle size distribution–sand content; SOM: soil organic matter; BD: bulk density; CEC: cation exchange capacity; PH: potential of hydrogen.
Table 2. Classification of meteorological drought for SPEI.
Table 2. Classification of meteorological drought for SPEI.
Degree of DroughtSPEI
Drought-free−0.5 < SPEI
Mild drought−1.0 < SPEI ≤ −0.5
Moderate drought−1.5 < SPEI ≤ −1.0
Severe drought−2.0 < SPEI ≤ −1.5
Extreme droughtSPEI ≤ −2.0
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Hu, H.; Zheng, M.; Niu, Y.; Shen, Q.; Ren, Q.; You, Y. Modeling of Drought-Induced Crop Yield Loss Based on Solar-Induced Chlorophyll Fluorescence by Machine Learning Methods. Atmosphere 2026, 17, 42. https://doi.org/10.3390/atmos17010042

AMA Style

Hu H, Zheng M, Niu Y, Shen Q, Ren Q, You Y. Modeling of Drought-Induced Crop Yield Loss Based on Solar-Induced Chlorophyll Fluorescence by Machine Learning Methods. Atmosphere. 2026; 17(1):42. https://doi.org/10.3390/atmos17010042

Chicago/Turabian Style

Hu, Han, Minxue Zheng, Yue Niu, Qiu Shen, Qinyao Ren, and Yanlin You. 2026. "Modeling of Drought-Induced Crop Yield Loss Based on Solar-Induced Chlorophyll Fluorescence by Machine Learning Methods" Atmosphere 17, no. 1: 42. https://doi.org/10.3390/atmos17010042

APA Style

Hu, H., Zheng, M., Niu, Y., Shen, Q., Ren, Q., & You, Y. (2026). Modeling of Drought-Induced Crop Yield Loss Based on Solar-Induced Chlorophyll Fluorescence by Machine Learning Methods. Atmosphere, 17(1), 42. https://doi.org/10.3390/atmos17010042

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop