Decoupling Dynamics, Utilization Efficiency, and Driving Mechanisms of Potash Fertilizer Inputs and Grain Production in China: Evidence from Provincial Panel Data, 2000–2024

Abstract

Potassium is an essential nutrient for crop growth and plays a critical role in regulating water metabolism, facilitating photosynthate transport, and improving agricultural product quality. The precise management of potash fertilizer inputs is therefore vital for enhancing agricultural productivity and promoting sustainable resource use. Using panel data for 31 provinces in China from 2000 to 2024, obtained from the China Statistical Yearbook, this study integrates the Tapio decoupling model, stochastic frontier analysis (SFA), fixed-effects models, and an XGBoost–BiLSTM hybrid model to investigate the dynamic relationship, utilization efficiency, and driving mechanisms of potash fertilizer inputs and grain production. The results indicate that the relationship between potash fertilizer inputs and grain production has shifted from an expansive negative decoupling state—characterized by faster growth in fertilizer inputs than in output—to a strong decoupling state, where fertilizer inputs decline while grain production continues to increase. This transition exhibits a clear spatial gradient, with improvements from eastern to northeastern and central regions. Potassium use efficiency (KUE) shows a steady upward trend, with significant regional heterogeneity, characterized by higher efficiency in the south, lower efficiency in the north, and notable differentiation in western regions, largely driven by climatic and soil variations. Despite these improvements, substantial potential for reducing fertilizer inputs remains across provinces. Potash fertilizer inputs exert a significant positive effect on grain production, while the cultivation of potassium-intensive crops, such as sugar crops, tobacco, and fruits, is a key driver of regional demand. Model projections suggest that from 2025 to 2030, grain production will grow at an annual rate of 1.2–1.5%, while potash fertilizer inputs will decline by 2–4% annually, indicating a transition toward greener agricultural development. These findings highlight the need for region-specific fertilization strategies, optimized fertilizer structures, and improved soil nutrient monitoring systems to ensure food security and sustainability.

1. Introduction

Food security constitutes a foundational pillar of socioeconomic stability and human well-being [1,2]. Against the backdrop of accelerating climate change, increasingly constrained arable land, and escalating ecological pressures, sustaining growth in food production on a finite land base has emerged as one of the most pressing challenges facing global agriculture [3,4,5]. Among essential plant nutrients, potassium plays an irreplaceable physiological role in regulating plant water relations, enhancing photosynthate translocation, and improving crop quality, thereby exerting a direct and substantial influence on crop productivity and resilience [6,7,8].

Since the second half of the twentieth century, rapid population growth and rising food demand have markedly intensified agriculture’s dependence on chemical fertilizers [9,10]. Global potassium fertilizer consumption increased from less than 10 million tons in 1960 to approximately 38.8 million tons in 2024 [11,12]. However, potassium input levels vary enormously across regions due to differences in economic development, agrarian structure, and technological capacity. Using FAO data, Wang et al. [13] reported that during 2016–2020, the global average potassium application rate for cereal crops was only 27 kg ha⁻¹, with values of 19 kg ha⁻¹ in developed countries, 41 kg ha⁻¹ in developing countries, and as low as 1.9 kg ha⁻¹ in the least developed countries. In striking contrast, China’s potassium fertilizer input reached approximately 80 kg ha⁻¹, far exceeding the global mean and underscoring a substantial anomaly in input intensity.

Potassium use efficiency (KUE) likewise exhibits pronounced spatial differentiation. Dhillon et al. [14] estimated that the global average KUE for cereal crops remains below 30%. While advanced regions such as North America and Western Europe have achieved utilization rates exceeding 60% through precision fertilization and scientifically informed management practices, KUE in many developing countries remains below 25% due to imbalanced nutrient application regimes and extensive field management practices. Excessive potassium application not only generates substantial resource waste but also exacerbates soil acidification and aquatic eutrophication, thereby imposing increasing ecological externalities [6,15,16].

As the world’s largest grain producer, China has realized sustained and stable growth in grain output since the initiation of economic reforms [17,18], a trajectory that has relied heavily on increasing fertilizer inputs, including potassium. Between 2000 and 2015, China’s potassium fertilizer consumption increased from 3.66 to 6.42 million tons, accounting for roughly 17% of global use during the same period [15], thereby revealing a structural pattern of “high input–low efficiency” production. More critically, China possesses only about 2% of the world’s proven potassium reserves, and its external dependence on potassium fertilizer imports has persistently exceeded 50% [7,19]. This pronounced import reliance renders domestic supply highly vulnerable to geopolitical shocks and volatility in international markets, thereby posing a substantial yet latent threat to national food security.

In recent years, policy initiatives such as the “Zero Growth of Fertilizer Use” and the “Fertilizer Reduction Action Plan” have driven a continuous decline in China’s total fertilizer inputs since 2015, signaling a structural transition from high-input expansion toward a quality-oriented development pathway emphasizing efficiency and input optimization [10,20,21]. Accumulating evidence indicates that scientifically guided fertilization combined with precision management can maintain high and stable yields while reducing fertilizer inputs. Using nationwide microdata from millions of smallholder farmers, Yin et al. [22] demonstrated that scientific fertilization coupled with technical guidance can simultaneously reduce fertilizer use and sustain high grain yields. Zhao et al. [12] further showed that optimized nutrient management and cropping systems can maintain yield stability under an average annual fertilizer reduction rate of approximately 1.5%.

However, despite these advances, a critical limitation persists. Existing studies overwhelmingly concentrate on the aggregate relationship between total fertilizer inputs and grain output, while systematic, nutrient-specific quantification of the mechanisms linking potassium fertilizer inputs to grain production remains notably absent. In particular, little is known about (i) whether and to what extent potassium inputs have decoupled from grain output growth, (ii) how potassium use efficiency varies across space and time, (iii) which structural and agronomic factors drive regional potassium demand, and (iv) whether future grain production can be sustained under declining potassium input trajectories. This lack of targeted evidence severely limits the development of rational, region-specific potassium fertilization strategies and weakens the empirical foundation for reconciling food security with an agricultural green transition.

The Tapio decoupling model is employed to characterize the dynamic coupling and decoupling states between potash fertilizer inputs and grain production, thereby providing a macro-level perspective on the evolution of agricultural sustainability [23]. Stochastic frontier analysis (SFA) enables a precise assessment of potassium use efficiency by disentangling random shocks from technical inefficiency. The fixed-effects model controls for time-invariant regional heterogeneity, allowing for the identification of structural factors influencing both grain output and fertilizer inputs. Furthermore, the XGBoost–BiLSTM hybrid deep learning framework captures complex nonlinear relationships and temporal dependencies within the agricultural system, significantly enhancing the accuracy and generalizability of future trend predictions.

In summary, based on panel data for 31 provinces in China from 2000 to 2024, this study integrates the Tapio decoupling model, stochastic frontier analysis, fixed-effects modeling, and an XGBoost–BiLSTM hybrid framework to systematically examine the decoupling relationship, utilization efficiency, and driving mechanisms of potash fertilizer inputs and grain production, as well as to forecast their trends for the period 2025–2030. By jointly incorporating input dynamics, utilization efficiency, and future trajectories, this study provides robust empirical evidence for the formulation of region-specific potash fertilizer management strategies, and offers policy insights for balancing grain production growth with efficient resource use and long-term agricultural sustainability.

2. Materials and Methods

2.1. Data Sources

All data used in this study were obtained from the China Statistical Yearbook 2001–2025 (http://www.stats.gov.cn). The study covers 31 provincial-level administrative units in mainland China (excluding Hong Kong, Macao, and Taiwan). Missing observations were imputed using linear interpolation in Stata 18 to maintain the completeness and continuity of the panel dataset, a balanced panel dataset was constructed for subsequent decoupling analysis, efficiency evaluation, econometric modeling, and machine-learning forecasting. The selected indicators include the following:

• Outputs of major agricultural products (grain, cotton, oil crops, fiber crops, sugar crops, tobacco, tea, and fruit);
• Sown area of grain crops;
• Input amounts of nitrogen (N), phosphorus (P), and potassium (K) fertilizers;
• Compound fertilizer inputs;
• Total agricultural machinery power;
• Effective irrigated area of cultivated land.

To ensure consistency and comparability, fertilizer input variables were measured in nutrient-equivalent terms using the standardized statistical definitions reported in the China Statistical Yearbook. Specifically, fertilizer use was recorded as the amount of agricultural chemical fertilizer applied after conversion to pure nutrient content, including nitrogen (N), phosphorus (P₂O₅), and potassium (K₂O). This approach reduces potential bias caused by differences in fertilizer types and nutrient concentrations across regions and over time.

To investigate regional heterogeneity in potassium fertilizer inputs under different levels of economic development, the 31 provincial units were classified into four major economic regions—eastern, central, western, and northeastern China—based on the regional development criteria issued by the Development Research Center of the State Council of China (Table 1). This regional grouping framework was used consistently throughout the empirical analysis to identify spatial gradients, structural differences, and region-specific driving mechanisms of potassium fertilizer use.

Table 1. Classification of Four Major Chinese Economic Regions.

Area	Provinces Contained
Eastern China	Beijing, Tianjin, Hebei, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, Hainan
Central China	Shanxi, Anhui, Jiangxi, Henan, Hubei, Hunan
Western China	Inner Mongolia, Guangxi, Chongqing, Sichuan, Guizhou, Yunnan, Tibet, Shaanxi, Gansu, Qinghai, Ningxia, Xinjiang
Northeastern China	Liaoning, Jilin, Heilongjiang

2.2. Research Methods

2.2.1. Tapio Decoupling Model

The Tapio decoupling model is a widely used analytical framework for evaluating the relationship between economic growth and resource consumption or environmental pressure [24,25]. To examine the coupling between potassium fertilizer inputs and grain production, we applied this model and identified the decoupling status by calculating the decoupling elasticity coefficient [26], which is defined as

$$e={\displaystyle \frac{\left(PF{U}_t- PF{U}_{t-1}\right) / PF{U}_{t-1}}{\left(GU{A}_t- GU{A}_{t-1}\right) / GU{A}_{t-1}}}={\displaystyle \frac{\mathit{\Delta }PFU / PF{U}_{t-1}}{\mathit{\Delta }GUA / GU{A}_{t-1}}}$$

(1)

where e denotes the decoupling elasticity coefficient; GUA_t and GUA_t−1 represent grain output per unit area in the current and base periods, respectively; PFU_t and PFU_t−1 denote potassium fertilizer input per unit area in the current and base periods, respectively; and $\mathit{\Delta }GUA = GU{A}_t - GU{A}_{t-1}$ and $\mathit{\Delta }PFU = PF{U}_t - PF{U}_{t-1}$ represent the changes in grain output and potassium fertilizer input between two consecutive periods, respectively.

Based on the value of the decoupling elasticity coefficient and the directional changes in grain output and potassium fertilizer inputs, the decoupling relationship was classified into three major categories—decoupling, coupling, and negative decoupling—and further subdivided into eight specific states [27]. The detailed classification criteria are presented in Table 2.

Table 2. Classification Criteria for Decoupling States.

Category	State	Diagnostic Condition	Economic Implication
Decoupling	Strong decoupling	ΔPFU < 0, ΔGUA > 0, e < 0	Potassium fertilizer input decreases while grain output increases (reduced input with increased production)
	Weak decoupling	ΔPFU > 0, ΔGUA > 0, 0 < e < 0.8	Both increase, but the growth rate of potassium fertilizer is lower than that of grain output
	Recessive decoupling	ΔPFU < 0, ΔGUA < 0, e > 1.2	Both decrease, but the reduction rate of potassium fertilizer is greater than that of grain output
Coupling	Expansive coupling	ΔPFU > 0, ΔGUA > 0, 0.8 ≤ e ≤ 1.2	Both increase at a similar pace
	Recessive coupling	ΔPFU < 0, ΔGUA < 0, 0.8 ≤ e ≤ 1.2	Both decrease at a similar pace
Negative decoupling	Expansive negative decoupling	ΔPFU > 0, ΔGUA > 0, e > 1.2	Both increase, but the growth rate of potassium fertilizer is higher than that of grain output (inefficient growth)
	Weak negative decoupling	ΔPFU < 0, ΔGUA < 0, 0 < e < 0.8	Both decrease, but the reduction rate of potassium fertilizer is smaller than that of grain output
	Strong negative decoupling	ΔPFU > 0, ΔGUA < 0, e < 0	Potassium fertilizer input increases while grain output decreases (increased input with reduced production)

Note: ΔGUA and ΔPFU denote the changes in grain output and potassium fertilizer use, respectively, between two consecutive periods. e represents the decoupling elasticity coefficient, defined as the ratio of the percentage change in potassium fertilizer use to the percentage change in grain output, and it is used to characterize the decoupling relationship between agricultural production and fertilizer input.

From the perspective of agricultural sustainable development, strong decoupling represents the most desirable state, characterized by a reduction in potassium fertilizer input alongside continued growth in grain output. This reflects the goal of green development—achieving higher agricultural productivity with lower resource inputs. In contrast, strong negative decoupling represents the least desirable state, where grain output declines despite increased potassium fertilizer input. This indicates that agricultural production fails to generate proportional output returns from increased input, leading to resource waste and greater environmental pressure. It reflects a significant deterioration in input–output efficiency and poses a serious challenge to the green transformation and sustainable development of agriculture.

2.2.2. Stochastic Frontier Analysis (SFA)

Stochastic frontier analysis (SFA) is a parametric efficiency evaluation method that specifies the functional form of a production or cost function in advance and decomposes the composite error term into a stochastic disturbance component and a non-negative technical inefficiency term, thereby enabling the quantitative identification of the relative technical efficiency of decision-making units [28,29]. Technical efficiency reflects the ability of a decision-making unit to achieve an optimal input–output allocation under given technological constraints, that is, the extent to which maximum output can be obtained from a fixed input level or a given output can be produced with minimal inputs [30,31]. The core SFA specification is expressed as

$$y_{it}=f \left(x_{it}; \beta \right) \times exp \left(v_{it}-{ u}_{it}\right)$$

(2)

where $y_{it}$ represents the observed output of decision unit i at time t; $f \left(x_{it}; \beta \right)$ denotes the theoretical maximum attainable output given the input vector $x_{it}$, i.e., the production frontier; $v_{it}$ is the stochastic disturbance term capturing measurement errors and uncontrollable random shocks, assumed to follow a normal distribution $N\left(0, {\sigma }_{\alpha }^2\right)$; and $u_{it}$ is a non-negative technical inefficiency term that measures the deviation of observed output from the production frontier [32,33,34,35].

In this study, the stochastic production frontier was specified in the Cobb–Douglas functional form, which is widely adopted in agricultural efficiency research and avoids potential multicollinearity issues associated with transcendental logarithmic functions [36,37]. Since potassium fertilizer inputs are included within the total chemical fertilizer metric, multicollinearity among explanatory variables was examined, and all variance inflation factors (VIFs) were found to be below commonly accepted thresholds, indicating the statistical robustness of the model specification.

Based on this framework, an input-oriented, time-varying decay model was employed to capture the dynamic evolution of technical efficiency. The stochastic frontier production function for potassium use efficiency was specified as

$$ln{K}_{it}={ \beta }_0+{ \beta }_{1}ln{Y}_{it}+{\beta }_{2}ln{W}_{it}+{\beta }_{3}ln{M}_{it}+{\beta }_{4}ln{F}_{it}+v_{it}+u_{it}$$

(3)

where i = 1, 2, …, n denotes the decision-making units; t denotes time; $Y_{it}$ is the grain output per unit area of unit i at time t; K_it, W_it, $M_{it}$ and $F_{it}$ represent potassium fertilizer inputs, irrigated area, agricultural machinery input, and total fertilizer inputs, respectively; and β0~β4 are the parameters to be estimated.

The technical inefficiency term μit was modeled using a time-varying decay specification:

$${\mu }_{it} ={ u}_i \times exp \left(-\eta \left(t - T\right) \right)$$

(4)

where μ_it is the unit-specific baseline inefficiency, η is the decay parameter representing the temporal trend of technical inefficiency, and T denotes the final period of the sample [38,39].

Based on the stochastic frontier model estimates, KUE was calculated as

$$KU{E}_{it}={\displaystyle \frac{E \left(Y_{it}\right)}{E \left(Y_{it}/u_{it}=0\right)}}=exp \left({-u}_{it}\right)$$

(5)

Substituting Equation (4) into Equation (5) yields

$$K{E}_{it}=exp \left(-u_i \times exp (-\eta (t-T\left)\right)\right)$$

(6)

where $K{UE}_{it}$ represents the ratio of the minimum feasible optimal potassium input to the actual input $K_{it}$, with a theoretical range of (0, 1]. Values approaching 1 indicate higher efficiency [32,33].

KUE directly determines the potential for input reduction. Accordingly, the feasible reduction ratio of potassium inputs RE_it was defined as

$${RE}_{it}=1-K{E}_{it}$$

(7)

The corresponding feasible reduction in potassium input can then be calculated as

$${RE=K_{it} \times RE}_{it}$$

(8)

In this framework, RE quantifies the amount by which actual potassium fertilizer application can be safely reduced without compromising the frontier output level, thereby providing an empirical basis for precision fertilization strategies.

2.2.3. Fixed-Effects Model

The fixed-effects model controls for time-invariant individual heterogeneity in panel data, effectively mitigating omitted variable bias and enabling the precise estimation of the causal effects of explanatory variables on the dependent variable [30,40]. In this study, four estimation approaches were applied for comparative analysis: pooled ordinary least squares (POLSs), individual fixed effects (FEs), random effects (REs), and two-way fixed effects (TWFEs) [41,42]. The corresponding model specifications are as follows:

$$Y_{it} = \alpha + \sum _{k=1}^k{\beta }_{ki}{X}_{kit} +{ u}_{it}$$

(9)

$$Y_{it}={ \alpha }_i+\sum _{k=1}^k{\beta }_{ki}{X}_{kit}+u_{it}$$

(10)

$$Y_{it}= \alpha +{\alpha }_i+\sum _{k=1}^k{\beta }_{ki}{X}_{kit}+u_{it}$$

(11)

$$Y_{it}={\alpha }_i+{\lambda }_i+\sum _{k=1}^k{\beta }_{ki}{X}_{kit}+u_{it}$$

(12)

Equations (9)–(12) correspond to POLS, FE, RE, and TWFE models, respectively. Here, i = 1, 2, …, n, denotes the provincial units; t = 1, 2, …, T represents the time periods; and Y_it is the dependent variable, representing grain output in province i at year t; α is the overall intercept. In the FE model, ${\alpha }_i$ represents the unit-specific fixed effect, while in the RE model, it is assumed to follow a normal distribution N (0, ${\sigma }_{\alpha }^2$) and is uncorrelated with the explanatory variables. ${\lambda }_i$ denotes the time fixed effect, capturing common temporal trends across all provinces. X_kit represents the k-th explanatory variable (e.g., potassium fertilizer input, nitrogen fertilizer input, irrigated area, etc.), β_ki is the corresponding regression coefficient, and u_it is the stochastic error term, assumed to follow N (0, ${\sigma }_{\alpha }^2$) [42,43,44].

2.2.4. XGBoost-BiLSTM

The accurate prediction of potassium fertilizer inputs and grain output provides a quantitative basis for analyzing the driving mechanisms of potassium use in grain production. While traditional econometric models capture linear relationships between variables, they are limited in representing the complex nonlinearities and temporal dependencies inherent in agricultural production systems. To address this limitation, an XGBoost-BiLSTM hybrid deep learning model was developed to enhance predictive accuracy and generalization capability.

XGBoost (eXtreme Gradient Boosting) is a highly efficient gradient boosting decision tree algorithm [45], which iteratively integrates multiple weak learners to construct a strong learner [46]. The algorithm incorporates a regularization term in the objective function to control model complexity and prevent overfitting. BiLSTM (Bidirectional Long Short-Term Memory) is a bidirectional recurrent neural network architecture composed of forward and backward LSTM layers, capable of capturing both forward and backward temporal dependencies in time-series data [47].

The XGBoost-BiLSTM hybrid model employs a hierarchical ensemble structure, integrating the feature extraction and selection strength of XGBoost with the temporal modeling capability of BiLSTM, thereby overcoming the limitations of single-model approaches [48]. This architecture effectively captures complex nonlinear relationships and dynamic temporal patterns in agricultural production data, providing technical support for precise prediction of potassium fertilizer inputs and grain output. The structure of the XGBoost-BiLSTM model used in this study is illustrated in Figure 1.

Figure 1. Schematic Architecture of the XGBoost-BiLSTM.

3. Results

3.1. Decoupling Relationship Between Potassium Fertilizer Inputs and Grain Output

A sliding-window analysis was conducted to examine the Tapio decoupling relationship between potassium fertilizer inputs and grain output across 31 provincial-level units in China. The 25-year study period was divided into four overlapping intervals: 2000–2010, 2005–2015, 2010–2020, and 2015–2024. The temporal and spatial evolution of the decoupling states is illustrated in Figure 2.

Figure 2. Spatiotemporal evolution of decoupling states between grain output and potassium fertilizer inputs in 31 Chinese provinces. (a) 2000–2010; (b) 2005–2015; (c) 2010–2020; (d) 2015–2024.

Figure 2 shows that from 2000 to 2010, the vast majority of Chinese provinces were in a state of expansive negative decoupling, indicating that potassium fertilizer inputs and grain output increased simultaneously, while the growth rate of potassium fertilizer inputs generally exceeded that of grain output. A clear positive association between the two variables was observed, with only limited regional variation, resulting in a high degree of overall consistency across the country.

Between 2005 and 2015, the decoupling relationship between potassium fertilizer inputs and grain output in China began to undergo a marked transformation, accompanied by an emerging trend of regional differentiation. In eastern coastal provinces such as Tianjin, Shanghai, Jiangsu, Zhejiang, and Anhui, the decoupling state gradually shifted from expansive negative decoupling to strong decoupling, indicating that grain output continued to increase despite a slowdown or even decline in potassium fertilizer inputs. In contrast, the decoupling process in the central and western regions lagged behind, where expansive negative decoupling remained the dominant pattern.

During 2010–2020, the decoupling pattern further evolved, and regional disparities continued to widen. Provinces including Tianjin, Jilin, Heilongjiang, Jiangsu, Zhejiang, and Anhui maintained relatively stable and coordinated relationships between potassium fertilizer inputs and grain output. Strong decoupling and expansive negative decoupling expanded from the eastern coastal region toward the northeast and central regions, showing a certain degree of spatial diffusion. Although a few provinces experienced temporary fluctuations during specific periods, the overall direction of decoupling evolution continued to improve.

From 2015 to 2024, a nationwide spatial pattern dominated by strong decoupling gradually emerged. More than 20 provinces exhibited strong decoupling, characterized by a continued decline in potassium fertilizer inputs alongside sustained growth in grain output. Although grain output declined slightly in some provinces, the reduction in potassium fertilizer inputs was even greater. At this stage, the coordination between input and output improved significantly, and the overall trend remained stable and robust.

Overall, from 2000 to 2024, the relationship between potassium fertilizer inputs and grain output in China underwent a transition from a pattern dominated by expansive negative decoupling to one characterized primarily by strong decoupling. The overall spatial pattern displayed a gradient of improvement progressing sequentially from the eastern region to the northeast and then to the central region.

3.2. Potassium Use Efficiency (KUE) Estimation

KUE serves as a critical indicator for evaluating the effective utilization of potassium fertilizer resources. The systematic measurement of KUE, combined with an analysis of spatial heterogeneity and temporal dynamics, facilitates a comprehensive assessment of the effectiveness of potassium inputs across regions and helps identify potential optimization opportunities.

3.2.1. Stochastic Frontier Model Estimation

The SFA estimation results are presented in Table 3. Core explanatory variables and efficiency-related parameters (γ, μ, η, etc.) were statistically significant at the 1% level, with generally high absolute t-values, indicating robust and unbiased parameter estimates. These results validate the appropriateness of the SFA model specification.

Table 3. Estimation Results of the KUE Production Function Based on SFA.

Production Function	Coefficients	Standard Error	t-Value
ln_Yit	−0.046	0.081	−0.568
ln_Mit	−0.117 ***	0.030	−3.900
ln_Wit	−0.002	0.007	−0.286
ln_Fit	1.415 ***	0.036	39.306
Constant	−3.541 ***	0.68	−5.207
ln σ2	−1.656 ***	0.297	−5.576
lgt γ	1.302 ***	0.381	3.417
μ	0.577 ***	0.108	5.343
η	0.014 ***	0.002	7.000
σ2	0.191	-	-
γ	0.786	-	-

Note: ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels.

Specifically, the Gamma (γ) value was 0.786 and significant at the 1% level, indicating that 78.6% of the variance in the stochastic frontier production function can be attributed to technical inefficiency, with the remaining 21.4% arising from uncontrollable random factors such as climatic variability or measurement errors. This confirms the model’s ability to separate controllable efficiency losses from stochastic disturbances. The total variance (σ² = 0.191) was moderate, indicating that the model captured output fluctuations appropriately without evidence of overfitting or underfitting. Furthermore, the estimated results are consistent with theoretical expectations and empirical findings in prior agricultural efficiency studies, providing additional validation for the model specification.

3.2.2. Potassium Use Efficiency (KUE)

The KUE values for the 31 Chinese provinces from 2000 to 2024, estimated based on the SFA model, are summarized in Table 4.

Table 4. KUE values of China’s provinces from 2000 to 2024.

Region	Province	2000	2005	2010	2015	2020	2024
Eastern Region	Beijing	0.228	0.252	0.278	0.304	0.330	0.351
	Tianjin	0.359	0.391	0.417	0.443	0.469	0.489
	Hebei	0.316	0.348	0.374	0.401	0.427	0.448
	Shanghai	0.181	0.209	0.233	0.258	0.283	0.304
	Jiangsu	0.218	0.248	0.273	0.299	0.325	0.346
	Zhejiang	0.360	0.392	0.418	0.444	0.470	0.490
	Fujian	0.902	0.910	0.916	0.921	0.927	0.930
	Shandong	0.355	0.387	0.413	0.439	0.465	0.485
	Guangdong	0.856	0.867	0.876	0.884	0.891	0.897
	Hainan	0.737	0.756	0.770	0.784	0.798	0.808
Northeastern Region	Liaoning	0.394	0.425	0.451	0.477	0.502	0.522
	Jilin	0.343	0.375	0.401	0.427	0.453	0.474
	Heilongjiang	0.821	0.835	0.845	0.855	0.865	0.872
Central Region	Shanxi	0.380	0.412	0.438	0.464	0.489	0.509
	Anhui	0.414	0.446	0.471	0.497	0.521	0.541
	Jiangxi	0.930	0.935	0.940	0.944	0.948	0.950
	Henan	0.344	0.375	0.402	0.428	0.454	0.475
	Hubei	0.319	0.350	0.377	0.403	0.430	0.450
	Hunan	0.883	0.892	0.899	0.905	0.912	0.916
Western Region	Inner Mongolia	0.375	0.407	0.433	0.459	0.485	0.505
	Guangxi	0.975	0.977	0.978	0.980	0.981	0.982
	Chongqing	0.252	0.282	0.308	0.334	0.361	0.382
	Sichuan	0.293	0.325	0.351	0.377	0.404	0.425
	Guizhou	0.494	0.524	0.548	0.571	0.594	0.612
	Yunnan	0.503	0.532	0.556	0.579	0.601	0.619
	Tibet	0.560	0.588	0.610	0.631	0.652	0.668
	Shaanxi	0.343	0.375	0.401	0.427	0.453	0.474
	Gansu	0.370	0.402	0.428	0.454	0.480	0.500
	Qinghai	0.257	0.288	0.314	0.340	0.366	0.388
	Ningxia	0.226	0.256	0.281	0.307	0.333	0.354
	Tibet	0.208	0.237	0.262	0.288	0.314	0.335
China		0.458	0.484	0.505	0.527	0.548	0.565

The results indicate that KUE exhibited a steady upward trend nationwide, with the average KUE increasing from 0.458 in 2000 to 0.565 in 2024. Spatially, KUE displayed a “high in the south, low in the north, and marked differentiation in the western region” pattern. Provinces such as Fujian, Guangdong, and Hainan consistently maintained high efficiency levels, whereas Beijing, Shanghai, and Jiangsu, although exhibiting relatively low efficiency levels, showed stable improvement trends. In the northeastern region, Heilongjiang exhibited the highest efficiency, while Liaoning and Jilin maintained moderate efficiency with steady growth. The central region exhibited a polarized pattern: Jiangxi and Hunan led in efficiency, whereas traditional major grain-producing provinces such as Henan and Hubei still had considerable room for improvement. The western region demonstrated the highest internal heterogeneity, with Guangxi ranking highest nationally, Guizhou, Yunnan, and Tibet maintaining above-average efficiency, and northwestern arid provinces such as Xinjiang, Ningxia, and Qinghai exhibiting lower efficiency but relatively higher growth rates, indicating a clear catching-up trend.

3.2.3. Potassium Fertilizer Reduction Potential (PFRP)

Based on the estimated potassium use efficiency (KUE) for each province (Table 4) and the average potash fertilizer input per unit area, and in conjunction with Equations (7) and (8), the average annual PFRP for each province over the period 2000–2024 was further calculated (Table 5).

Table 5. Average KUE, APFI and PFRP in each province from 2000 to 2024.

Region	Province	KUE	APFI (kg ha−1)	PFRP (kg ha−1)
Eastern Region	Beijing	0.286	32.862	23.291
	Tianjin	0.425	42.197	24.244
	Hebei	0.382	36.108	22.227
	Shanghai	0.241	25.242	19.226
	Jiangsu	0.281	34.898	25.164
	Zhejiang	0.426	50.700	28.842
	Fujian	0.917	192.534	15.671
	Shandong	0.421	53.962	31.507
	Guangdong	0.878	170.324	20.471
	Hainan	0.774	169.559	36.725
Northeastern Region	Liaoning	0.458	33.037	17.831
	Jilin	0.409	24.892	14.697
	Heilongjiang	0.848	23.911	3.558
Central Region	Shanxi	0.445	24.548	13.485
	Anhui	0.479	42.560	22.364
	Jiangxi	0.941	54.082	3.239
	Henan	0.409	52.212	30.758
	Hubei	0.385	59.030	36.148
	Hunan	0.900	76.180	7.538
Western Region	Inner Mongolia	0.441	22.362	12.250
	Guangxi	0.979	161.670	3.356
	Chongqing	0.316	19.989	13.436
	Sichuan	0.359	22.458	14.253
	Guizhou	0.554	24.106	10.588
	Yunnan	0.562	41.178	17.592
	Tibet	0.616	19.293	7.389
	Shaanxi	0.409	53.583	30.841
	Gansu	0.436	21.784	12.020
	Qinghai	0.322	9.388	6.415
	Ningxia	0.289	24.425	16.868
	Tibet	0.270	57.533	40.870
China		0.511	54.084	18.802

The analysis reveals marked regional disparities in KUE, fertilization intensity, and fertilizer reduction potential across provinces. Provinces with high KUE are generally associated with high potassium fertilizer inputs per unit area. Guangxi, Jiangxi, Fujian, and Hunan all exhibit KUE values above 0.90, while their per-hectare potassium fertilizer inputs remain significantly above the national average. Despite their high efficiency levels, these provinces still show measurable fertilizer reduction potential due to relatively large baseline application rates; the reducible potassium fertilizer amounts reach 15.7 kg ha⁻¹ in Fujian and 20.5 kg ha⁻¹ in Guangdong.

Provinces with medium to low KUE display substantially greater reduction potential. Xinjiang, Hubei, Hainan, Shandong, Shaanxi, and Henan record KUE values between 0.27 and 0.42, with per-hectare reducible potassium fertilizer amounts generally exceeding 30 kg ha⁻¹. Xinjiang shows the highest reduction potential, at 40.9 kg ha⁻¹.

3.3. Driving Mechanisms of Potassium Fertilizer Input and Grain Production

3.3.1. Drivers of Grain Production

To examine the mechanisms and effect magnitudes of various production factors on grain output, this study introduces control variables—such as nitrogen and phosphorus fertilizers—to eliminate confounding effects and mitigate potential bias in estimating the marginal contribution of potash fertilizer arising from omitted variables. Furthermore, pooled ordinary least squares (POLSs), fixed-effects (FEs), random-effects (REs), and two-way fixed-effects (TWFEs) models are sequentially estimated to ensure the robustness of the results (Table 6).

Table 6. Response of grain production to key factors.

Variable	POLS	FE	RE	TWFE
K	−1.382 (1.026)	10.041 *** (2.790)	5.150 ** (2.296)	15.682 *** (2.989)
Area	0.508 *** (0.006)	0.733 *** (0.017)	0.590 *** (0.013)	0.715 *** (0.018)
Power	0.277 *** (0.010)	0.266 *** (0.015)	0.298 *** (0.015)	0.331 *** (0.023)
Water	−0.003 (0.006)	−0.0002 (0.005)	0.0003 (0.005)	−0.001 (0.005)
N	−0.730 (0.409)	−1.331 ** (0.609)	−1.951 *** (0.576)	−1.843 *** (0.634)
P	−1.892 * (0.992)	−3.902 *** (1.436)	−3.237 ** (1.411)	−3.936 *** (1.438)
FHF	3.868 *** (0.324)	2.486 *** (0.481)	2.987 *** (0.456)	3.051 *** (0.489)
Time fixed effect	No	No	No	Yes
Individual fixed effects	No	Yes	Yes	Yes
Sample size	775	775	775	775
R2	-	0.864	0.853	0.872

Note: ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively; values in parentheses are standard errors; Hausman test: χ² = 124.62, p < 0.01; joint test for time effects: F = 1.75, p < 0.05; POLS, FE, RE, and TWFE denote pooled ordinary least squares, fixed effects, random effects, and two-way fixed effects models, respectively.

Hausman test results (χ² = 124.62, p < 0.01) strongly rejected the RE model, supporting the use of the FE specification. Joint F-tests for time effects (F = 1.75, p < 0.05) indicated the significance of time-fixed effects, which were accordingly included. Across all models, the TWFE model exhibited the highest explanatory power (R² = 0.872).

The model results indicate that cultivated land area and agricultural machinery power exert significant positive effects on grain production (p < 0.01). Specifically, the two-way fixed-effects model shows that a 1% increase in cultivated land area is associated with an approximately 0.72% increase in grain output, while a 1% increase in total agricultural machinery power leads to an approximately 0.33% increase in grain production, highlighting the critical roles of land input and mechanization in promoting yield growth. With respect to fertilizer inputs, compound fertilizers exhibit a robust positive effect on grain production, whereas the coefficients for nitrogen and phosphorus fertilizers are significantly negative in most model specifications. In contrast, potash fertilizer inputs display a significant positive effect after controlling for individual fixed effects, suggesting that optimizing the input structure of nitrogen, phosphorus, and potassium fertilizers plays an important role in enhancing grain productivity in these regions.

3.3.2. Drivers of Potassium Fertilizer Input

To further assess the impact of crop structure on regional potassium fertilizer demand, eight major crop categories—grain, cotton, oil crops, hemp, sugar crops, tobacco, tea, and fruits—were analyzed (Table 7). Among these, sugar crops, tobacco, and fruits are potassium-intensive crops, characterized by high potassium requirements during growth stages.

Table 7. Response of potassium fertilizer input to major crops.

Variable	POLS	FE	RE	TWFE
Grain	0.003 *** (0.001)	−0.0004 (0.001)	−0.0003 (0.001)	−0.0006 (0.001)
Cotton	0.024 * (0.013)	0.022 (0.015)	0.023 (0.014)	0.020 * (0.012)
Oil	0.018 ** (0.007)	0.011 (0.011)	0.011 (0.011)	0.010 (0.010)
Hemp	−0.311 *** (0.095)	−0.226 *** (0.083)	−0.263 *** (0.067)	−0.139 * (0.073)
Sugar	0.006 *** (0.001)	0.002 *** (0.001)	0.003 *** (0.001)	0.002 *** (0.001)
Tobacco	0.087 (0.078)	0.274 *** (0.085)	0.220 *** (0.077)	0.130 ** (0.056)
Tea	0.112 (0.115)	−0.091 * (0.049)	−0.086 * (0.049)	−0.027 (0.053)
Fruits	0.009 *** (0.002)	0.004 *** (0.001)	0.004 *** (0.001)	0.003 ** (0.002)
Time fixed effect	No	No	No	Yes
Individual fixed effects	No	Yes	Yes	Yes
Sample size	775	775	775	775
R2	—	0.555	0.553	0.718

Note: ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively; values in parentheses are cluster-robust standard errors; Hausman test: χ² = 34.05, p < 0.01; joint test for time effects: F = 44.74, p < 0.01; POLS, FE, RE, and TWFE denote pooled ordinary least squares, fixed effects, random effects, and two-way fixed effects models, respectively.

The Hausman test (χ² = 34.05, p < 0.01) supports the use of the fixed-effects model, while the joint test for time effects (F = 44.74, p < 0.01) indicates that time fixed effects should be included. The two-way fixed-effects (TWFEs) model achieves an R² of 0.718, demonstrating relatively strong explanatory power, and it is therefore adopted for further analysis. The TWFE results show that sugar crops, tobacco, and fruits exert significant positive effects on potash fertilizer inputs (p < 0.01), indicating that the expansion of potassium-intensive cash crops is a key driver of regional fertilizer demand growth. In contrast, the effects of grain crops, oil crops, and tea are not statistically significant, while fiber crops exhibit a significant negative effect and cotton shows a weakly significant positive effect. Overall, crop planting structure—particularly the spatial distribution of potassium-intensive cash crops—is an important factor contributing to regional disparities in potash fertilizer inputs.

3.4. Response of Grain Production to Potassium Fertilizer Input

3.4.1. Model Performance Evaluation

Five models—BiLSTM, TFT-BiLSTM, CNN-BiLSTM, Transformer-BiLSTM, and XGBoost-BiLSTM—were constructed to compare their predictive performances. The training and evaluation results for grain yield and potassium fertilizer inputs are summarized in Table 8 and Table 9.

Table 8. Training and evaluation metrics for grain yield prediction.

Model	Training-Yield			Test-Yield
	MAE	RMSE	R2	MAE	RMSE	R2
BiLSTM	104.972	145.077	0.993	110.717	138.737	0.942
TFT-BiLSTM	261.835	359.636	0.955	225.299	316.579	0.969
CNN-BiLSTM	91.172	122.026	0.995	130.519	159.916	0.923
Transformer-BiLSTM	322.428	489.209	0.870	522.495	846.405	0.815
XGBoost-BiLSTM	84.071	141.283	0.993	45.955	78.781	0.981

Note: MAE = mean absolute error; RMSE = root mean square error; R² = coefficient of determination.

Table 9. Training and evaluation metrics for potassium fertilizer input prediction.

Model	Training-Yield			Test-Yield
	MAE	RMSE	R2	MAE	RMSE	R2
BiLSTM	1.208	1.505	0.991	1.562	1.979	0.946
TFT-BiLSTM	6.172	8.677	0.699	6.510	8.700	0.719
CNN-BiLSTM	1.166	1.660	0.990	1.126	1.387	0.973
Transformer-BiLSTM	3.028	4.552	0.915	5.5520	8.138	0.666
XGBoost-BiLSTM	2.403	3.049	0.965	2.282	2.505	0.913

Note: MAE = mean absolute error; RMSE = root mean square error; R² = coefficient of determination.

Table 8 and Table 9 indicate substantial differences among models for both grain yield and potassium fertilizer prediction tasks. For grain yield, the XGBoost-BiLSTM model outperformed the others, achieving a test set MAE of 45.955, RMSE of 78.781, and R² of 0.981. Compared with the BiLSTM model (MAE = 110.717, R² = 0.942), the XGBoost-BiLSTM reduced the MAE by 58.5% and increased R² by 4.15%, demonstrating superior predictive accuracy and generalization capability. Moreover, its training and testing performances were well balanced, indicating strong model stability and effective mitigation of overfitting.

For potassium fertilizer input prediction, although the CNN-BiLSTM model achieved the best test set performance (MAE = 1.126, R² = 0.973), the difference between its training and test MAE reached 3.5%, suggesting potential overfitting. In contrast, the XGBoost-BiLSTM model showed more balanced performance across training and test sets (Test-K—R² = 0.913). Considering both prediction tasks, the hybrid XGBoost-BiLSTM model, by integrating XGBoost’s nonlinear feature learning and BiLSTM’s temporal modeling capabilities, demonstrated superior overall performance in predictive accuracy, generalization, and robustness, making it suitable for multi-objective agricultural production prediction tasks.

Figure 3 shows the comparison results between the predicted data of the XGBoost-BiLSTM hybrid model and the actual data for the period 2020–2024. From grain yield predictions, the model demonstrated an excellent fit, with R² = 0.993 and a mean absolute percentage error (MAPE) of 9.22%. The scatter points were closely distributed around the 1:1 reference line, indicating that spatial yield variations among provinces were accurately captured. For potassium fertilizer input (K), the model achieved R² = 0.974, capturing overall trends well; however, MAPE was relatively high at 193.7%, primarily due to low input provinces such as Tibet and Qinghai, where small base values amplified relative errors. These outliers had minimal impact on national or regional-scale trend assessments. On an annual average basis, both actual and predicted grain yields showed consistent year-on-year increases during 2020–2024, while potassium fertilizer input exhibited a continuous decline, accurately reflecting the direction and magnitude of changes. Overall, the XGBoost-BiLSTM model demonstrated high reliability for both yield and fertilizer input predictions, providing a robust reference for regional agricultural decision-making.

Figure 3. Scatter Plot of Predicted vs. Actual Values from 2020 to 2024 Based on XGBoost-BiLSTM Hybrid Model.

3.4.2. Forecast Results

Figure 4 shows the predicted relationship between grain production and potassium fertilizer inputs in China for the period 2025–2030, based on the XGBoost-BiLSTM model.

Figure 4. Trends of Grain Production and Potassium Fertilizer Inputs in China from 2000 to 2024 and Forecasts for Subsequent Years.

From 2000 to 2024, total grain production in China demonstrated a continuous upward trend, rising from 508.4 Mt in 2000 to 695.4 Mt in 2024, with an average annual increment of approximately 7.8 Mt. Total potassium fertilizer inputs followed an inverted U-shaped pattern, increasing from 3.66 Mt in 2000 to a historical peak of 6.42 Mt in 2015, followed by a gradual decline thereafter.

Forecast results indicate that from 2025 to 2030, grain production is projected to continue increasing at an average annual rate of approximately 1.1%, reaching 739.7 Mt by 2030. Conversely, potassium fertilizer input is expected to decrease at an average annual rate of 3.1%, declining from 4.81 Mt in 2024 to 4.0 Mt in 2030. A minor rebound in potassium fertilizer input in 2030 compared to 2029 is primarily attributed to the expansion of cultivated arable land.

4. Discussion

4.1. Driving Mechanisms of Decoupling Evolution

The observed shift in the decoupling relationship between potassium fertilizer input and grain production in China is consistent with the study of Xiong and Zhao [27], who reported dynamic interactions between fertilizer input and crop yield. This alignment further validates that post-2015 Chinese agriculture has transitioned from a high-input mode to a green development model characterized by “reducing inputs while increasing output”. This transformation results from coordinated policy, technological, and market forces [49]. At the policy level, the 2015 issuance of the “Action Plan for Zero Growth of Fertilizer Use by 2020” marked a pivotal milestone, shifting national agricultural objectives from solely ensuring supply and increasing output to a multidimensional focus on yield stability, quality enhancement, and ecological security [50]. Technological innovations, including controlled-release fertilizers, integrated water–fertilizer management, and variable-rate fertilization, have provided the technical foundation for reducing inputs while maintaining or increasing output, thereby improving nutrient use efficiency [51]. Market-driven price fluctuations have further encouraged farmers to optimize fertilizer application, reducing excessive use and improving the cost-effectiveness of input–output relationships, indirectly contributing to the decoupling of potassium input from grain yield [52].

The observed spatial gradient in decoupling states, following an Eastern–Northeastern–Central pattern, reflects the uneven progression of agricultural modernization. Eastern provinces, benefiting from higher economic development and greater accessibility to agricultural services, were the first to adopt precision fertilization at scale [53]. Northeastern provinces leveraged large-scale farming to enhance potassium use efficiency, whereas parts of the Central and Western regions experienced delayed decoupling due to geographic constraints and weaker technology dissemination networks [54,55].

4.2. Mechanisms Underlying Spatial Variations in Potassium Use Efficiency

Differences in regional natural conditions are a key factor contributing to the significant spatial heterogeneity in potash fertilizer use [56]. The southern regions are predominantly characterized by acidic red soils, with low baseline potassium content and strong leaching effects, but favorable water and heat conditions and high cropping intensity. The central and northeastern plains are dominated by black soils and meadow soils, with good basic fertility and concurrent rainfall and heat. The northwestern region is mainly composed of desert soils, where water availability is the primary limiting factor.

The differences in natural conditions lead to variations in planting structures, which are the main factors driving the divergence in potash fertilizer inputs between the north and south. In the southern provinces, potassium-intensive cash crops, such as sugarcane, citrus, and tobacco, dominate, and their potassium demand is much higher than that of the grain crops, such as wheat and corn, in the north [57,58]. Additionally, due to the low baseline potassium levels in the soil, more potash fertilizer is required to ensure normal crop growth. Moreover, compared to the north, southern regions tend to have more precise field management and higher levels of agricultural intensification, which collectively contribute to the efficient use of potash fertilizer, forming a “high in the south, low in the north” spatial pattern of KUE [59,60].

Provinces such as Guangxi, Jiangxi, Fujian, and Hunan exhibit a “high input—high output—high efficiency” characteristic, where both potash fertilizer use efficiency and input amounts are high. This phenomenon results from the combined effect of regional planting structures and agricultural management practices [61,62]. On one hand, these regions widely cultivate potassium-intensive cash crops, and the acidic red soils naturally prompt farmers to increase potash fertilizer inputs per unit area [63]. On the other hand, the application of modern agricultural techniques such as soil testing and formula fertilization, and integrated water and fertilizer management, has effectively improved the absorption and conversion efficiency of potash fertilizer by crops, thus forming a positive feedback loop. This suggests that efficient potash fertilizer use does not necessarily depend on low-input models but can be achieved by optimizing crop planting structures and improving agricultural management practices, leading to a synergistic increase in both fertilizer input and efficiency [64].

4.3. Yield Effects of Input Factors

Fertilizer structure and agricultural practices are core determinants of crop yield [65]. Compound fertilizers have a significant positive effect on yield, while single nitrogen and phosphorus fertilizers exhibit negative impacts. This conclusion is empirically supported by the fixed-effects model results in Table 6. This apparent contradiction reflects structural imbalances in fertilizer application in certain provinces, where excessive single-nutrient inputs cause antagonistic effects and soil acidification, thereby limiting crop uptake of other nutrients [66]. Compound fertilizers, by optimizing N–P–K ratios and supplying essential microelements, better match crop nutrient requirements and soil nutrient supply, promoting synergistic nutrient uptake and efficient nutrient conversion [10,67].

Changes in agricultural practices, particularly mechanization, further enhance yield. Mechanized fertilization enables precise nutrient application, adjusting both rate and depth according to field heterogeneity and crop growth stage, thereby reducing nutrient losses and improving application efficiency [68]. Mechanization also increases field management efficiency, reduces labor costs, and supports yield improvement [69]. However, the observed higher output elasticity of cultivated area compared with mechanization indicates that yield increases in China still partially depend on land expansion, highlighting that some regions have not fully transitioned from “extensive expansion” to “intensive optimization” [70].

4.4. Mechanisms of “Increased Output with Reduced Input” and Comparative Analysis

The intensity of fertilizer use in China has shown a phased transition from an increasing to a decreasing trend over the long term. Since the implementation of policies such as the “Zero Growth Action on Fertilizer Use” in 2015, total fertilizer input has continued to decline, and the agricultural growth model has gradually shifted from being factor-driven to efficiency-driven [63,71]. However, as of 2024, China’s fertilizer application intensity still exceeds the international safety threshold, indicating that while fertilizer reduction policies have achieved some success, the pressure on resources and the environment has not been fundamentally alleviated. Forecasts suggest that in the future, China’s grain production and potash fertilizer input will continue to exhibit an “increased output with reduced input” optimization trend, reflecting China’s confidence in improving agricultural resource use efficiency and signaling a transition toward a more efficient and greener agricultural production model [6,7].

However, the path to reducing fertilizer use relies not only on improvements in technical efficiency but also on changes in crop planting structures, which exhibit significant regional heterogeneity [15]. Additionally, soil potassium deficiencies and fluctuations in international potash fertilizer prices present exogenous constraints that may undermine the stability of the reduction path. Therefore, potash fertilizer management strategies must shift from a singular focus on reduction goals to optimal inputs under constraint conditions. By adjusting fertilizer structures, promoting agricultural mechanization, improving crop varieties, and optimizing field management, a multifaceted approach can collectively drive the transformation of agricultural production from “extensive growth reliant on high-input factors” to “intensive growth reliant on efficient resource utilization,” thus achieving a balance between input efficiency and ecological constraints [40].

4.5. Limitations

Although this study has constructed a relatively systematic analytical framework, there are still certain limitations. Based on data from the China Statistical Yearbook, the study analyzes the driving mechanisms of potash fertilizer input and grain production at a macro scale, providing initial empirical evidence for related research. However, due to data availability constraints, the study does not fully incorporate heterogeneous factors such as regional climate, soil nutrient conditions, and field management practices. As a result, the identification of fine-grained mechanisms, the testing of regional applicability, and the predictive capacity under different environmental conditions still need improvement. Future research could build upon this macro-level analysis, incorporating regional differences to conduct more detailed and mechanistic studies. Additionally, long-term field trials and multi-environment validation should be introduced to systematically examine the interactive effects of different soil types, irrigation conditions, and crop varieties on potash fertilizer use efficiency. Such efforts would enhance the reliability of predictive models and provide more accurate evidence for scientific fertilization and grain production management.

5. Conclusions and Recommendations

5.1. Conclusions

Based on agricultural panel data for China’s 31 provinces from 2000 to 2024 (China Statistical Yearbook), this study applied the Tapio decoupling model, SFA, fixed-effects models, and an XGBoost-BiLSTM hybrid deep learning model to systematically investigate the dynamic relationship, efficiency, and future trends of potassium fertilizer use and grain production. The main findings are as follows:

• Fundamental shift in decoupling relationship: Between 2000 and 2024, the overall decoupling pattern changed from predominantly “expansive negative decoupling” to “strong decoupling,” with an “Eastern—Northeastern—Central” spatial gradient, reflecting the influence of regional modernization and technology dissemination on decoupling evolution.
• KUE steadily improved with significant spatial variation: National KUE increased overall, displaying a “high South—low North” pattern. Southern high-yield regions, despite high efficiency, maintain a substantial base of fertilizer application, leaving room for reduction, whereas low- to mid-efficiency provinces have even greater reduction potential.
• Input structure optimization drives yield gains: Sown area, total agricultural mechanization, potassium, and compound fertilizer inputs positively influence yield. The cultivation of potassium-demanding crops (sugarcane, tobacco, fruits) is a key driver of regional potassium demand.
• Sustained trend of “reducing input while increasing efficiency”: Projections indicate that from 2025 to 2030, grain production will grow at an annual rate of 1.2–1.5%, while potassium input will decline at 2–4% annually.

5.2. Recommendations

To further optimize potassium use and ensure food security, the following recommendations are proposed:

• Implement region-specific fertilization strategies: In line with the “high South—low North” KUE pattern, southern provinces should exploit reduction potential from high input bases, northern major grain-producing areas should accelerate the adoption of soil-testing-based precision fertilization, and northwest arid regions should integrate water-fertilizer management to determine optimal potassium thresholds.
• Optimize fertilizer input structure: Considering the diminishing marginal returns of excessive single-component nitrogen and phosphorus fertilizers, policy guidance and technical demonstrations should encourage gradual adjustment toward higher compound fertilizer and potassium ratios, correcting the historical “high N-P, low K” bias and ensuring balanced nutrient supply.
• Establish soil nutrient monitoring and early-warning systems: To mitigate risks from decreasing potassium application, dynamic soil nutrient monitoring networks should be constructed in major agricultural regions, with scientifically defined potassium balance thresholds to prevent soil fertility decline.
• Strengthen technological innovation and policy coordination: Stakeholders should promote research, development, and the adoption of controlled-release potassium fertilizers and intelligent fertilization technologies; improve extension services; and enhance strategic potassium reserves and recycling to reduce external dependency and ensure food security.