Agriculture has long been regarded as a high-risk sector, and early studies attributed vulnerability to seasonal climate variability, pest and disease outbreaks, and price uncertainty (Ifejika Speranza, 2013). Over time, interacting risks have exposed agricultural systems to multiple shocks, including extreme weather, supply-chain reconfiguration, and institutional change (Berardi et al., 2011; Gascón and Mamani, 2022; Sundstrom et al., 2023). Disruptions to global food supply chains during the COVID-19 pandemic highlighted a critical point: conventional risk-resistance frameworks are often insufficient to explain collapse and delayed recovery under successive shocks. Agriculture should not be viewed as a passive recipient of risk; rather, it is a dynamic system with capacities for adaptation, reorganization, and transformation (Zampieri et al., 2020; Quandt, 2023). Accordingly, resilience assessment should incorporate resistance capacity, adaptation capacity, and innovation capacity (Zeng et al., 2025), which are discussed in detail in Section 3.

Consequently, maintaining stability, facilitating recovery, and enabling adaptation under uncertainty has become a central agenda in resilience research (Sundstrom et al., 2025). The literature can be broadly grouped into two strands. The first emphasizes deepening vulnerability driven by external shocks, focusing primarily on climatic hazards such as droughts and floods that frequently reduce yields and generate economic losses (Carter et al., 2018). The magnitude of these impacts depends on the resources and technologies that agricultural systems can mobilize in response to emerging risks (Ward, 2022; Das et al., 2025). The second strand highlights resilience deficits arising from internal resource or structural constraints—for example, population pressure, limited land resources, fragmented sectoral organization, and pronounced regional heterogeneity in China (Bizikova et al., 2019; Li and He, 2025). While documenting China’s vulnerabilities, existing studies also underscore potential pathways to strengthen resilience, including technological innovation and agricultural digitalization, infrastructure investment, subsidy and disaster insurance programs, and industrial restructuring (Luo et al., 2024; Zhang et al., 2025; Dong et al., 2025).

Quantifying agricultural industry resilience is a prerequisite for uncovering its underlying mechanisms. Three approaches dominate the literature: the core-variable approach, composite index construction, and systemic risk methods. First, the core-variable approach follows the tradition of Rose (2007) and Martin (2012) by tracking the relative deviation of key variables that are highly sensitive to shocks before and after events. While straightforward and conducive to cross-country and intertemporal comparisons, it oversimplifies system complexity and may fail to capture overall elasticity under complex economic fundamentals (Belhadi et al., 2024). Second, composite index methods are widely used. They integrate multidimensional sub-capacities into a unified framework via entropy–TOPSIS, principal component analysis, or factor analysis to provide static measurement and cross-sectional comparisons (Xiao et al., 2020; Xie et al., 2024; Wang et al., 2025). Although useful for synthesizing key macro variables and offering a systematic perspective (Bolson et al., 2022), such methods often overlook nonlinear responses of micro-level agents (e.g., households and firms) near thresholds (Huang et al., 2023). Third, systemic risk methods employ impulse-response analysis and complex networks to characterize contagion intensity, duration, and cascading pathways (Zhou et al., 2020; Chen Y. et al., 2025). Overall, existing measurement frameworks remain relatively narrow and often cannot identify which indicators matter most or how indicators mechanistically relate to resilience, leaving apparent paradoxes (e.g., “high index–low recovery” or “high risk–high innovation”) insufficiently explained.

The rise of big data and artificial intelligence has accelerated the use of algorithms such as random forests, backpropagation neural networks, and support vector machines (SVM) in resilience research (Tang et al., 2025). Compared with traditional approaches, these tools can identify latent patterns in high-dimensional data with complex nonlinearities and interaction effects without imposing a priori functional forms (Peters et al., 2014; Wang et al., 2020). They have achieved notable progress in ecological applications (e.g., forests and wetlands) and urban resilience assessment (Wu et al., 2025; Yin et al., 2025). Building on this, XGBoost improves predictive performance through iterative residual learning and regularization, while mitigating overfitting, and has been shown to be particularly suitable for agricultural contexts characterized by “small samples and high-dimensional features” (Alam et al., 2025). Nonetheless, the black-box nature of many machine-learning models has limited researchers’ ability to interpret internal mechanisms (Kim and Kim, 2022). The integration of SHAP with machine-learning models addresses this challenge by combining predictive accuracy with process transparency (Lundberg and Lee, 2017). The resulting framework can identify key thresholds and quantify heterogeneous responses across regions and agents, providing an interpretable evidence base for targeted interventions. Although this approach has been applied to crop yield prediction and risk-factor diagnostics in agriculture (Martin et al., 2024), it has not yet been systematically extended to agricultural resilience, leaving substantial room for methodological and empirical advancement.

Despite significant advancements in defining and measuring agricultural resilience, two critical issues remain largely overlooked in the literature. First, existing evaluation frameworks are often predicated on linear assumptions, utilizing mainstream measurement methods such as entropy-weighting to measure and rank resilience (Xiao et al., 2020; Xie et al., 2024; Wang et al., 2025). However, simple linear weighting approaches face substantial challenges in accurately reflecting high-dimensional and structurally complex resilience data. Second, regarding the identification of driving mechanisms, existing literature frequently relies on pre-specifying mechanism variables for verification and interpretation (Luo et al., 2024; Zhang et al., 2025; Dong et al., 2025). There is a notable lack of tools capable of automatically traversing feature importance and discovering latent mechanisms within the data.

Consequently, this study aims to address the following research gaps. First, to construct a measurement framework capable of handling high-dimensional, non-linear data without sacrificing accuracy, applied to the estimation of complex and diverse agricultural resilience data across China’s 31 provinces. Second, to reveal the non-linear relationships and threshold effects of key driving factors, thereby fully explaining how these factors influence variations in agricultural resilience. Third, to comprehensively and systematically assess the changes in resilience across different geographical regions in China over a long period, identifying regional imbalances and dynamic trends.

In summary, this study constructs a comprehensive analytical framework integrating XGBoost and SHAP methods. This framework optimizes the measurement of high-dimensional structured data while simultaneously elucidating the non-linear relationships and threshold effects between resilience and its driving mechanisms, thereby providing a more accurate measurement standard and a deeper, mechanism-oriented explanation for the disaster resistance capability of China’s agricultural industry.

Figure 1 summarizes the research procedure. First, we preprocess and validate the data, establish a three-capacity indicator system, and conduct outlier treatment, Pearson correlation tests, and dataset partitioning. Second, we measure the resilience index using a combined “entropy-weighting and coefficient of variation” method and examine its spatiotemporal evolution. Third, we feed high-dimensional structured data into multiple machine-learning models, select the best-performing XGBoost specification, and refine the resilience index accordingly. Finally, we apply SHAP to identify key drivers and nonlinear relationships, thereby providing a mechanism-based interpretation of the observed spatiotemporal patterns.

The XGBoost model is configured as follows: max_depth = 5, learning_rate = 0.1, min_child_weight = 1, and subsample = 0.95, while colsample and num_parallel_tree are kept at their default values of 1. These hyperparameters are chosen to balance model flexibility and generalization. The learning rate controls the step size of parameter updates during boosting; a smaller value typically requires more boosting iterations to achieve comparable performance. Tree depth determines the complexity of each base learner: deeper trees can capture richer nonlinear patterns but require more data and computation and may increase the risk of overfitting. The number of boosting iterations governs overall model complexity; an appropriate number of rounds improves learning capacity without excessive variance.

To examine the model’s performance on the full dataset under this configuration, we apply five-fold cross-validation (Ye et al., 2023). The sample is split into five subsets; in each fold, four subsets are used for training and the remaining subset is used for testing, repeated five times so that each subset serves as the test fold once. The MAE values across folds are 0.0012, 0.0010, 0.0011, 0.0009, and 0.0011, with an average MAE of 0.0011. The consistently low errors indicate that XGBoost performs well and generalizes favorably in this dataset.

Based on the trained XGBoost model, we input the resilience indicators for 2000–2022 as features to predict RCAI. The rationale for this optimization lies in the generalization capability of machine learning algorithms. Unlike the static calculation of the composite weighting method, XGBoost captures the complex, non-linear mapping relationships between the multidimensional input indicators and the resilience capacity through iterative residual learning. By generating predictions from this trained model, we obtain a “re-estimated” index that reflects the structural regularities of the agricultural system rather than raw arithmetic aggregations. This process effectively filters out random noise and anomalies inherent in the initial data, ensuring that the index represents the systematic resilience level.

The resulting predictions constitute the refined index for the resilience of China’s agricultural industry, as reported in Table 3. Two features are noteworthy. First, the refinement mainly involves local adjustments rather than wholesale changes: the predicted values represent modest numerical corrections, and provincial rankings by resilience remain broadly stable. Second, the refined RCAI effectively mitigates bias induced by extreme values. For instance, in Sichuan in 2008, disaster relief-related variables surged following the Wenchuan earthquake, yielding an abnormally high value of 0.2124 under the original composite-weight method. In contrast, the XGBoost-based prediction is 0.0319—still relatively high, but no longer inflated by orders of magnitude relative to other provinces. This demonstrates that the XGBoost-refined RCAI is less sensitive to outliers and provides a more reliable measure when high-dimensional indicators and extreme observations coexist.

Accordingly, this step addresses the first concern raised in the Introduction: conventional index-construction approaches may struggle to maintain robustness and plausibility in the presence of high-dimensional data and extreme values. Having obtained the refined RCAI, we next use it to identify key drivers and interpret the mechanisms underlying resilience dynamics.

To address the black-box nature of machine-learning models, we employ SHAP to identify the key determinants of RCAI and to interpret their underlying mechanisms. Figure 6 presents (a) the SHAP feature-importance (contribution) plot and (b) the SHAP summary plot. As shown in Figure 6a, IAFA (agricultural fixed-asset investment) accounts for 30.6% of the total contribution and ranks first among all indicators, indicating that IAFA plays a pivotal role in shaping RCAI and is the most influential driver of variation in the resilience of China’s agricultural industry. A plausible explanation is that fixed-asset investment accelerates the construction and upgrading of agricultural infrastructure, thereby directly strengthening system-wide risk resistance (Zhou et al., 2025). In particular, investment in irrigation and water conservancy facilities can mitigate yield losses from droughts and floods, while investment in high-standard farmland construction enhances overall production capacity (Jiao et al., 2025). Together, these investments improve production stability and continuity and reduce production disruptions induced by external shocks such as natural hazards and market fluctuations.

PCCE and PCIR (rural residents’ expenditure and income) are also highly influential, contributing 14.3 and 10.4%, respectively. Increases in household income and consumption capacity enhance farmers’ ability to finance production inputs and facilitate technology adoption and innovation. Higher rural income enables greater investment in modern machinery, improved seed varieties, and other productivity-enhancing inputs, and increases both willingness and capacity to adopt new technologies, thereby improving efficiency and output quality. Rising income may also strengthen risk-management capacity by supporting agricultural insurance uptake and encouraging income diversification, which reduces exposure to shocks affecting a single production activity.

Other indicators also contribute substantially to RCAI. Ranked by contribution, GOGA (land productivity) accounts for 10.2%, directly capturing output efficiency per unit of land. Higher land productivity implies greater agricultural output under the same resource endowment, which not only strengthens economic performance but also enhances the capacity to absorb external shocks. Improvements in land productivity are typically associated with technological progress and more efficient resource use, supporting a transition toward more efficient and sustainable agricultural development. DRS (disaster relief expenditure for rural residents) contributes 6.9% and reflects public financial assistance provided in response to disaster-related agricultural losses. Such support can help affected households resume production more rapidly and reduce income losses associated with production interruptions. TIL (transportation infrastructure; 6.1%) and PTRG (postal and telecommunications service volume; 5.0%) further highlight the importance of connectivity and information flows. Under natural hazards or market fluctuations, better transport infrastructure facilitates timely resource mobilization and stabilizes production and distribution. Higher postal and telecommunications activity implies improved access to market, technical, and policy information, enabling producers to make more informed decisions.

The SHAP summary plot in Figure 6b further illustrates the direction of each feature’s effect on RCAI. Each point represents an observation, with color indicating feature value; SHAP values above (below) zero imply positive (negative) contributions relative to the baseline prediction (Gao et al., 2025; Pinichka et al., 2025). Overall, IAFA, PCCE, PCIR, GOGA, DRS, TIL, and PTRG exhibit predominantly positive effects, indicating a generally positive association between these indicators and RCAI, though the magnitude of this effect varies nonlinearly. Taken together, improvements in these factors can enhance production efficiency and quality, strengthen disaster recovery capacity, improve circulation efficiency, and promote information diffusion and market integration, thereby raising the resilience of China’s agricultural industry. Notably, some observations display negative SHAP values, suggesting the presence of threshold effects or other nonlinearities, which motivates a more detailed variable-specific investigation in the subsequent analysis.

We further select the six most influential features and use SHAP dependence plots to examine the complex nonlinear relationships between these factors and RCAI (Figure 7). The analysis reveals that the enhancement of agricultural resilience is governed by distinct structural thresholds, where different factors specifically activate the system’s capacities for resistance, adaptation, and innovation in non-uniform ways.

Agricultural Fixed Asset Investment (IAFA) demonstrates characteristics of increasing marginal returns. The dependence plot indicates that below a threshold of 475.20, the contribution of capital investment to resilience is negligible, likely because low-level investments are absorbed by basic depreciation and maintenance. However, surpassing this inflection point triggers a qualitative leap in system capability (Xue et al., 2025). High-intensity investment serves as a dual driver: it fortifies resistance capacity through the construction of high-standard farmland and robust infrastructure, while simultaneously unlocking innovation capacity by providing the material basis for technological upgrading and equipment renewal (Jiang and Peng, 2026).

Similarly, rural consumption (PCCE) and income (PCIR) exhibit threshold behaviors consistent with the relaxation of liquidity constraints. Both variables show a distinct “hockey-stick” trajectory, with resilience gains accelerating sharply only after exceeding thresholds of 5653.60 and 5944.06, respectively. This suggests that below these levels, household resources are primarily allocated to subsistence. Once income crosses this safety threshold, the financial buffer allows farmers to diversify livelihoods and adopt climate-smart practices, thereby significantly boosting their adaptation capacity to market and climatic volatility, as well as supporting post-disaster recovery (Mnukwa et al., 2025).

Transportation Infrastructure (TIL) displays a pattern indicative of network externalities. The impact on resilience remains limited until the index reaches 11.94, implying that fragmented road networks contribute little to systemic stability. Surpassing this critical density enables the rapid release of resilience benefits. A connected network coordinates the system’s defenses: it enhances adaptation capacity by ensuring efficient market circulation during fluctuations and strengthens resistance capacity by facilitating the rapid mobilization of emergency resources during extreme events (Hossain and Kashem, 2025; Mthembu et al., 2025).

In contrast, Land Productivity (GOGA) and Disaster Relief Subsidies (DRS) show signs of diminishing returns or passive effects. GOGA exhibits a ceiling effect, indicating that relying solely on yield maximization provides only baseline resistance but cannot sustain long-term resilience growth without structural transformation. Meanwhile, DRS functions primarily as an ex-post compensatory mechanism. While positively correlated with RCAI, it supports short-term recovery but contributes less to active, ex-ante adaptive or innovative capabilities compared to proactive investments.

Overall, these findings provide a supplement to the traditional linear narrative about the construction of resilience. They suggest that resilience enhancement is not a continuous linear accumulation but a step-wise evolution. Achieving higher resilience requires key capacities—specifically capital stock, financial liquidity, and physical connectivity—to accumulate beyond critical tipping points. Only by crossing these thresholds can the agricultural system structurally activate its intrinsic abilities to resist shocks, adapt to changes, and innovate for the future.

To assess the robustness of the empirical results, we reconstruct the resilience index using two alternative composite methods—entropy–TOPSIS and entropy–VIKOR—and then refit XGBoost using the newly computed RCAI as the target variable. The results are reported in Table 4. Under the two alternative indices, the test-set R² values are 0.9342 and 0.9536, respectively, indicating that model fit remains strong after replacing the index-construction method. The corresponding test-set RMSE values are 0.0001 and 0.0248, suggesting that prediction errors remain low. In addition, five-fold cross-validation yields mean MAE values of 9.8796 × 10 − 5 and 0.0136, confirming that the model performs well when applied to the full dataset.

We further examine whether the results are sensitive to indicator selection. Based on the feature-importance ranking from the Pearson correlation analysis, we remove indicators with correlations below 0.3, reconstruct RCAI, and refit the XGBoost model. The resulting MAE (0.0016) and RMSE (0.0905) are lower than those of the baseline specification, and R² increases to 0.9553, indicating improved predictive accuracy and overall fit. These findings suggest that the main results are not driven by indicator-selection bias.

This study measures and evaluates the resilience of China’s agricultural industry over the period 2000–2022, refines the resilience index using XGBoost, and employs SHAP to identify key drivers and nonlinear relationships. The main findings are as follows.

First, the resilience of China’s agricultural industry exhibits pronounced spatiotemporal heterogeneity. It shows an overall upward trajectory over time, alongside substantial regional disparities. The evolution of resilience follows a clear pattern of “fragmentation–regional clustering–spatial contiguity.” Driven jointly by policy reforms, technological upgrading, and agricultural structural adjustment, China’s overall capacity to withstand risks has improved markedly. High-resilience regions are concentrated in North China and the southeastern coastal provinces, whereas resilience improvement is relatively limited in the northwest due to constraints such as climate and resource endowments. Among the three dimensions, resistance and adaptation capacity contribute to gradual resilience enhancement, while innovation capacity acts as the key force behind resilience upgrading.

Second, comparative evaluation across multiple machine-learning models indicates that XGBoost demonstrates robust performance in fitting and predicting the relationship between RCAI and the underlying indicator system. Using the trained XGBoost model to reconstruct the index yields a refined RCAI that is more coherent and reliable. This refinement mitigates shortcomings of conventional index-construction methods in handling high-dimensional structured data and provides a more robust treatment of extreme observations, thereby reducing the risk of overall measurement bias induced by outliers. Consequently, the proposed “index construction–XGBoost–SHAP” framework not only improves measurement plausibility but also addresses the interpretability challenge of machine-learning models, offering strong potential for broader application in other evaluation and composite-index studies.

Third, the explainable machine-learning framework effectively identifies key determinants and nonlinearities of RCAI. SHAP results show that agricultural fixed-asset investment, rural residents’ income and expenditure, transportation infrastructure, and disaster relief expenditure contribute most to RCAI. These factors improve regional agricultural infrastructure, production capacity, and adaptive capability, thereby promoting resilience enhancement. At the same time, the effects of several drivers display clear threshold behavior, indicating that resilience gains may accelerate only after certain capacity levels are reached.

Based on the identification of key drivers and threshold mechanisms, this study proposes three targeted policy implications to enhance the resilience of China’s agricultural industry.

First, investment strategies should focus on reaching the “critical mass” of capital accumulation to trigger scale effects. Since agricultural fixed-asset investment exhibits a distinct threshold effect, scattered or insufficient investments may fail to generate observable resilience gains. Policymakers should prioritize high-intensity investments in key infrastructure—such as high-standard farmland and water conservancy projects—ensuring that regional investment levels surpass the effective threshold to trigger a qualitative leap in resistance and innovation capacities.

Second, enhancing the financial adaptability of rural households is essential for bottom-up resilience. The findings reveal that rural income and consumption must cross specific safety thresholds to unlock significant improvements in adaptation capacity. Therefore, policies should go beyond post-disaster relief to focus on ex-ante capacity building. This includes diversifying rural income sources and strengthening financial support systems, enabling farmers to overcome liquidity constraints and invest in resilience-enhancing technologies and diversified livelihoods.

Third, differentiated resilience-building pathways should be adopted based on regional endowments. Recognizing the “East–West” disparity, the eastern regions should leverage their advantages to foster transformation and innovation, focusing on digital agriculture and structural upgrading. Conversely, the western regions, constrained by ecological fragility, should prioritize resistance and adaptation. Interventions in these areas should focus on overcoming binding constraints (e.g., water scarcity) through region-specific technologies (e.g., water-saving irrigation) rather than blindly replicating the capital-intensive models of the eastern plains.