Study on the Lagged Response Mechanism of Vegetation Productivity Under Atypical Anthropogenic Disturbances Based on XGBoost-SHAP

Highlights

What are the main findings?

• NPP increased across the BTH region during lockdown, especially in urban cores.
• Vegetation responses shifted toward more immediate climate sensitivity with weakened lag effects.

What are the implications of the main findings?

• Short-term anthropogenic cessation amplified vegetation responsiveness to concurrent environments.
• Insights into spatial–temporal lag mechanisms aid urban ecological regulation under abrupt disturbances.

Abstract

The abrupt COVID-19 lockdown in early 2020 offered a unique natural experiment to examine vegetation productivity responses to sudden declines in human activity. Although vegetation often responds to environmental changes with time lags, how such lags operate under short-term, intensive disturbances remains unclear. This study combined multi-source environmental data with an interpretable machine learning framework (XGBoost-SHAP) to analyze spatiotemporal variations in net primary productivity (NPP) across the Beijing-Tianjin-Hebei region during the strict lockdown (March–May) and recovery (June–August) periods, using 2017–2019 as a baseline. Results indicate that: (1) NPP showed a significant increase during lockdown, with 88.4% of pixels showing positive changes, especially in central urban areas. During recovery, vegetation responses weakened (65.31% positive) and became more spatially heterogeneous. (2) Integrating lagged environmental variables improved model performance (R² increased by an average of 0.071). SHAP analysis identified climatic factors (temperature, precipitation, radiation) as dominant drivers of NPP, while aerosol optical depth (AOD) and nighttime light (NTL) had minimal influence and weak lagged effects. Importantly, under lockdown, vegetation exhibited stronger immediate responses to concurrent temperature, precipitation, and radiation (SHAP contribution increased by approximately 7.05% compared to the baseline), whereas lagged effects seen in baseline conditions were substantially reduced. Compared to the lockdown period, anthropogenic disturbances during the recovery phase showed a direct weakening of their impact (decreasing by 6.01%). However, the air quality improvements resulting from the spring lockdown exhibited a significant cross-seasonal lag effect. (3) Spatially, NPP response times showed an “urban-immediate, mountainous-delayed” pattern, reflecting both the ecological memory of mountain systems and the rapid adjustment capacity of urban vegetation. These findings demonstrate that short-term removal of anthropogenic disturbances shifted vegetation responses toward greater immediacy and sensitivity to environmental conditions. This offers new insights into a “green window period” for ecological management and supports evidence-based, adaptive regional climate and ecosystem policies.

1. Introduction

In early 2020, the COVID-19 outbreak in China prompted the government to implement unprecedented lockdown measures, including the suspension of transportation, shutdown of industrial production, and strict restrictions on population mobility [1]. These measures led to a sharp short-term reduction in human activities, which in turn induced significant changes in various environmental factors. During the lockdown, the population-weighted average concentration of PM_2.5 across China decreased by −15 μg/m³ [2], and aerosol optical depth (AOD) levels also exhibited a pronounced decline [3,4]. Meanwhile, canopy radiation increased by 3.50% in spring 2020 [5], while nighttime light (NTL) showed a decrease followed by a recovery [6]. In addition, springtime temperature (TEM) and precipitation (PRE) in urban areas across China generally decreased [7]. Since vegetation growth is jointly regulated by multiple environmental drivers, its responses often reflect the combined effects of these interacting variables [8]. Therefore, the abrupt environmental perturbations during the lockdown may have triggered ecological responses in vegetation growth. Previous studies have reported that vegetation in China exhibited an “earlier greening” trend during the lockdown, with enhanced vegetation index (EVI) values increasing nationwide [5,9]. Moreover, net primary productivity (NPP) within urban areas increased by 6.50% [10]. Taken together, these findings suggest that the lockdown provided a natural experimental setting to investigate the ecological impacts of abrupt reductions in human activity.

Previous studies have demonstrated that environmental factors exert significant time-lag effects on vegetation, meaning that when environmental changes exceed the adaptive capacity of vegetation growth, the responses may be delayed [11]. At the global scale, experiments linking normalized difference vegetation index (NDVI) with temperature, precipitation, and solar radiation have shown that accounting for lag effects can improve the explanatory power for vegetation growth by 11% [12]. In the high-altitude regions of Tibet, the average lag period of solar radiation on vegetation leaf area index (LAI) was approximately 2.4 months [13]. Clarifying the mechanisms underlying lag effects is critical for improving our understanding of vegetation growth dynamics and ecosystem regulatory processes. However, most existing studies have focused on long-term time series, and the short-term lag responses to sudden disturbances remain insufficiently explored. From a theoretical perspective, anthropogenic pollutants typically act as external stressors that constrain resource availability and physiological efficiency [14]. The abrupt removal of such stressors may trigger a shift in vegetation physiology from a stress-tolerant mode to an active acquisition mode. This transition theoretically allows vegetation to respond more rapidly to current favorable environmental conditions, thereby compressing the traditional lag effects observed under baseline conditions. In particular, under atypical anthropogenic disturbances such as the COVID-19 lockdown, current research has mainly examined the correlations or nonlinear relationships between vegetation and concurrent environmental variables [15,16], while little attention has been paid to how lag effects of environmental drivers on vegetation changed during the lockdown. This knowledge gap constrains our systematic understanding of the ecological response speed, cumulative effects, and recovery processes of vegetation under abrupt shifts in human activity. Therefore, investigating the lagged response mechanisms of vegetation to environmental drivers during the lockdown is essential for identifying the “green window period” of ecological recovery. In this study, we define this concept as a specific time window during which, following a reduction in sudden anthropogenic disturbances (such as lockdowns), the sensitivity of vegetation responses to climatic factors increases and the lag period shortens due to the release of environmental stress.

Existing studies have employed a variety of methods to analyze vegetation drivers and lag effects, including linear regression [17], correlation and partial correlation analyses [18], lagged regression models, and generalized additive models (GAM). For example, a study focusing on Qilian Mountain National Park applied GAM to reveal the lag effects of climatic variables on NDVI across different elevations and vegetation types, demonstrating that incorporating the “optimal lag period” significantly enhanced model explanatory power [19]. In another case, distributed lag nonlinear models (DLNMs) were developed to investigate the global spatial patterns of drought-induced lag effects on terrestrial vegetation [20]. Due to the inherent complexity of ecological data, traditional approaches often struggle to capture potential interactions and intricate temporal dependencies [21]. In contrast, machine learning algorithms, particularly XGBoost (an efficient gradient boosting model), can efficiently process big data and uncover complex nonlinear relationships, demonstrating strong stability and robustness [22]. However, the “black-box” nature of machine learning poses challenges for interpretability [23]. The advent of SHAP (Shapley Additive exPlanations) has addressed this issue by combining the predictive power of machine learning with improved interpretability, enabling the identification of the directionality of drivers and capturing their nonlinear responses, thereby greatly enhancing model transparency [24]. Consequently, by constructing independent features for each lag period, the XGBoost-SHAP framework enables the quantitative comparison of their respective SHAP contributions, thereby identifying the optimal lag time corresponding to the feature with the highest importance.

The Beijing-Tianjin-Hebei (BTH) urban agglomeration is one of the regions in China with the highest levels of economic development, industrial activity, and urbanization, characterized by high population density and intense human activities [25]. During the COVID-19 lockdown in 2020, strict policy enforcement led to substantial reductions in pollutant concentrations [26], making BTH an ideal natural laboratory for examining the ecological impacts of the lockdown. NPP, defined as the difference between carbon absorbed through photosynthesis and carbon released via respiration, directly reflects the productive capacity of vegetation under natural environmental conditions [27]. Vegetation in the BTH region plays a crucial role in carbon sequestration, air quality improvement, and the provision of ecosystem services, and understanding its response mechanisms has important implications for regional ecological planning [28]. Therefore, this study focuses on two policy-consistent periods in 2020: the strict lockdown phase (March–May) and the recovery phase (June–August). These periods correspond to well-documented national control stages with relatively homogeneous intervention intensity. Its objectives are: (1) to quantify the spatiotemporal variations of vegetation NPP in the BTH region before and after the lockdown; (2) to apply an XGBoost-SHAP framework incorporating lagged environmental variables to reveal the explanatory power of major drivers across different lag periods; (3) to identify the temporal dynamics and spatial heterogeneity of vegetation responses under abrupt reductions in human activity, and to delineate potential “green window periods”. This study does not aim to quantify the long-term effects of climate change on vegetation productivity. Instead, it focuses on how the temporal patterns and lag structures of vegetation responses change under abrupt, short-term anthropogenic disturbances. COVID-19 lockdowns constituted a high-intensity yet transient natural experiment [29], offering a unique opportunity to investigate shifts in vegetation responsiveness and ecological memory under atypical conditions. By achieving these aims, this study advances theoretical understanding of ecological responses and recovery mechanisms under short-term disturbances and provides data-driven evidence to support ecological management and policy-making.

2. Materials and Methods

2.1. Study Area

The Beijing-Tianjin-Hebei (BTH) region is located in northern China, spanning 113°27′–119°50′E and 36°05′–42°40′N. It comprises Beijing, Tianjin, and all cities within Hebei Province, and is one of the most densely populated and economically dynamic urban agglomerations in China. Topographically, the northern part of the region is dominated by the Yanshan Mountains, the southern part by the North China Plain, the western part by the Taihang Mountains, and the eastern part by the Bohai Bay [30]. Elevations in the northwest generally range from 500 to over 2000 m, with pronounced relief, whereas the central and southern plains are flat and lie below 200 m. The transition zone between mountains and plains forms a distinct environmental gradient, providing the geographical basis for regional variations in vegetation productivity. The region has a warm temperate, semi-humid to semi-arid continental monsoon climate, with four distinct seasons [31]. The lowest and highest temperatures occur in January and July, respectively, and annual precipitation ranges from 400 to 800 mm, concentrated mainly in July and August.

The central-southern cities of Hebei Province, such as Shijiazhuang, Tangshan, and Handan, as well as Beijing and Tianjin, have a dense population, high levels of urbanization, and strong human activity (Figure 1) [32]. Eastern coastal areas of Hebei, together with industrial hubs such as Tangshan and Handan, host a concentration of heavy industries including steel and coal-fired power, while major transportation nodes such as Tianjin Port play a crucial role in national logistics and energy supply [33]. Forests, grasslands, and croplands across the BTH region jointly constitute an important carbon sink system, which is vital for carbon balance in the North China Plain and even at the national scale [34].

2.2. Data Used

This study was conducted on the Google Earth Engine (GEE) cloud platform, using multiple satellite products to investigate vegetation productivity dynamics during the 2020 lockdown and recovery periods, as well as their lagged responses to environmental factors. A three-year baseline period (2017–2019) was adopted to smooth out the local impacts of extreme events and seasonal anomalies on vegetation [15], and was therefore considered representative of conditions without lockdown influences. In addition, since diurnal variations of parameters during the study period are generally small and trends are not significant, the temporal resolution was standardized to one month. Given the fact that most datasets are available at 1000 m resolution, the spatial resolution was uniformly resampled to 1000 m. All datasets were reprojected to the WGS84/UTM coordinate system. A detailed list of datasets is provided in

Table 1. Datasets catalog introduction.

Dataset	Source	Link
500 m PSNnet	GEE: MOD17A2H	https://lpdaac.usgs.gov/products/mod17a2hv061/ (accessed on 25 October 2025)
1000 m TEM	Qinghai-Tibet Plateau/Third Pole Environment Data Center	http://data.tpdc.ac.cn/ (accessed on 25 October 2025)
1000 m PRE	Qinghai-Tibet Plateau/Third Pole Environment Data Center	http://data.tpdc.ac.cn/ (accessed on 25 October 2025)
5000 m PAR	GEE: MCD18A2	https://lpdaac.usgs.gov/products/mcd18a2v061/ (accessed on 25 October 2025)
0.004° NTL	Environmental Sciences Chinese Academy of Sciences	http://data.tpdc.ac.cn/ (accessed on 25 October 2025)
1000 m AOD	GEE: MCD19A2	https://lpdaac.usgs.gov/products/mcd19a2v061/ (accessed on 25 October 2025)
500 LULC	GEE: MCD12Q1	https://lpdaac.usgs.gov/products/mcd12q1v061/ (accessed on 25 October 2025)

.

2.2.1. Net Primary Productivity Data

Net primary productivity (NPP) is a key component of the terrestrial carbon cycle and directly reflects the productive capacity of vegetation under natural environmental conditions [35]. Since GEE does not provide a ready-to-use monthly NPP dataset, we derived NPP from the net photosynthesis (PSNnet) band. The PSNnet data are obtained from the MOD17A2 product, which provides global vegetation monitoring at a 500 m resolution with an 8-day interval [36]. The accuracy and reliability of this product have been extensively validated worldwide and it has been widely applied in vegetation productivity assessments in China [37]. According to the MOD17A2 algorithm, NPP is calculated as the difference between gross primary productivity (GPP) and autotrophic respiration (Ra), where Ra is further partitioned into maintenance respiration (Rm) and growth respiration (Rg) [38]. PSNnet is defined as GPP minus Rm [39], while Rg can be empirically approximated as 25% of NPP [40]. Therefore, NPP can be derived from PSNnet through a series of equations (Supplementary Materials, Formulas (S1)–(S6)).

2.2.2. Environmental Factor Data

Temperature (TEM) and precipitation (PRE) are two major factors that significantly influence vegetation growth [41]. Monthly TEM and PRE data for China were obtained from the Tibetan Plateau Data Center (https://data.tpdc.ac.cn/) (accessed on 25 October 2025), which provides a 1 km monthly dataset of mean TEM and PRE. This dataset was generated using the Delta downscaling method, combining the CRU TS v4.02 dataset with the WorldClim dataset. Furthermore, its reliability for climate change studies in China has been validated against observations from 496 national meteorological stations [42].

Photosynthetically active radiation (PAR) refers to the portion of solar radiation within the wavelength range of 400–700 nm that can be utilized by vegetation for photosynthesis. The PAR data used in this study were obtained from the MCD18A2 product, a gridded Level-3 dataset in sinusoidal projection with a spatial resolution of 5 km. Validation against ground-based measurements has demonstrated that the MCD18A2 product is reliable and highly accurate [43], and it has been widely applied in radiation assessments across China [44].

The nighttime light (NTL) data used in this study were obtained from the National Oceanic and Atmospheric Administration (NOAA), acquired by the Visible Infrared Imaging Radiometer Suite (VIIRS) onboard the Suomi National Polar-orbiting Partnership satellite [6]. The dataset was further processed by the Resource and Environmental Science Data Center (RESDC), Chinese Academy of Sciences, to generate nationwide monthly NTL brightness data at a spatial resolution of 0.004°.

The aerosol optical depth (AOD) data were obtained from the MCD19A2 V6 gridded Level-2 AOD product, which is a combined dataset from MODIS/Terra and MODIS/Aqua. It is retrieved using the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm and provides daily products at a 1 km spatial resolution [45]. Previous studies have demonstrated a high level of consistency between this product and ground-based measurements in China, and it has been widely applied in air quality research across the country [46].

This study obtained monthly PSNnet data from March to August during 2017–2020, as well as monthly environmental factor data from December 2016 to August 2020. Due to the influence of satellite orbits, weather conditions, and retrieval algorithm limitations, AOD data in the BTH region exhibit unavoidable gaps in spatial coverage, particularly in winter and in areas with high surface reflectance [47,48]. Direct use of these data would result in inconsistencies in spatiotemporal coverage. To avoid potential biases caused by spatial coverage differences among datasets, a unified masking approach was applied across all months and periods. Only pixels valid in all data sources were retained, thereby ensuring the reliability and comparability of subsequent model construction and spatiotemporal analyses.

2.2.3. Land Use and Land Cover Data (LULC)

The data were derived from the “LC_Type1” classification within the MCD12Q1 v006 annual global land cover type dataset, featuring a spatial resolution of 500 m. The processing steps included: reprojection, resampling using the mode calculation method via the reduceResolution function, retention of dominant categories from the original classification data, and masking of the study area. Reclassification was performed according to the IGBP classification system into six categories: forest, cropland, grassland, impervious surfaces, water bodies, and other. Due to the short study time span, 2020 land use data were used for statistical analysis.

2.3. Methods

2.3.1. Difference Analysis

To quantitatively assess the changes in vegetation productivity during the lockdown and recovery periods, this study constructed a difference index (ΔNPP) to represent the relative deviation between 2020 and the baseline period (multi-year mean of 2017–2019). The calculation formula is as follows:

$$∆{\mathrm{N}\mathrm{P}\mathrm{P}}_{\mathrm{i}}={\mathrm{N}\mathrm{P}\mathrm{P}}_{2020,\mathrm{i}}-\overline{{\mathrm{N}\mathrm{P}\mathrm{P}}_{2017-2019,\mathrm{i}}}$$

(1)

where ${\mathrm{N}\mathrm{P}\mathrm{P}}_{2020,\mathrm{i}}$ denotes the NPP value of the i-th pixel during the corresponding period in 2020, and $\overline{{\mathrm{N}\mathrm{P}\mathrm{P}}_{2017-2019,\mathrm{i}}}$ represents the multi-year mean NPP of the same period during 2017–2019. This approach minimizes the influence of interannual fluctuations at the pixel scale, thereby highlighting deviations in 2020 attributable to the lockdown and associated environmental changes.

2.3.2. Construction of Lag Windows

Given that the temporal effects of environmental factors on vegetation at the monthly scale are generally believed to persist for less than one quarter [12,49], and to capture short-term physiological regulatory mechanisms while avoiding cross-seasonal confounding factors [50], this study considered a maximum lag of three months to investigate the lagged responses of vegetation NPP to environmental factors during the COVID-19 lockdown and recovery periods. Accordingly, four lag windows (lag0–lag3) were constructed for both periods, linking NPP with environmental drivers at different time lags. For each window, the values of AOD, PAR, TEM, and NTL were represented by their mean levels, while PRE was represented by its cumulative value. For each environmental factor (e.g., TEM), the four time lag windows are defined as follows:

TEM_lag0: Impact factor level for the current period (for the Lockdown period: March–May; for the Recovery period: June–August); TEM_lag1: Impact factor level for the period one month prior to the current period (for the Lockdown period: February–April; for the Recovery period: May–July); TEM_lag2: Impact factor level of the period two months prior to the current period (for Lockdown period: January–March, for Recovery period: April–June); TEM_lag3: Impact factor level of the period three months prior to the current period (for Lockdown period: December–February, for Recovery period: March–May).

All environmental factor data are organized into four lag scenarios for subsequent model training:

Scenario Lag = 0 (no lag scenario): Composed of the lag0 window for each environmental factor;

Scenario Lag ≤ 1 (one-month lag scenario only): Composed of lag0 and lag1 windows for each environmental factor;

Scenario Lag ≤ 2 (two-month lag scenario only): Composed of lag0, lag1, and lag2 windows for each environmental factor;

Scenario Lag ≤ 3 (three-month lag scenario): Composed of lag0, lag1, lag2, and lag3 windows for each environmental factor.

2.3.3. Extreme Gradient-Boosting Model

To quantify the explanatory power of environmental and anthropogenic factors on vegetation NPP variation, this study employed the Extreme Gradient Boosting (XGBoost) algorithm. XGBoost is an ensemble learning method based on the gradient boosting framework. Its core principle involves iteratively constructing multiple regression trees to progressively refine residuals from previous models, thereby enhancing predictive accuracy [51]. Compared to traditional gradient boosting, XGBoost incorporates regularization parameters to control model complexity and prevent overfitting [52]. It demonstrates advantages in handling multidimensional input variables, capturing nonlinear relationships, and identifying ecological response thresholds through its iterative learning approach [49]. XGBoost is particularly well-suited for predicting and conducting attribution analyses of ecosystems under multifactorial influences [53].

To investigate the dynamic responses of vegetation to environmental factors across different periods, we trained independent model frameworks for four distinct timeframes. Specifically, the models were categorized as follows: (1) 2020 Lockdown Period: Training data comprised NPP from March to May 2020 and corresponding environmental factors, with four lag scenarios (Lag = 0, Lag ≤ 1, Lag ≤ 2, Lag ≤ 3); (2) Baseline-year Lockdown period: Training data comprised NPP from March to May 2017–2019 and corresponding four lag scenarios of environmental factors; (3) 2020 Recovery period: Training data comprised NPP from June to August 2020 and corresponding four lag scenarios of environmental factors; (4) Baseline-year Recovery period: Training data comprised NPP from June to August 2017–2019 and corresponding four lag scenarios of environmental factors.

To ensure optimal model generalization and prevent overfitting, we implemented a rigorous hyperparameter tuning process using Grid Search combined with 5-fold cross-validation. The dataset was randomly partitioned into training (70% of the samples) and testing (the remaining 30%) sets. Within the training set, we optimized the number of trees, maximum tree depth to control model complexity, learning rate to prevent getting stuck in local minima, and subsampling ratios to introduce randomness and enhance robustness [54]. The final model’s performance is evaluated using two metrics: the coefficient of determination (R²) and the root mean square error (RMSE). The hyperparameter combination yielding the lowest average RMSE across the five validation folds was selected as the final model configuration.

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\mathrm{\text{'}}})}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\mathrm{y})}^2}}$$

(2)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\mathrm{\text{'}}})}^2}{\mathrm{n}}}}$$

(3)

where n denotes the sample size, $y_i$ represents the NPP value of the sample, ${\mathrm{y}}_{\mathrm{i}}^{\mathrm{\text{'}}}$ denotes the NPP value simulated by the NPP-XGB model, and $\mathrm{y}$ represents the mean value. R² serves as a key indicator for evaluating model performance, with values ranging from 0 to 1. An R² close to 1 indicates the model possesses good fitting capability and high reliability. RMSE is commonly used to assess the closeness between observed and simulated values, with a range of [0, ∞]. A smaller RMSE value signifies more accurate predictions [23].

2.3.4. Ablation Study

The ablation study is a method used to evaluate the contribution of different components within a machine learning system to its overall performance, aiming to provide a deeper understanding of the role of each module [55]. By systematically removing feature variables one by one and calculating the loss in model performance, the relative importance of each component can be assessed. The procedure is defined as follows:

$${\mathsf{\Delta }\mathrm{R}}_{\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h}\mathrm{o}\mathrm{u}\mathrm{t}\_\mathrm{i}}^2={\mathrm{R}}^2-{\mathrm{R}}_{\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h}\mathrm{o}\mathrm{u}\mathrm{t}\_\mathrm{i}}^2$$

(4)

where ${\mathrm{R}}^2$ represents the coefficient of determination for the model constructed using all feature variables, ${\mathrm{R}}_{\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h}\mathrm{o}\mathrm{u}\mathrm{t}\_\mathrm{i}}^2$ denotes the coefficient of determination for the model after removing variable i, and ${\mathsf{\Delta }\mathrm{R}}_{\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h}\mathrm{o}\mathrm{u}\mathrm{t}\_\mathrm{i}}^2$ indicates the loss in the coefficient of determination for the model after removing variable i. A larger loss value signifies a more significant impact of that variable on the model [56].

2.3.5. Interpretation: Shapley Additive Explanation

To address the “black-box” nature of complex machine learning models and enhance interpretability, this study adopted the Shapley Additive Explanations (SHAP) method. Originating from cooperative game theory, SHAP quantifies the marginal contribution of each input factor to the model prediction at the pixel or sample scale, while distinguishing between positive and negative effects. Compared with traditional feature importance ranking approaches, SHAP has the advantage of providing both a global assessment of factor contributions and revealing local driving mechanisms behind individual predictions [57]. By decomposing model predictions, SHAP accounts for the relationships between input and output variables, partitioning each prediction into a baseline expectation and the average marginal contribution of each input feature (SHAP value). For a given sample i, the prediction can be expressed as:

$${\mathrm{y}}_{\mathrm{i}}={\mathrm{y}}_{\mathrm{b}\mathrm{a}\mathrm{s}\mathrm{e}}+{\sum }_{\mathrm{j}=1}^{\mathrm{M}}{\mathsf{\Phi }}_{\mathrm{i}\mathrm{j}}$$

(5)

where ${\mathrm{y}}_{\mathrm{b}\mathrm{a}\mathrm{s}\mathrm{e}}$ denotes the baseline value, M is the total number of features, and ${\mathsf{\Phi }}_{\mathrm{i}\mathrm{j}}$ represents the SHAP value of feature j for sample i, reflecting the incremental contribution of that feature relative to the baseline prediction. When ${\mathsf{\Phi }}_{\mathrm{i}\mathrm{j}}$ > 0, the feature has a positive effect on the prediction, whereas ${\mathsf{\Phi }}_{\mathrm{i}\mathrm{j}}$ < 0 indicates a negative effect. By calculating the absolute values |${\mathsf{\Phi }}_{\mathrm{i}\mathrm{j}}$| and averaging them across all samples, an overall importance measure for each feature can be obtained.

SHAP assigns each variable a numerical value to quantify its contribution to the vegetation change model, where larger absolute values indicate a stronger influence on the prediction, and the sign reflects whether the variable exerts a promoting or inhibiting effect. In this study, the TreeSHAP framework, specifically designed for tree-based models, was employed to efficiently compute SHAP values. Applying SHAP to the machine learning models with lagged terms not only clarifies the importance of each environmental factor for vegetation NPP, but also illustrates how these effects unfold across different lag periods and reveals their dynamic directions [58]. Additionally, to visualize the specific response mechanisms of vegetation to environmental factors, we generated SHAP dependency curves and interaction term analyses. These results plot the feature values of each pixel on the x-axis and the corresponding SHAP values on the y-axis, thereby revealing the patterns of marginal contribution variation with feature values and the synergistic or antagonistic effects between factors under different scenarios. To identify threshold characteristics, locally estimated scatterplot smoothing (LOESS) was further applied to examine the nonlinear relationships between factor values and their corresponding SHAP values, thereby determining the turning points where effects shift from positive to negative [59].

2.3.6. Spatial Autocorrelation Analysis

Spatial autocorrelation is an important manifestation of spatial dependence, used to assess whether the attribute values of a given element are significantly associated with those of its neighboring spatial units [60]. Spatial autocorrelation can be either positive or negative: positive autocorrelation indicates that neighboring units exhibit similar trends in attribute values, whereas negative autocorrelation suggests that they vary in opposite directions.

Global spatial autocorrelation is primarily used to characterize the spatial distribution patterns of attribute values across the entire study area. Among the various methods for measuring global spatial autocorrelation, Moran’s I is the most widely applied. It quantifies the degree of interdependence among spatial features, with values ranging from −1 to 1. A positive Moran’s I indicates positive spatial correlation; a negative value indicates negative spatial correlation; and a value of zero indicates no significant spatial correlation. Local spatial autocorrelation measures the correlation characteristics exhibited by each spatial feature’s attribute within its local context. By configuring different types of spatial adjacency matrices, it effectively identifies spatial association patterns within local regions. Local Moran’s I is one method for assessing local spatial autocorrelation, defined as a Local Indicator of Spatial Association (LISA) [61].

3. Results

3.1. Spatiotemporal Variation Characteristics of NPP

To compare the spatiotemporal variations of vegetation NPP between the Lockdown and Recovery periods, we calculated the NPP distribution in 2020 and the corresponding multi-year mean (2017–2019), and used the difference index (ΔNPP) to characterize spatial heterogeneity of NPP changes.

During the Lockdown phase (March–May), the spatial distribution of NPP in 2020 and in the baseline years showed a broadly consistent pattern, with plains and suburban areas exhibiting generally higher NPP levels than mountainous and semi-arid regions (Figure 2A). ΔNPP clearly revealed a significant overall increase in NPP under strict lockdown conditions in 2020, with 88.4% of pixels showing positive changes (Figure 2B). Positive responses were concentrated in northeastern cities such as Beijing, Tianjin, Tangshan, Qinhuangdao, and Hengshui, with particularly pronounced increases in densely populated urban fringe areas, especially around Beijing and Tianjin (Figure 2C). In contrast, negative changes were mainly observed in western high-altitude regions, with Zhangjiakou being a typical example. During the Recovery phase (June–Aug), the overall NPP level increased, and the spatial distribution was again similar to the baseline years, with high values concentrated in the western and northern regions, specifically the Yanshan Mountains and Taihang Mountains (Figure 2D). ΔNPP indicated that 65.31% of pixels exhibited positive changes, reflecting an overall increasing trend, though the magnitude was weaker than in the Lockdown phase (Figure 2E). Positive changes were concentrated in southern agricultural areas (Handan, Xingtai, Hengshui), where the positive NPP response was further amplified during this period. In contrast, northeastern coastal areas (Qinhuangdao, Tangshan, Tianjin) experienced a marked negative shift (Figure 2F).

In summary, vegetation in the BTH region exhibited a general increase in NPP during both the Lockdown and Recovery phases of 2020. However, the spatial patterns differed significantly: increases during the Lockdown period were more extensive and stronger, whereas positive effects weakened during the Recovery period and were accompanied by localized reversals. This distinct spatial heterogeneity suggests potential divergence in the underlying driving mechanisms, warranting further quantitative attribution analysis to decouple the specific temporal response patterns. The monthly spatiotemporal variations of NPP and associated environmental factors are presented in Supplementary Materials, Figures S1–S6.

3.2. Performance Enhancement Through Incorporating Lagged Environmental Variables

To examine the lagged effects of environmental factors on NPP variations, we constructed XGBoost models and compared the performance between models without lags (Lag = 0) and those incorporating lagged variables of 0–3 months (Figure 3A–D). The results show that including up to a 3-month lag significantly improved model performance across different periods, as evidenced by the gradual increase in R² (average increase of 0.071) and the concurrent decrease in RMSE (average decrease of 0.457).

Further ablation study, conducted by comparing the ΔR² values after removing lagged terms of each factor, revealed the relative importance of different lagged drivers (Figure 3E). Overall, the ΔR² values of PRE, TEM, and PAR were substantially higher than those of AOD and NTL, indicating that vegetation productivity was primarily driven by lagged climatic factors. During the 2020 Lockdown phase, the lagged effect of PRE was most critical, while the lagged contributions of TEM and PAR were markedly stronger than in the baseline period. In the Recovery phase, PAR and PRE played more prominent lagged roles, with ΔR² values generally higher than those of the baseline. By contrast, the lagged contributions of AOD and NTL remained limited in both periods, suggesting that improvements in air quality and changes in human activity intensity influenced NPP mainly through immediate rather than lagged effects.

3.3. Interpreting Lagged and Threshold Responses of NPP Using SHAP Analysis

Based on SHAP values, the importance of different factors and their lagged terms was ranked, and both mean |SHAP| heatmaps and feature dependence plots were generated to identify the explanatory power of various factors across lag periods and their threshold characteristics (Figure 4 and Figure 5).

During the Lockdown phase, the global SHAP results for 2020 indicated that the immediate effects of all five factors were the strongest. NPP was primarily driven by the contemporaneous effects of TEM (SHAP contribution: 12.83%), PAR (11.29%), and PRE (10.00%), followed by NTL_lag0 and AOD_lag0 (Figure 4A). Based on observations from the feature dependency plot’s smoothed trend lines: Lower values of TEM, PRE, and NTL typically correspond to positive SHAP values (i.e., positive contributions to NPP), while AOD and PAR exhibit nonlinear relationships with NPP. The average |SHAP value| heatmap indicates that instantaneous effects dominate, but TEM and PRE maintain significant explanatory power at lag3 (contributing 8.83% and 7.18%, respectively) and lag2 (10.58%), respectively, suggesting meteorological factors undergo delayed regulation at monthly scales (Figure 4B). Conversely, AOD and NTL demonstrate low explanatory power across all lag periods. PAR maintains high explanatory power at lag2 (10.58%), indicating monthly-scale delayed regulation of meteorological factors (Figure 4B). Conversely, AOD and NTL exhibit low explanatory power across all lag periods, suggesting limited delayed regulation of NPP. Feature dependence curves further reveal threshold characteristics: the TEM trendline crosses zero around 15 °C, where its contribution shifts from positive to negative; PAR exhibits diminishing marginal gains under high radiation conditions (indicating saturation effects); PRE reaches peak contribution at precipitation levels around 70 mm, with SHAP values declining beyond this threshold; and AOD transitions from negative to positive effects at a threshold of 0.4. By comparison, during the same period in 2017–2019, AOD_lag0 (1.006) and TEM_lag0 (0.721) showed the highest explanatory power, while the peak lags of PAR and PRE appeared at lag1 and lag3, respectively, while NTL had only limited influence (Figure 4H,I). This indicates that under multi-year baseline conditions, the immediate effects of air quality in spring and the lagged effects of climatic factors were more prominent.

Overall, the comparison highlights that during the strict lockdown, abrupt reductions in human activities appear to have significantly amplified the immediate response of NPP to environmental drivers, with the contemporaneous effects of temperature, precipitation, and radiation being particularly pronounced (The contribution of the immediate effects of the three factors increased by 7.05% compared to the baseline period). At the same time, the lagged influence of precipitation weakened considerably. Moreover, the explanatory power of NTL increased, reflecting the direct effects of changes in human activity.

During the Recover period, the SHAP results for 2020 indicated that TEM_lag0 (SHAP contribution: 34.18%) had a substantially stronger explanatory power than other factors, followed by PRE_lag0 and PAR_lag0 (11.60% and 10.71%, respectively), with overall weaker lag effects (Figure 5A,B). In contrast, during the same period in 2017–2019, TEM_lag0 (36.31%) also dominated, but the importance of PAR_lag3 and PRE_lag1 was relatively higher (Figure 5H,I). Feature dependence curves reveal shifts in response thresholds compared to the lockdown period: temperature thresholds rose to approximately 25 °C during the 2020 recovery phase. Precipitation effects exhibited greater complexity, with positive contributions concentrated in low-precipitation ranges while medium-to-high precipitation intervals corresponded to negative SHAP values. PAR maintained a consistent positive influence at lag0, though its stimulatory effect approached saturation under extremely high radiation conditions.

The comparison results indicate that during June–August, NPP was primarily regulated by climatic factors. As seasonal temperatures rose, the regulatory effect of temperature became markedly stronger. During the post-lockdown recovery period, the lag effects of climatic factors weakened, while immediate influences became more pronounced, and the explanatory power of precipitation increased (The instantaneous effect contribution of the three factors TEM, PRE, and PAR increased by 22.37% compared to the period of strict lockdown). In contrast to the strict lockdown period, the contributions of NTL and AOD declined (The combined contribution of the two decreased by 6.01%). Moreover, the relatively higher explanatory power of AOD_lag3 compared with its immediate term indicates that air quality improvements during the spring lockdown exerted a delayed regulatory influence on NPP during the recovery period.

3.4. Spatial Heterogeneity of NPP Response Time

To reveal the spatial heterogeneity of vegetation NPP lag responses to environmental factors, we identified the optimal lag time with the strongest explanatory power for each pixel based on SHAP values and mapped the spatial distribution of the optimal lag period for each factor (Figure 6 and Figure 7). Additionally, the statistics on the proportion of lag response times across different land use types are presented in Table S1.

During the 2020 lockdown period (Figure 6A–F), meteorological factors PAR and TEM generally exhibited shorter lag times (lag0–lag1), concentrated in the central and eastern plains; whereas PRE response times showed spatial differentiation. The instantaneous effect of PRE dominated the northeastern cities of Tangshan and Qinhuangdao and, together with AOD, jointly controlled NPP variations in the southern cities of Shijiazhuang, Xingtai, and Handan. The northwestern mountainous areas showed more pronounced lag effects, with TEM_lag2-3 dominating northern Chengde and PAR_lag2-3 being most influential in Zhangjiakou. NTL and AOD were primarily characterized by instantaneous effects (>75%), with NTL dominating densely populated urban areas such as Beijing and Tianjin. Single-factor analyses revealed that PAR and TEM were mainly driven by instantaneous effects (60–70%), while PRE_lag3 explained a larger portion of NPP variation (36.9%) in the central plains. In comparison, during the 2017–2019 baseline (Figure 6G–L), AOD (92%) and TEM (76.7%) exhibited stronger instantaneous effects, with TEM dominating the northern Yanshan Mountains and AOD prevailing over the southern plains. The 1–3 month lag effects of PRE and PAR were more spatially extensive, particularly around the suburban areas of Beijing and the Bohai coastal region.

Compared with baseline, the overall lag effects of meteorological factors weakened during the lockdown period, while instantaneous responses became more pronounced. Specifically, the immediate effects of PAR and PRE expanded markedly across the eastern and southern regions, indicating a more rapid vegetation response to changes in radiation and precipitation. In contrast, TEM exhibited stronger lag effects along the Yanshan Mountains, likely reflecting the buffering influence of high-altitude climatic conditions. Meanwhile, the instantaneous effect of NTL intensified substantially, particularly around major urban centers, highlighting the strengthened immediate response of vegetation productivity to abrupt reductions in human activity. Conversely, AOD showed enhanced 1–3 month lag effects in some areas, especially across the plains, which suggested that the ecological responses to improved air quality occurred with a temporal delay. Overall, instantaneous effects dominated in plains and urban areas (with immediate effects from dominant impervious surfaces increasing by nearly 2% compared to baseline periods), whereas mountainous regions exhibited stronger lag responses (delayed effects from dominant forest and grassland areas increasing by 2.34% and 2.37%, respectively) (Table S1).

During the 2020 Recover phase (Figure 7A–F), instantaneous effects of TEM dominated most areas across the BTH region, while immediate influences of PRE and PAR were particularly pronounced along the Taihang Mountains and within the Beijing-Tianjin-Langfang urban cluster. AOD exhibited dominant lag0 effects in central core cities and eastern coastal areas but showed 2–3 month lag responses across approximately 52.4% of the region, suggesting a delayed vegetation response to air quality changes. The instantaneous effect of NTL was mainly distributed in suburban zones expanding outward from urban centers, reflecting the rapid vegetation response to the resumption of human activities. In contrast, during the baseline period (Figure 7G–L), TEM remained the dominant instantaneous factor, while PAR and PRE exhibited stronger lagged effects over Beijing, Tianjin, and the Taihang Mountains. Single-factor analyses of AOD, PAR, and NTL during the baseline period all indicated predominant 2–3 month lags. From Jun to Aug, instantaneous effects under different land use types showed an upward trend compared to the baseline period (Table S1). Comparing the lockdown and recovery phases, vegetation consistently exhibited high instantaneous sensitivity to meteorological factors, with the instantaneous dominance range of temperature further expanding during recovery. However, the response patterns of anthropogenic factors AOD and NTL underwent significant spatiotemporal shifts, transitioning from instantaneous dominance during lockdown to pronounced lag (2–3 months) across more than half of the regions, predominantly distributed in non-core urban areas.

Overall, the response time of NPP to environmental factors exhibited a pattern of “immediate effects in urban areas and delayed effects in mountainous regions”. During the Lockdown period, the immediate effects of meteorological factors were markedly enhanced, while mountainous areas still retained some lagged responses, and NTL and AOD primarily manifested as immediate controls. In the Recovery period, the dominant role of temperature became even more pronounced, precipitation and radiation mostly showed immediate effects, whereas AOD and NTL displayed lagged responses in certain coastal and suburban areas.

3.5. Spatial Clustering of Response Characteristics

To examine whether lagged responses exhibit spatial structuring, this study conducted global and local spatial autocorrelation analyses based on the “strongest lag period” and the “dominant factor-lag combination”, and further calculated the proportion of dominant types and the mean lag period by administrative region (Figure 8).

During the 2020 Lockdown period, the global Moran’s I was significantly positive (0.850), indicating that lag periods exhibited a spatially clustered distribution (Figure 8(A1)). LISA results revealed two main high-high (HH) clusters: one spanning the urban corridor of central Beijing, downtown Tianjin, and Tangshan-Qinhuangdao. And the other covering the Yanshan-Taihang Mountains and the northern Hebei hilly regions, indicating that both urban and mountainous ecosystems displayed monthly-scale delayed responses under strict lockdown. By administrative region, central urban areas were dominated by lag0, whereas northern mountainous Zhangjiakou was primarily characterized by lag2 (Figure 8(A4)). In terms of the two-dimensional factor-lag combination (Figure 8(B1–B4)), the global Moran’s I was also significantly positive, showing strong spatial autocorrelation. TEM mainly concentrated its influence in northern Hebei cities such as Chengde, Beijing, and Baoding, with stronger lag effects in mountainous areas, while AOD and NTL instantaneous effects notably influenced the western hilly regions including Shijiazhuang, Xingtai, and the Tianjin Binhai area. In contrast, the 2017–2019 baseline period, although also significant, showed weaker overall clustering (Figure 8(C1–C4,D1–D4)), with HH patches more fragmented, suggesting that the lockdown amplified the spatial structuring of lagged effects, particularly in the Yanshan-Taihang mountain regions.

During the 2020 Recovery period, the global Moran’s I remained positive (0.570 and 0.853). Local spatial autocorrelation analysis revealed that while the instantaneous effect (lag0) remained dominant overall, its High-High clusters exhibited a more dispersed spatial distribution (Figure 8(C2,C3)). Notably, the instantaneous effect of PRE formed significant clusters in the Beijing-Tianjin-Tangshan urban agglomeration and along the Taihang Mountains (Figure 8(C5–C8)). In the 2017–2019 baseline period, spatial clustering was also evident, with a pronounced lagging effect cluster evident in the Yanshan Mountains region (Figure 8(D1–D8)). Overall, during the recovery period, lag periods generally exhibited low-value clustering, while the instantaneous effect of precipitation was enhanced in the coastal industrial belt.

4. Discussion

The strict lockdown measures implemented during the COVID-19 pandemic provided a unique natural experiment to investigate vegetation responses to sudden changes in human activities and environmental conditions. By integrating multi-source remote sensing data, machine learning models, and interpretable SHAP analyses, this study revealed the spatiotemporal heterogeneity and temporal lag effects of various environmental factors on NPP in the BTH region. These findings should not be regarded as an alternative explanation for long-term vegetation-climate relationships, but rather provide new insights into understanding the mechanisms of vegetation dynamics under short-term intense anthropogenic disturbances.

4.1. Lagged Vegetation Responses Under Changes in Human Activities

The variations in lagged effects among different influencing factors and their underlying mechanisms are crucial for understanding vegetation dynamics. Our ablation experiments and SHAP analysis consistently indicate that meteorological factors (PRE, TEM, PAR) play a dominant role in driving NPP changes, with their lagged effects being significantly more important than those of AOD and NTL (Figure 3, Figure 4 and Figure 5). This finding aligns with previous studies emphasizing that water-heat conditions and radiation predominantly control vegetation productivity at regional scales [62].

During the lockdown period (March–May), the immediate effects (Lag0) of TEM, PAR, and PRE in 2020 were the strongest drivers of NPP (Figure 4A,B). The enhanced immediate response of meteorological factors may be attributed to unusually clear skies and reduced aerosol pollution, which likely strengthened the direct influence of solar radiation and temperature on photosynthesis during the early stages of large-scale lockdowns [5]. This finding is consistent with previous research, which similarly observed that the reduction in aerosol optical depth during the lockdown period in eastern China significantly enhanced the increase in total primary productivity of vegetation [63]. Furthermore, the significant increase in NTL_lag0 SHAP contribution during the 2020 lockdown period compared to the baseline period (Figure 4A) should primarily be interpreted as a proxy for changes in human activity intensity. This reflects the abrupt alleviation of anthropogenic disturbances (e.g., traffic congestion, industrial emissions) resulting from lockdown measures, thereby reducing external environmental stresses on urban vegetation. It highlights the indirect promotion of vegetation recovery through the suspension of human activities [64]. Interestingly, compared with the 2017–2019 baseline, the lagged effects of PRE during the 2020 lockdown were weaker than its immediate effects, while the overall immediate effects of meteorological factors increased (Figure 4). This suggests that under conditions of sharply reduced human activity, vegetation growth became more sensitive to concurrent meteorological conditions, possibly due to the removal of confounding anthropogenic disturbances. Additionally, the observed shortening of lag time can be attributed to the fact that under baseline conditions prior to lockdown, atmospheric pollution loads may have imposed long-term stress on photosynthesis, compelling vegetation to allocate resources toward defense mechanisms (manifested as delayed responses) [65]. The abrupt reduction in human activity alleviated these constraints, effectively unlocking the physiological potential of vegetation. With the removal of light limitations and the easing of pollution stress, vegetation rapidly shifted to a resource-utilization state [66], synchronizing its growth with the improved immediate climatic conditions.

During the recovery period (June–Aug), the dominance of TEM_lag0 was further reinforced in both 2020 and the baseline years, reflecting that temperature during the peak growing season remains a key limiting factor for vegetation growth (Figure 5A,H) [67]. However, the lagged effects of PAR and PRE were still significant in the baseline period (Figure 5I), whereas in 2020, the immediate effects of these factors became more prominent (Figure 5B). This shift toward immediate effects during the post-lockdown recovery phase may indicate a compression of ecological response time, possibly due to the cumulative effects of favorable conditions, such as improved air quality during the lockdown, preconditioning the vegetation to respond more rapidly to current weather [68]. Notably, compared with immediate effects, the lagged effects of AOD, particularly AOD_lag3, became more pronounced during the 2020 recovery period (Figure 5A), suggesting that air quality improvements during the spring lockdown had a delayed positive impact on summer vegetation growth, potentially by reducing diffuse radiation fertilization effects or alleviating pollution-induced physiological stress on plants [69,70].

The spatial heterogeneity of lag effects exhibited a distinct pattern of “immediate responses in urban areas and delayed responses in mountainous regions” (Figure 6 and Figure 7). This ecosystem-type-based lag difference aligns with a global-scale study indicating significant variations in the lag response time to climatic factors across different vegetation types, with forest ecosystems typically exhibiting longer lags than croplands or grasslands [71]. This pattern can be attributed to differences in ecosystem types, vegetation characteristics, and environmental conditions. Vegetation in highly urbanized and human-activity-intensive areas tends to respond more sensitively to environmental changes, particularly during the lockdown, when abrupt reductions in anthropogenic activities enhanced the immediacy of ecosystem responses. In metropolitan cores such as Beijing and Tianjin, where human activity intensity is high [72] and the lockdown impact was strongest, vegetation mainly consists of urban green spaces and managed farmlands [73]. These vegetation types are directly influenced by management practices, microclimatic fluctuations, and immediate human activities (dominated by NTL_lag0), resulting in an average lag time close to zero months.

In contrast, in the Yanshan-Taihang mountainous region (e.g., Zhangjiakou and Chengde), ecosystems dominated by natural forests and grasslands exhibited stronger “ecological memory” effects [74]. The cold and humid conditions in these high-elevation regions slow biochemical processes such as soil nutrient mineralization and plant phenology, leading to clear lagged responses (dominated by TEM_lag1-2 and PRE_lag1-2) with an average lag of 1–2 months. Along the coastal industrial belt (Qinhuangdao-Tangshan-Tianjin Binhai), the dominant factors were spatially fragmented during the lockdown but shifted toward stronger control by AOD and PRE during the recovery period (Figure 7D), consistent with their enhanced explanatory power revealed by the SHAP analysis. This may relate to the significant negative anomalies in both AOD and PRE during that period (Supplementary Materials, Figures S1 and S5). In the agricultural zone of southern and central Hebei (Handan, Xingtai, Hengshui), PRE_lag0 became the dominant factor during the lockdown (Figure 6E), likely due to a significant decrease in precipitation (Supplementary Materials, Figure S6), which constrained positive NPP changes (Figure 2). During the recovery phase, PAR_lag0 dominance and a shortened average lag time (Figure 7C) reflected the rapid growth response of cropland ecosystems to favorable light and heat conditions once agricultural activities resumed, a hallmark of managed agroecosystems [75]. Furthermore, in regions beyond the central urban cores, both AOD and NTL exhibited pronounced lagged effects during the recovery period, suggesting that the residual impacts of improved air quality during the spring lockdown propagated outward to surrounding areas, exerting delayed positive influences on vegetation growth in the summer.

This spatial heterogeneity in vegetation response times helps to explain the observed ΔNPP patterns (Figure 2). The widespread positive ΔNPP during the lockdown period, particularly across urban areas and the northeastern plains, aligns well with immediate positive responses to improved radiation levels (PAR) and reduced human disturbance (NTL). In contrast, the negative ΔNPP observed in western mountainous regions may reflect lagged effects of pre-lockdown adverse conditions, such as consecutive temperature declines (Supplementary Materials, Figure S4). During the recovery phase, the southward shift of positive ΔNPP toward agricultural zones corresponded with shortened lag times and strengthened immediate responses of PRE and PAR, suggesting that cropland ecosystems rapidly benefited from favorable climate conditions and the resumption of agricultural activities. Although SHAP analysis confirms that temperature and precipitation remain the dominant drivers of NPP, the temporal structure of vegetation response is significantly modulated by the intensity of anthropogenic disturbance. If purely meteorological anomalies were at play, their effects should exhibit regional consistency [76]. However, we observe pronounced spatial differentiation: the elimination of lag effects is confined to urban and agricultural areas experiencing abrupt reductions in human activity, while mountain ecosystems with lower human disturbance retain the lag characteristics observed during the baseline period. The inconsistency between the regional homogeneity of meteorological anomalies and the pronounced spatial variation in vegetation responses provides compelling evidence that the anthropogenic pressure release triggered by lockdowns—rather than isolated climate fluctuations—is the primary driver behind the enhanced immediate responsiveness of urban vegetation. These findings highlight that accounting for the spatial heterogeneity of lag effects is crucial for accurately interpreting and predicting vegetation dynamics under abrupt environmental changes. Short-term disturbances can reveal underlying ecosystem response mechanisms often masked under long-term equilibrium conditions, highlighting shifts in response immediacy and ecological memory, thereby providing complementary insights to traditional long-term climate studies. From an ecological management perspective, identifying such “green window periods” can inform the development of targeted strategies. For instance, early-spring pollution control measures could maximize vegetation growth benefits in densely populated urban and industrial areas, while the delayed diffusion of air-quality improvements could synergize with favorable light and thermal conditions to enhance vegetation productivity in southern agricultural regions during summer. Meanwhile, the delayed responses of high-elevation ecosystems, driven by topographic isolation, should also be incorporated into adaptive management and restoration planning.

4.2. SHAP Analysis and Interactions Among Environmental Factors

SHAP analysis not only quantified the relative importance of influencing factors but also revealed key nonlinear thresholds and interaction effects, deepening our understanding of the underlying driving mechanisms [77]. During the lockdown period in 2020, feature dependence plots (Figure 4C–G) identified several critical thresholds. For example, TEM_lag0 exhibited a clear turning point around 15.03 °C, beyond which its effect on NPP shifted from positive to negative, suggesting the onset of heat stress. This threshold aligns with previous findings that elevated temperatures can damage photosynthetic machinery and increase respiratory carbon loss [78]. Notably, this negative temperature effect was amplified under high AOD_lag1 levels, indicating that combined heat and pollution stress exacerbates productivity loss [79]. Similarly, PAR_lag0 showed a nonlinear relationship with NPP. SHAP values rose rapidly at first, plateaued at moderate levels, and declined under extremely high radiation, reflecting both a saturation effect and potential photoinhibition under excessive light intensity [80].

The interaction effects further revealed that under high cumulative radiation conditions (represented by interaction terms), the negative trend of NPP response became steeper. The positive effect of PRE_lag0 on NPP was effective only below the threshold of approximately 70.72 mm, alleviating water limitation; however, beyond this level, its effect turned negative, likely due to reduced solar radiation under cloudy conditions or soil waterlogging causing root hypoxia [81]. Additionally, under high AOD conditions, the positive impact of moderate precipitation was more pronounced, possibly due to the combined effects of atmospheric cleansing by rainfall and enhanced soil moisture availability [82]. Interestingly, AOD_lag0 itself exhibited a threshold around 0.4, where its effect shifted from negative to positive, and this positive influence was amplified under warmer temperature conditions. Such a complex response suggests that, beyond a certain level, aerosols may exert a fertilization effect through enhanced diffuse radiation, particularly when temperatures are favorable [83,84]. This phenomenon aligns with global-scale research indicating that moderate aerosol loads can enhance photosynthetic efficiency by increasing diffuse light penetration within the canopy [85], although the overall effect of AOD remains limited compared to meteorological factors. Meanwhile, NTL_lag0 consistently showed a slight negative correlation with NPP, likely reflecting the general inhibitory effect of intense human activities on vegetation growth. Additionally, it is worth noting that both temperature and precipitation exhibit significantly enhanced interactions with air quality levels lagged by one month. Furthermore, the positive effect of radiation on NPP gradually intensifies before reaching a certain threshold. This reflects the profound impact of the concentrated lockdown measures implemented prior to spring 2020 on both air quality and radiation levels.

During the 2020 Recover period (Figure 5C–G), the threshold for the negative temperature effect increased to approximately 25 °C, consistent with findings indicating a seasonal adaptive shift in the optimal temperature for photosynthesis in Northern Hemisphere vegetation [86]. The interaction between TEM_lag0 and AOD_lag0 was stronger, with higher AOD exacerbating the negative impact on NPP. High temperatures intensify air pollution, highlighting the severity of combined summer heat and poor air quality stress [87]. Precipitation (PRE_lag0) exhibits negative effects on NPP within a specific range (approximately 332–437 mm), but shifts to positive effects outside this range, with a significant interaction with PAR. This suggests that excessive rainfall during the growing season may not be beneficial if it comes at the expense of solar radiation, as it impairs photosynthesis. AOD_lag3 exhibited higher explanatory power than AOD_lag0 during the 2020 recovery period (Figure 5A), strongly supporting the delayed positive impact of improved air quality during the spring lockdown on summer vegetation growth.

4.3. Limitations

First, although AOD data represent a high-quality product, unavoidable missing values persist, particularly during winter and in regions with high reflectance. While restricting analysis to valid pixels ensures robustness, this slightly reduces spatial coverage. Future research may employ data fusion techniques to generate gap-free AOD datasets. Second, this study focuses on monthly-scale data and lag effects up to three months. Finer temporal resolution and longer lag periods may reveal additional short-term dynamics or cross-seasonal cumulative effects. Future research should extend this baseline period to rigorously account for interannual climate variability and ensure the reference state is climatologically representative. Furthermore, while the XGBoost model demonstrates high performance, it remains a correlation-based approach. Although SHAP enhances interpretability, causal inferences should be made with caution. In addition, the COVID-19 lockdown served as an unplanned ‘natural experiment’ rather than a strictly controlled manipulation. While we utilized a multi-year baseline to isolate anomalies, this observational approach cannot fully eliminate the influence of unmeasured confounding variables or the inherent contingency of climate variability specific to 2020. Therefore, the observed shifts in vegetation responsiveness should be interpreted as a case study of abrupt disturbance rather than a generalized rule. Future research could benefit from integrating process-based ecosystem models to construct rigorous counterfactual scenarios, thereby further validating the causal mechanisms between anthropogenic cessation and ecosystem recovery. Moreover, employing such process-based models (e.g., CASA) would address the potential bias arising from the simplified fixed-coefficient method used for NPP inversion, thereby refining the quantitative accuracy of respiration estimates. Additionally, at the 1 km resolution, the mixed-pixel effect in highly urbanized areas may dilute vegetation signals with non-vegetation surfaces (e.g., impervious layers), potentially confounding the observed immediate physiological responses with the rapid thermal properties of built environments. Regarding the characterization of anthropogenic disturbance, most socioeconomic indicator datasets (such as population migration data) are typically aggregated at the administrative unit level, resulting in a significant spatial scale mismatch with the grid resolution of remote sensing analyses. Future research could employ spatial downscaling techniques to integrate these statistical socioeconomic data with raster datasets, thereby constructing more multidimensional indices of human activity intensity. Finally, this study focuses on the Beijing-Tianjin-Hebei region; the generalizability of findings to other areas with different climatic zones and vegetation types requires further investigation.

5. Conclusions

This study quantitatively analyzed the delayed response characteristics of vegetation NPP in the Beijing-Tianjin-Hebei region during the COVID-19 lockdown and recovery periods. Based on multi-source remote sensing data and the XGBoost-SHAP method, the following key conclusions were drawn:

(1)

Phased spatiotemporal responses: During the strict lockdown period in 2020, net primary productivity showed a significant increase, with 88.4% of pixels exhibiting positive changes. Anomalies were primarily concentrated in urban core areas such as Beijing and Tianjin. In contrast, the recovery phase exhibited weaker overall growth and greater spatial heterogeneity.

(2)

Lag patterns reshaped by anthropogenic disturbances: The lockdown amplified the immediate effects of climatic factors (TEM, PRE, PAR) while markedly weakening their lagged contributions compared to the baseline, with instantaneous effects increasing by 7.05%. Anthropogenic factors (NTL) and aerosols (AOD) primarily exerted instantaneous influences.

(3)

Nonlinear thresholds and interaction mechanisms: SHAP analysis revealed critical physiological thresholds for climatic drivers (e.g., temperature turning points) and identified synergistic interactions, such as the combined stress of high temperatures and high aerosol loading, which exacerbated NPP inhibition.

(4)

Distinct spatial divergence in response timing: Vegetation exhibited an “urban-immediate, mountainous-delayed” pattern. Urban and agricultural vegetation responded rapidly to improved air quality and enhanced radiation, whereas mountain ecosystems retained prolonged lagged effects driven by ecological memory.