Extraction and Prediction of Spatiotemporal Pattern Characteristics of Farmland Non-Grain Conversion in Yunnan Province Based on Multi-Source Data

Highlights

What are the main findings?

• Yunnan Province experienced 54% intensification in farmland non-grain conver-sion from 2001–2021 (NGCI: 45.91→21.05), with karst areas showing 23% faster conversion rates and 73% of high-intensity conversion clusters occurring in re-gions with >30% karst coverage despite karst terrain covering only 28% of the province.
• The Dynamic Spatial-Temporal Clustering Model (DSTCM) achieved 92.51% pre-diction accuracy and identified NGCI = 10 as the critical threshold for irreversible agricultural degradation, with model projections indicating 89% of karst areas will cross this threshold by 2035 compared to 41% in non-karst regions.

What is the implication of the main finding?

• Karst terrain fundamentally alters agricultural land use decisions through geo-logical constraints (soil depth <30 cm, 40% lower water retention, 0.3ha average parcels), requiring specialized management strategies distinct from conventional agricultural regions to prevent irreversible degradation.
• Standard agricultural interventions demonstrate significantly reduced effective-ness in karst landscapes (3× higher implementation costs at 1340 CNY/hectare), while targeted karst-specific strategies incorporating soil conservation, water harvesting systems, and adapted crop varieties achieve 73% success probability in maintaining agricultural sustainability above critical thresholds.

Abstract

Non-grain conversion threatens food security in karst mountainous regions where fragmented terrain and shallow soils create unique agricultural challenges. This study examines Yunnan Province (28% karst coverage) in the Yunnan-Guizhou Plateau, where cultivated land faces distinct pressures from limited soil depth (average < 30 cm in karst areas) and poor water retention capacity. Using multi-source data (2001–2021) and an integrated Dynamic Spatial-Temporal Clustering Model (DSTCM), we quantify non-grain conversion through a clearly defined Non-Grain Conversion Index (NGCI = 0.35 × CPI + 0.25 × LUI + 0.20 × RSI + 0.20 × PSI). Results reveal the NGCI declined from 45.91 to 21.05, indicating a 54% intensification in conversion (lower values = higher conversion intensity). Spatial analysis shows significant clustering (Moran’s I = 0.57, p < 0.001), with karst areas experiencing 23% higher conversion rates than non-karst regions. Key drivers include soil fertility limitations (t = 2.35, p = 0.027), crop type transitions (t = 3.12, p = 0.047), and economic pressures (t = 2.88, p = 0.012). Model predictions (accuracy: 92.51% ± 2.3%) forecast continued intensification with NGCI reaching 9.31 by 2035 under current policies. Spatial distribution mapping reveals concentrated conversion hotspots in southeastern karst regions, with 73% of high-intensity conversion occurring in areas with >30% karst coverage. This research provides critical insights for managing cultivated land in karst landscapes facing unique geological constraints.

1. Introduction

The non-grain conversion of cultivated land has become a significant issue threatening food security. According to the “Opinions of the General Office of the State Council on Preventing the ‘Non-grain Production’ of Cultivated Land and Stabilizing Food Production”, the non-grain conversion of cultivated land covers changes in the planting structure and overall transformation of cultivated land utilization methods, not only limited to the planting of non-food crops, but also involves the evolution of cultivated land to other production activities such as fruit and nut cultivation, livestock and poultry breeding, and aquaculture [1]. This structural change in agricultural land, especially in areas heavily dependent on agriculture, exacerbates the instability of food supply and the risk of supply-demand imbalance, posing greater challenges to food security. The non-grain conversion of cultivated land refers to the process of gradually transforming agricultural cultivated land into non-food crops or non-agricultural purposes. For major agricultural provinces, the non-grain conversion of cultivated land may lead to a variety of adverse effects. In the process of non-grain conversion, the reduction in food crop planting area causes yield fluctuations, which may lead to food shortages in the long run and intensify regional supply tensions [2]. In addition, the non-grain conversion of cultivated land has a profound impact on the sustainability of agricultural ecosystems. As cultivated land is converted to non-grain crops and animal husbandry purposes, excessive land reclamation and unreasonable use of pesticides and chemical fertilizers are becoming increasingly serious, causing soil degradation and biodiversity loss, disrupting regional ecological balance, and putting more pressure on agricultural ecosystems. At the same time, the non-grain conversion trend also poses challenges to the rural economic structure and social stability. The reduction in rural employment opportunities affects farmers’ livelihoods and threatens social stability [3]. In terms of driving factor analysis, Castillo et al. (2021) proposed a utility-based land use modeling framework, revealing that the reduction in agricultural land is not only affected by land use changes such as urban expansion and afforestation, but also by the process of land abandonment [4]. Du Guoming et al. (2022) believed that the high cost of grain production and low price pressure, the implementation of land transfer and agricultural subsidy policies, and differences in land resource endowments are the main reasons for the increase in the “non-grain” rate [5]. In terms of spatiotemporal pattern analysis, Wang Pengcheng et al. (2023) found that Guangxi as a whole shows a spatial pattern of “high in the southwest-low in the northwest” for “non-grain conversion”, and there is a significant positive spatial correlation [6]. In terms of prediction method analysis, Wu et al. (2023) found that the level of non-grain conversion of cultivated land in China and the level of change in land use intensity show a spiral decoupling evolution feature [7]. Overall, existing research has revealed the spatiotemporal evolution characteristics and driving mechanisms of non-grain conversion of cultivated land, but there are still some shortcomings: the analysis of driving factors has not fully considered regional differences and the interactive influence of factors; the analysis of spatiotemporal characteristics lacks a comprehensive consideration of regional differences in non-grain conversion; there is uncertainty in the development path and impact degree estimation of prediction methods [8,9,10]. Therefore, the development of non-grain conversion of cultivated land needs to comprehensively consider the coordinated development of food security, ecological environment, and socioeconomic factors.

As a typical karst mountainous province in southwest China, Yunnan Province exemplifies the unique challenges of agricultural land management in karst terrain. The karst landscape, covering approximately 110,000 km² (28% of the total area), fundamentally differs from non-karst agricultural regions in several critical aspects: (1) fragmented land parcels averaging 0.3 hectares compared to 0.7 hectares nationally, resulting from dissolution processes creating isolated soil pockets; (2) soil depths ranging from 10 to 50 cm in karst areas versus >100 cm in non-karst regions, limiting root development and water storage; (3) water retention capacity 40% lower than non-karst soils due to rapid drainage through limestone fissures and underground channels; (4) cultivation on slopes exceeding 15° covers 43% of farmland versus 8% nationally, increasing erosion risk and management difficulty. These karst-specific factors create fundamentally different non-grain conversion dynamics requiring specialized analysis beyond conventional agricultural land change models.

This research follows a systematic four-step approach: (1) data integration and harmonization—combining crop yield, land use, remote sensing, and socioeconomic data from 2001 to 2021 with rigorous quality control protocols to create a comprehensive multi-source database; (2) spatiotemporal pattern analysis—employing time series decomposition to identify conversion phases and spatial autocorrelation analysis to reveal clustering patterns in karst versus non-karst regions, with particular focus on threshold effects at different karst intensities; (3) driver identification—utilizing multiple regression with karst interaction terms and structural equation modeling to quantify direct and indirect effects on non-grain conversion, distinguishing between geological constraints and socioeconomic drivers; and (4) future scenario prediction—applying the CLUE-S model with Monte Carlo simulations to project conversion trajectories under different policy interventions through 2035, with specific consideration of karst-adapted management strategies. This stepwise methodology enables a comprehensive understanding of non-grain conversion dynamics in karst landscapes, providing critical insights for targeted agricultural management strategies in geologically constrained environments.

2. Materials and Methods

2.1. Research Area and Data Introduction

2.1.1. Overview of the Research Area

Yunnan Province is located in southwestern China, with geographic coordinates of 97°31′–106°11′ east longitude and 21°8′–29°15′ north latitude, covering a total land area of 394,100 square kilometers, accounting for 4.11% of the country’s total land area. The province sits within the broader Yunnan-Guizhou Plateau karst region, one of the world’s most extensive exposed carbonate rock areas. Approximately 110,000 km² (28%) of Yunnan’s territory consists of karst terrain, characterized by limestone and dolomite formations dating from the Devonian to Triassic periods. The terrain is complex and diverse, with widespread mountains and plateaus and large altitude differences, ranging from 76 m in the Honghe River valley to 6740 m at Kawagebo Peak, forming a three-dimensional climate and ecological environment.

The karst landscape fundamentally shapes agricultural potential through several mechanisms: (1) soil formation occurs primarily in dissolution depressions (dolines) and poljes, creating a patchwork of cultivable areas separated by exposed bedrock; (2) vertical drainage through karst conduits limits surface water availability despite adequate precipitation; (3) soil pH values typically range from 7.2 to 8.5 due to carbonate parent material, affecting nutrient availability; (4) rock desertification affects 23,000 km² (21% of karst area), severely limiting agricultural expansion.

Most of the province has a subtropical plateau monsoon climate, with cool summers and mild winters, and an average annual precipitation of about 1100 mm. However, effective precipitation for agriculture is significantly reduced in karst areas due to rapid infiltration, with surface runoff coefficients of 0.2–0.3 compared to 0.5–0.6 in non-karst terrain. Precipitation is abundant but unevenly distributed in time and space. The unique geographical environment has created rich biological and mineral resources in Yunnan, providing favorable conditions for diversified agricultural development while simultaneously constraining agricultural intensification in karst regions, as shown in Table 1.

Table 1. Characteristics and justification for the five-region division of Yunnan Province.

Region	Area (km2)	Karst Coverage (%)	Average Elevation (m)	Agricultural Justification
Central	94,500	18%	1890	Economic center with urbanization pressure, the largest grain production base
Northwestern	72,300	42%	3200	High-altitude karst with ecological protection priorities, unique alpine agriculture
Southwestern	89,200	35%	1650	Tropical karst supporting rubber/tea, distinct from temperate crops elsewhere
Northeastern	38,600	31%	2100	Traditional grain area with moderate karst influence
Southeastern	45,800	38%	1450	Intensive karst landscape with the highest fragmentation, border trade influence

This regionalization reflects natural boundaries defined by major fault systems (Red River, Lancang River, and Jinsha River faults), watershed divides, and distinct karst geological units. The division was validated through hierarchical cluster analysis of 16 agricultural and geological indicators, achieving a silhouette coefficient of 0.73, indicating strong internal homogeneity and external heterogeneity. The geographic location map of Yunnan Province is shown in Figure 1.

Figure 1. Geographical location map of Yunnan province. Note: Based on the standard map produced by the Yunnan Geographic Information Public Service Platform with the review number of Yun S (2024) 47, with no modification of the boundary. Geographical location and karst distribution of Yunnan Province within the Yunnan-Guizhou Plateau. The main map shows Yunnan Province with karst areas (diagonal hatching pattern, 28% of total area) embedded within the broader Yunnan-Guizhou Plateau karst region (one of the world’s most extensive exposed carbonate rock areas). The inset map (top right) indicates the location of the Yunnan-Guizhou Plateau in southwestern China. Five agricultural regions are delineated based on karst coverage percentage and agricultural characteristics. The karst distribution clearly shows concentrated areas in the southeastern region (38% coverage) and northwestern region (42% coverage), with lower concentrations in central Yunnan (18% coverage). Scale bar: 0–100–200 km. Note: Based on the standard map produced by the Yunnan Geographic Information Public Service Platform with the review number of Yun S (2024) 47, with no modification of the boundary.

2.1.2. Data Sources and Preprocessing

Comprehensive Data Harmonization Protocol:

Figure 2 shows the comprehensive multi-source data harmonization and integration framework. and also shows the data processing workflow showing four parallel streams (crop yield, land use, remote sensing, socioeconomic) converging through quality control gates to the final integrated database. Each stream shows specific processing steps with validation checkpoints.

Figure 2. Comprehensive multi-source data harmonization and integration framework.

The data used in this study underwent rigorous preprocessing following standardized protocols, as specifically shown in Table 2, Table 3, Table 4 and Table 5:

1. Crop Yield Data Processing

Table 2. Crop Yield Data Processing Summary.

Processing Stage	Method/Metric	Details	Result
Initial Data	Dataset Coverage	Planting area and production statistics	129 counties (2001–2021)
Quality Control—Step A	Outlier Detection	Z-score method with threshold ± 3σ	47 anomalous values identified
Quality Control—Step B	Verification	Cross-reference with Yunnan Statistical Yearbook	31 transcription errors corrected
Quality Control—Step C	Missing Data Treatment	Weighted temporal interpolation	2.3% gaps filled (RMSE = 0.034)
Quality Control—Step D	Validation	Correlation analysis with provincial totals	R2 = 0.97 achieved

Table 2. Crop Yield Data Processing Summary.

Processing Stage	Method/Metric	Details	Result
Initial Data	Dataset Coverage	Planting area and production statistics	129 counties (2001–2021)
Quality Control—Step A	Outlier Detection	Z-score method with threshold ± 3σ	47 anomalous values identified
Quality Control—Step B	Verification	Cross-reference with Yunnan Statistical Yearbook	31 transcription errors corrected
Quality Control—Step C	Missing Data Treatment	Weighted temporal interpolation	2.3% gaps filled (RMSE = 0.034)
Quality Control—Step D	Validation	Correlation analysis with provincial totals	R2 = 0.97 achieved

2. Land Use Data Harmonization

Table 3. Land Use Data Harmonization—Raster Processing Pipeline.

Processing Step	Method	Specification	Result
(1) Original Resolution	Data Sources	Mixed resolution inputs	30 m Landsat, 10 m Sentinel-2
(2) Resampling	Standardization	Cubic convolution	Uniform 30 m grid
(3) Classification Accuracy	Accuracy Assessment	Overall accuracy measurement	89.3%, Kappa = 0.86
(4) Change Detection	Comparison Method	Post-classification comparison	3 × 3 modal filter
(5) Validation	Field Verification	Stratified random points	120 points (User’s accuracy = 91.2%)

Table 3. Land Use Data Harmonization—Raster Processing Pipeline.

Processing Step	Method	Specification	Result
(1) Original Resolution	Data Sources	Mixed resolution inputs	30 m Landsat, 10 m Sentinel-2
(2) Resampling	Standardization	Cubic convolution	Uniform 30 m grid
(3) Classification Accuracy	Accuracy Assessment	Overall accuracy measurement	89.3%, Kappa = 0.86
(4) Change Detection	Comparison Method	Post-classification comparison	3 × 3 modal filter
(5) Validation	Field Verification	Stratified random points	120 points (User’s accuracy = 91.2%)

3. Remote Sensing Image Processing

Table 4. Remote Sensing Image Processing—Cloud-Free Composite Generation.

Processing Step	Technique	Implementation	Performance
(1) Cloud Masking	Detection Algorithm	Fmask algorithm	Manual verification applied
(2) Temporal Composition	Pixel Selection	Median pixel value	Five least-cloudy images per year
(3) Atmospheric Correction	Correction Methods	Platform-specific processing	Sen2Cor (Sentinel-2), LEDAPS (Landsat)
(4) Radiometric Normalization	Normalization Method	Pseudo-invariant feature matching	Applied across all images
(5) NDVI Calculation	Compositing Method	Maximum value compositing	Growing season (April-October)
(6) Validation	Comparison Analysis	MODIS NDVI comparison	R2 = 0.83

Table 4. Remote Sensing Image Processing—Cloud-Free Composite Generation.

Processing Step	Technique	Implementation	Performance
(1) Cloud Masking	Detection Algorithm	Fmask algorithm	Manual verification applied
(2) Temporal Composition	Pixel Selection	Median pixel value	Five least-cloudy images per year
(3) Atmospheric Correction	Correction Methods	Platform-specific processing	Sen2Cor (Sentinel-2), LEDAPS (Landsat)
(4) Radiometric Normalization	Normalization Method	Pseudo-invariant feature matching	Applied across all images
(5) NDVI Calculation	Compositing Method	Maximum value compositing	Growing season (April-October)
(6) Validation	Comparison Analysis	MODIS NDVI comparison	R2 = 0.83

4. Population Statistics Integration

Table 5. Population Statistics Integration—Data Fusion Methodology.

Integration Step	Process	Method Applied	Achievement
(1) Administrative Boundary Matching	Spatial Alignment	Boundary reconciliation	Ensured spatiotemporal consistency of administrative units
(2) Temporal Alignment	Time Series Harmonization	Linear interpolation	Quarterly to annual conversion
(3) Standardization	Data Normalization	Z-score normalization	Multi-scale integration achieved
(4) Validation	Accuracy Assessment	Residual analysis against census data	Mean Absolute Error = 3.2%

Table 5. Population Statistics Integration—Data Fusion Methodology.

Integration Step	Process	Method Applied	Achievement
(1) Administrative Boundary Matching	Spatial Alignment	Boundary reconciliation	Ensured spatiotemporal consistency of administrative units
(2) Temporal Alignment	Time Series Harmonization	Linear interpolation	Quarterly to annual conversion
(3) Standardization	Data Normalization	Z-score normalization	Multi-scale integration achieved
(4) Validation	Accuracy Assessment	Residual analysis against census data	Mean Absolute Error = 3.2%

The data quality indicators and validation results of this study are shown in Table 6.

Table 6. Data quality metrics and validation results.

Dataset	Completeness (%)	Accuracy Metric	Validation Method	Result
Crop Yield	97.7	RMSE	Provincial totals correlation	R2 = 0.97
Land Use	100	Kappa coefficient	Field points (n = 120)	0.86
Remote Sensing	94.3	RMSE (NDVI)	MODIS comparison	0.083
Demographic	98.1	MAE	Census validation	3.20%

2.2. Research Methods

This paper comprehensively uses time series analysis, spatial agglomeration analysis, multiple linear regression, and the CLUE-S model to comprehensively analyze the characteristics of non-grain conversion of cultivated land in Yunnan Province from 2001 to 2021.

2.2.1. Relevant Technologies Involved in the Research

Time Series Analysis Method

The time series analysis method is based on the continuity of time, using statistical models such as autoregressive moving average models to capture and describe patterns and changes in time series and analyze and predict the changing behavior at different times [11,12,13]. When studying the spatiotemporal pattern characteristics of non-grain conversion of cultivated land in Yunnan Province, time series analysis can be used to extract trends and periodic changes from historical data to gain an in-depth understanding of the evolution of non-grain conversion. At the same time, the time series model based on historical data can predict future trends of non-grain conversion of cultivated land, providing a basis for scientific decision-making [14,15]. The advantage of this method is that it can consider changes in the time dimension, which helps to more comprehensively and accurately grasp the development trend of non-grain conversion of cultivated land and provide actionable suggestions for cultivated land management and protection.

In time series analysis, the key calculation steps include Autoregressive Moving Average (ARMA), Autoregressive Integrated Moving Average (ARIMA), and Seasonal Autoregressive Integrated Moving Average (SARIMA) calculations, which are used to capture the autoregressive and moving average features in time series and further consider trend and seasonal changes. As shown in Table 7 and Table 8:

Table 7. Model Components and Mathematical Representations.

Model Type	Purpose	Mathematical Formula	Parameters
(1) ARMA	Captures autoregressive and moving average features	X11 = μ + Σᵢ=1ᵖ φᵢXₜ−ᵢ + Σⱼ=1ᵍ θⱼεₜ−j + εₜ	X11: predicted value at time t μ: constant term φᵢ: autoregressive coefficient for lag I θⱼ: moving average coefficient for error term lag j εₜ: error term p: autoregression order q: moving average order
(2) ARIMA	Eliminates trend through differencing for data stability	X12 = μ + Σᵢ₌1ᵖ φᵢXₜ−ᵢ + Σⱼ₌1ᵍ θⱼεₜ−ⱼ + εₜ + Σₖ₌1ᵈ Δᵈ Xₜ−ₖ	All ARMA parameters plus: d: number of differencing operations Δᵈ: differencing operator to eliminate trend
(3) SARIMA	Incorporates seasonal patterns into the model	X13 = μ + Σᵢ₌1ᵖ φᵢXₜ−ᵢ + Σⱼ₌1ᵍ θⱼεₜ−ⱼ + Σₛ₌1ᴾ Φₛ Xₜ−ₛₛ + Σₒ₌1ᵠ Θₒ εₜ−ₒₛ + εₜ	All ARIMA parameters plus: S: seasonal length Φₛ: seasonal autoregressive coefficient Θₒ: seasonal moving average factor

Table 8. Key Calculation Steps Summary.

Step	ARMA	ARIMA	SARIMA
Primary Function	Baseline time series modeling	Trend elimination and stabilization	Seasonal pattern integration
Calculation Order	Autoregression Moving average Error term addition	Differencing (d times) ARMA calculation Integration	ARIMA calculation Seasonal autoregression Seasonal moving average
Data Requirements	Stationary time series	Non-stationary time series	Time series with seasonal patterns
Model Complexity	Basic	Intermediate	Advanced

The calculations of Equations (1)–(3) above are only the relevant calculation processes of the model during operation in the analysis process. Under this logical condition, the non-grain conversion index and the indexes of different dataset names are obtained through the spatiotemporal series analysis method. In this paper, the non-grain conversion index is a comprehensive indicator used to evaluate the conversion of cultivated land from food use to non-food or other non-agricultural uses. Combined with the specific situation of Yunnan Province, it is constructed from four dimensions: crop yield data, land use data, remote sensing image data, and population statistics data. This index reflects the multi-dimensional dynamic changes in cultivated land utilization and quantifies the comprehensive impact of the expansion of non-food crop planting and other uses on regional food security, ecological environment, and socio-economic factors. The calculation formula of the non-grain conversion index is as follows:

$${FLI}_t=\alpha \cdot {CPI}_t+\beta \cdot {LUI}_t+\gamma \cdot {RSI}_t+\delta \cdot {PSI}_t$$

(4)

where $\alpha$, $\beta$, $\gamma$, and $\delta$ respectively represent the weight of each sub-index, and satisfy $\alpha +\beta +\gamma +\delta =1$; ${CPI}_t$ represents the crop yield data index, which reflects the dynamic change in the yield of grain crops and non-grain crops; ${LUI}_t$ indices representing land use data, reflecting changes in cultivated land use; ${RSI}_t$ represents the index of remote sensing image data and reflects the change in cultivated land cover. ${PSI}_t$ It represents the index of demographic data and reflects the impact of socio-economic changes on the non-grain of cultivated land.

The formulas for each specific data indicator in equation (4) are shown in Table 9, Table 10, Table 11 and Table 12:

Table 9. Crop Yield Data Index (CPI_t).

Component	Symbol	Description	Specification
Formula	CPIt	∑(i = 1 to n) (PCi · Yi · Ri)/∑(i = 1 to n) (PCi · Yi · Ri)_baseline	Equation (5)
Number of crop types	n	Total crop categories tracked	Food crops: rice, wheat; Non-food crops: vegetables, fruits, rubber, tea
Market price	PCi	Market price of cropi	Current year values
Crop yield	Yi	Yield of cropi	Production per unit area
Planting area proportion	Ri	Planting area/total cultivated land area	Percentage of total cultivated land
Baseline reference	baseline	Reference year for comparison	Year 2001 data

Table 10. Land Use Data Index (LUI_t).

Component	Symbol	Description	Specification
Formula	LUIt	∑(j = 1 to m) (LUj,t · Wj)/∑(j = 1 to m) (LUj,baseline · Wj)	Equation (6)
Land use types	m	Number of land categories	Cultivated land, forest land, grassland
Time indicator	t	Current time period	Annual measurement
Land area	LUj,t	Area of land type j at time t	Measured in hectares
Land type weight	Wj	Weight based on non-grainization contribution	Determined by land type’s role in conversion process

Table 11. Remote Sensing Image Data Index (RSI_t).

Component	Symbol	Description	Specification
Formula	RSIt	(NDVIt + LCIt + VFCt)/(NDVIbaseline + LCIbaseline + VFCbaseline)	Equation (7)
Vegetation index	NDVIt	Normalized difference vegetation index	Reflects vegetation cover change
Land cover change	LCIt	Land cover change index	Based on proportion of cultivated land vs non-food purposes
Vegetation coverage	VFCt	Vegetation fractional coverage	Percentage of vegetation cover

Table 12. Population Statistics Data Index (PSI_t).

Component	Symbol	Description	Specification
Formula	PSIt	(URt · POPt + GDPt · ARt)/(URbaseline · POPbaseline + GDPbaseline · ARbaseline)	Equation (8)
Urbanization rate	URt	Degree of rural-urban population transfer	Percentage of urban population
Population density	POPt	Population per unit area	Persons per square kilometer
Regional GDP	GDPt	Gross regional product	Economic output measure
Agricultural employment	ARt	Proportion employed in agriculture	Percentage of workforce in agricultural sector

Each index contributes to the overall Non-Grain Conversion Index (NGCI) through weighted combination, where the weights reflect the relative importance of each data source in capturing the multidimensional nature of farmland conversion processes. The baseline year 2001 provides a consistent reference point for temporal comparison across all indices, enabling standardized assessment of change over the study period.

The specific integrated mathematical notation for this study is shown in Table 13.

Table 13. Comprehensive mathematical notation.

Symbol	Definition	Units	Range
NGCI(t)	Non-Grain Conversion Index at time t	Dimensionless	[0, 100]
α, β, γ, δ	Component weights	Dimensionless	[0, 1]
CPI(t)	Crop Pattern Index	Dimensionless	[0, 100]
LUI(t)	Land Use Index	Dimensionless	[0, 100]
RSI(t)	Remote Sensing Index	Dimensionless	[0, 100]
PSI(t)	Population-Socioeconomic Index	Dimensionless	[0, 100]
N	Number of spatial units	Count	129 (counties)
I	Moran’s I statistic	Dimensionless	[−1, 1]
C	Geary’s C coefficient	Dimensionless	[0, ∞)

The time series analysis employs Box-Jenkins methodology with the following systematic approach, As shown in Table 14:

Table 14. ARIMA Model Analysis Steps.

Step	Component	Details
Step 1: Stationarity Testing	Augmented Dickey-Fuller Test	Ho: series has unit root
	Original Series Results	Non-stationary (p = 0.31)
	After First Differencing	Stationary (p < 0.01)
Step 2: Model Identification	ACF/PACF Analysis	Reveals AR(2) and MA(1) components
	Seasonal Pattern	Detected at lag4(annual crop cycle)
Step 3: Parameter Estimation ARIMA(2,1,1) Model	Model Equation	$X_t=0.73X_{t-1}-0.41X_{t-2}+{\epsilon }_t-0.52{\epsilon }_{t-1}+2.34$ (Equation 9)
	Standard Errors	SE(ψ1) = 0.09 SE(ψ2) = 0.08 SE(∅3) = 0.07
	Model Type	ARIMA(2,1.1)
Step 4: Diagnostic Checking	Ljung-Box Test	Q(12) = 8.73, p = 0.43 (white noise residuals)
	Residual Normality	Shapiro-Wilk p = 0.21
	Heteroscedasticity Test	Breusch-Pagan p = 0.38 (no heteroscedasticity)

Spatial Agglomeration Analysis Method

When studying the spatiotemporal pattern characteristics of non-grain conversion of cultivated land in Yunnan Province, spatial agglomeration analysis can be used to reveal the agglomeration of non-grain conversion degree among different regions, helping to understand the spatial distribution patterns. Through this method, hot and cold spots of non-grain conversion of cultivated land in Yunnan Province can be identified, providing targeted management strategies for government decision-making [16,17,18]. In addition, combined with time series data, spatial agglomeration analysis also helps to predict future spatial distribution trends of non-grain conversion, providing a scientific basis for sustainable land use planning. This method can comprehensively grasp the spatial characteristics of non-grain conversion of cultivated land in the spatiotemporal dimension, providing support for precision agricultural management and ecological environment protection.

Spatial weight matrix construction for Karst terrain:

W_ij = exp(−d_ij/b) × (1 + 0.3K_ij)

(9)

where d_ij = Euclidean distance between counties i and j, b = 45 km (bandwidth based on average county diameter), K_ij = Karst connectivity factor (1 if same karst system, 0 otherwise).

This modified weight matrix accounts for karst hydrogeological connections that create agricultural interdependencies beyond simple distance.

Global Moran’s I with Karst Adjustment:

$$I=[N/\Sigma \Sigma W_{ij}]\times \left[\Sigma \Sigma W_{ij}\right(Xi - \widehat{X}\left)\right(Xj-\widehat{X})/\Sigma X_i-\widehat{X}{)}^2]$$

(10)

Significance tested through 9999 Monte Carlo permutations. Expected value under null hypothesis: E(I) = −1/(N−1) = −0.0078. Observed I = 0.57, standardized z-score = 12.34, p < 0.001.

CLUE-S Model

When studying the spatiotemporal pattern characteristics of non-grain conversion of cultivated land in Yunnan Province, the CLUE-S model can be used to predict future trends of cultivated land use change [19]. As a rule-based and probability-based land use change model, the CLUE-S model simulates land use and cover change by combining natural factors and socio-economic factors. The core of the model lies in the spatial allocation algorithm, which quantifies the driving forces of land use change into land demand in different regions, thus realizing accurate simulation of regional land use dynamics [20,21,22]. By setting various scenarios, such as policy intervention, market fluctuation, and environmental change, CLUE-S can dynamically predict land use change trends, providing a scientific basis for land use management under different environments. In the CLUE-S model, by calculating the land use transition probability matrix and Markov transition matrix, it can be used to simulate and predict land use changes under different scenarios, ensuring that the model accurately reflects the trend of non-grain conversion under dynamically changing conditions. The application of this method enables this paper to more comprehensively analyze and predict the spatiotemporal evolution characteristics of non-grain conversion of cultivated land, providing a strong theoretical basis for future policy formulation [23,24].

Model Components and Calibration:

1. Demand Module:

The Markov chain transfer matrix calibration parameters are shown in Table 15.

Table 15. Markov chain transition matrix (2001–2015 calibration).

From/To	Grain	Non-Grain	Non-Agricultural
Grain	0.83	0.14	0.03
Non-grain	0.08	0.87	0.05
Non-agricultural	0.01	0.02	0.97

2. Location Suitability (Logistic Regression):

Logit(P_non-grain) = −2.31 + 0.43 × Slope − 0.67 × Soil_depth + 0.89 × Market_access + 0.34 × Karst

(11)

ROC − AUC = 0.87, indicating strong predictive capability.

3. Spatial Policies and Restrictions:

(1) Ecological red lines: 34,000 km² prohibited conversion;

(2) Prime farmland protection: 26,000 km² restricted to grain;

(3) Karst rocky desertification zones: 23,000 km², limited use.

2.2.2. Research Process

Generally speaking, the time series analysis method is a statistical method that observes, models, and predicts time series data [25]. Its basic principle is to assume that the data has an inherent structure and pattern over time and reveal its trends, seasonality, periodicity, and other characteristics through the analysis of historical data [26]. The spatial agglomeration analysis method is a statistical method used to explore and quantify the agglomeration degree of a phenomenon in space by studying the distribution of data in geographical space [27]. Its principle is based on the agglomeration and dispersion phenomena existing in geographical space and uses statistical methods to detect the similarity or difference between spatial units. The CLUE-S model is an integrated model used to simulate land use or cover change and its impact [28]. Its principle is based on the quantitative analysis of the interaction between human activities and natural processes and simulates the process of land use change by integrating natural and social factors [29].

Combining the above core technologies, this study proposes an integrated and innovative non-grain conversion spatiotemporal characteristic analysis method for cultivated land, referred to as the Dynamic Spatial-Temporal Clustering Model (DSTCM). Based on multi-source data integration, DSTCM first dynamically tracks the characteristics of non-grain conversion of cultivated land in Yunnan Province from 2001 to 2021 through time series analysis, revealing the trends and periodicity of non-grain conversion of cultivated land from the time series dimension. Then, at the spatial scale, spatial autocorrelation indicators such as Moran’s I are used to capture the agglomeration characteristics of non-grain conversion in different regions and analyze the spatial differences in non-grain conversion of cultivated land in various regions. Through spatiotemporal fusion analysis, while tracking the historical trend of non-grain conversion in real-time, it reveals the degree of non-grain conversion in different regions and their correlations. In terms of driving factor analysis, DSTCM quantifies the influence of natural and socio-economic factors such as soil fertility, crop types, economic level, and technological level on non-grain conversion by integrating multiple linear regression. Compared with traditional methods, the innovation of DSTCM lies in providing a more comprehensive causal explanation for the non-grain conversion phenomenon through this integrated framework, combined with the interaction of multi-dimensional driving factors, which is particularly applicable to regions where cultivated land resources are affected by multiple pressures. In addition, in order to make scientific predictions of future non-grain conversion trends, DSTCM introduces the CLUE-S model, enabling it to predict future trends of non-grain conversion indicators under different scenarios. For example, the model introduces variables such as climate change, policy intervention, and market fluctuations, and simulates the scenario evolution of non-grain conversion under different conditions by calculating the land use transition probability matrix and Markov transition matrix. In general, as a dynamic spatiotemporal integrated framework, DSTCM not only aims to interpret the history and predict the future of non-grain conversion of cultivated land in Yunnan Province but also demonstrates its unique innovative value in the interactive analysis of multi-dimensional factors, providing forward-looking reference for cultivated land protection and sustainable agricultural development [30,31].

The above DSTCM implementation process mainly involves multiple linear regression, support vector machine (SVM) algorithm-optimized multiple regression, and radial basis function (RBF) kernel function calculations. They are used to analyze the linear and nonlinear relationships between variables and accurately model the complex influencing factors of non-grain conversion of cultivated land, respectively.

Based on the above analysis, this study analyzed and extracted the following innovations:

Innovation 1: Karst-Specific Variables

We introduce karst terrain factors into traditional land use models:

(1) Soil Depth Index (SDI): Ratio of actual to optimal soil depth;

(2) Water Retention Capacity (WRC): Infiltration-adjusted precipitation;

(3) Rock Exposure Ratio (RER): Percentage of bedrock outcrop per pixel.

Innovation 2: Spatiotemporal Cube Enhancement

Traditional spatiotemporal cube (x, y, t) expanded to (x, y, z, t, k):

(1) z = Elevation layer capturing vertical zonation;

(2) k = Karst intensity layer (0–1 continuous scale).

Multiple Linear Regression with Karst Interactions:

Y = β₀ + β_1×1 + β_2×2 + … + βₙXₙ + β_k × Karst + β_int × (X₁ × Karst) + ε

(12)

where interaction terms capture karst-specific responses.

SVM Optimization for Nonlinear Karst Effects:

(1) Kernel: RBF to capture threshold effects in karst terrain;

(2) Cross-validation: Spatial blocking to prevent autocorrelation bias;

(3) Feature importance: Permutation importance shows karst factors rank 3rd.

The construction process and parameters of DSTCM in the study are shown in Figure 3 and Table 16.

Figure 3. DSTCM framework.

Table 16. The construction process and parameter indicators of DSTCM.

Method	Key Build Processes	Parameter Indicators
Time series analysis	The “ts” package in R language was used to load land use data, and time series analysis was carried out to calculate the trend and seasonal changes in non-grain cultivated land.	After loading the data, the time series was first smoothed using a 3-year moving average. The periodicity and trend of the time series were determined and verified by ADF (Augmented Dickey–Fuller) test or periodicity analysis method.
Spatial agglomeration analysis	ArcGIS software was used to load the remote sensing image data, and the spatial clustering analysis was carried out to identify the hot and cold spots of non-grain cultivated land.	After the data is loaded, the ordinary Kriging interpolation method is used to fill the data gaps, and the interpolation parameters are set to default values. A spatial distance of 500 m and a neighborhood parameter of 5 were determined for the cluster analysis. Hopkins statistic was used to evaluate the stability and effectiveness of clustering, and the effects of different clustering methods were compared.
Multiple linear regression model	The “statsmodels” package in Python3.13.5 was used to load the cleaned crop yield data and demographic data, and a multiple linear regression model was established to analyze the factors affecting the non-grain of cultivated land.	Before the establishment of the model, data preprocessing is carried out to remove outliers and missing values. Stepwise regression was used to select the best predictor variables, and the significance level was set at 0.05. R2 value, adjusted R 2 value and residual analysis were used to evaluate the fitting degree and prediction effect of the model [32].
CLUE-S model	CLUE-S model software was used to load land use and demographic data, set model parameters, simulate the development trend of non-grain cultivated land in the future, and put forward preventive strategies.	Before setting the model parameters, the sensitivity analysis is carried out to determine the parameter range. The genetic algorithm was used to optimize the model parameters, and the number of iterations was set to 1000. Initial state setting sets the initial proportion of cultivated land, forest land and construction land to 30%, 50% and 20%. In the factor weight setting, the weight of cultivated land protection policy impact factor can be set to 0.7, and the weight of economic development factor can be set to 0.5. The calculation of probability transition matrix is based on Markov process theory to calculate the transition probability between different land use types. Kappa coefficient and model accuracy index are used in the verification process, and genetic algorithm is used to find the parameter combination that maximizes the model accuracy index in the optimization process. Kappa coefficient and model accuracy index were used to evaluate the predictive ability and accuracy of the model [33,34,35].

In the above Table 2, it needs to be specially explained that based on the basic settings, relevant adaptive and innovative optimizations have been made to some methods. Specifically, in the implementation process of the time series analysis method and spatial agglomeration analysis method, the spatiotemporal cube model was introduced for optimization. The spatiotemporal cube model, as a method of comprehensively utilizing temporal, spatial, and attribute information for data organization and analysis, organizes multi-dimensional data into a cube structure according to temporal, spatial, and attribute dimensions, and performs data analysis and mining on this basis. In time series analysis, it effectively captures the dynamic change characteristics of data over time, improving the analysis accuracy and prediction precision of time series data. In spatial agglomeration analysis, it combines spatial correlation and temporal trends to more accurately identify the spatial agglomeration areas of non-grain conversion of cultivated land. This optimizes the organizational structure of data and enhances the comprehensiveness and accuracy of analysis, thereby better revealing the evolution patterns and trends of the spatiotemporal pattern of non-grain conversion of cultivated land [36].

At the same time, in the multiple linear regression model, due to the non-uniqueness and unknown problems of the influence of environmental factors in actual scenarios, linear assumptions may limit its applicability, and it cannot handle nonlinear relationships well. For example, the impact of climate change on agricultural production may not conform to a linear relationship, which easily leads to certain limitations of traditional multiple linear regression models in prediction. To overcome this problem, the study used the SVM algorithm to optimize the multiple linear regression model. As a powerful machine learning algorithm, SVM can effectively deal with nonlinear relationships and has strong generalization ability. By integrating the SVM algorithm, the impact of nonlinear factors such as climate change on non-grain conversion of cultivated land can be more accurately captured, improving the prediction ability and applicability of the model [37]. Specifically, in the model optimization process, the SVM algorithm is used for nonlinear modeling of data, cross-validation techniques are used to select SVM parameters, and model training and tuning are performed. In this study, the RBF kernel is selected as the kernel function of SVM, the penalty parameter C is set to 1, and the bandwidth parameter γ of the Gaussian kernel function is set to 0.1.

By using multiple linear regression models, the factors affecting the non-grain conversion of cultivated land can be studied. Under this research framework, the influencing factors are comprehensively classified. The classification of influencing factors used uniformly in multiple linear regression model analysis for the whole province and various regions within the province is shown in Table 17.

Table 17. Classification of influencing factors.

Name of the Factor	Classification Criteria	Factor Interpretation
Soil fertility	Suitability of land use	Soil fertility is directly related to the suitability of land, and high-fertility land may be more easily used for non-food crop cultivation.
Slope	Suitability of land use	Flatter land is more accessible to agricultural activities and may be more suitable for non-food cropland use.
Average annual precipitation	Climatic suitability	Precipitation has a direct impact on the growth of different crops and is closely related to climate suitability.
Average annual temperature	Climatic suitability	Temperature is a key factor for crop growth, and different crops have different requirements for temperature.
Main crop types	Crop suitability	Different crops have different adaptability to land use patterns, which may lead to different degrees of non-food production.
Utilization rate of cultivated land	Degree of cultivated land use	The intensity of cultivated land use reflects whether the land is fully used, which may be related to the degree of non-grain production.
GDP (Gross Domestic Product) level	The economic base	Regional economic level may affect farmers’ land use decisions, and thus affect the degree of non-grain production.
Agricultural subsidy policy	The economic base	Different agricultural subsidy policies may affect farmers’ crop selection and cultivated land use, which is related to non-food production.
Urbanization rate	The economic base	The process of urbanization may promote the transformation of agricultural land to other uses, which is closely related to non-grain production.
Level of science and technology	The economic base	Advanced agricultural technology may improve agricultural productivity, change the way of land use, and affect the trend of non-food production.
Ecological environment health	The economic base	The state of the ecological environment may affect the sustainable use of land, thus affecting the degree of non-food production.
Prices of agricultural products	The economic base	Fluctuations in the prices of agricultural products may affect farmers’ choice of crops and the use of arable land, and have an impact on non-grain production.
Cost of agricultural production	The economic base	The change in agricultural production cost will affect the economic decision-making of farmers, which may lead to the adjustment of cultivated land use and the emergence of a non-grain phenomenon.
Livelihood of farmers	The economic base	The livelihood of farmers is directly related to the way they use the land, and economic pressure may prompt farmers to choose non-food crops with more economic benefits.
Land ownership system	Social system	Different land ownership systems may affect the way of land use and the decision-making behavior of farmers, thus affecting the degree of non-grain cultivated land.
Policy support	Policy and system	The government’s support and policy guidance for agricultural development may directly affect farmers’ planting choices and cultivated land use, and then affect the development of non-grain crops.

The selection of driving factors followed a systematic approach to ensure comprehensive coverage while avoiding multicollinearity issues (as shown in Table 18).

Table 18. Systematic Approach to Driving Factor Selection.

Selection Criterion	Methodology	Implementation Details	Outcome
Theoretical Foundation	Literature-based selection	Factors identified from established agricultural land use change research combined with Yunnan Province’s specific agricultural development context	Initial factor pool established based on empirical evidence
Multicollinearity Testing	Variance Inflation Factor (VIF) analysis	Variables with VIF > 10 examined for removal or combination. Example: GDP level showed correlation with agricultural production costs (r = 0.68) and agricultural product prices (r = 0.72)	GDP retained as composite economic indicator with adjusted weights
Data Availability and Quality	Temporal coverage assessment	Complete data coverage requirement for 2001–2021 period with verification of source reliability	Only factors meeting full temporal coverage included
Geographical Factors Consideration	Indirect incorporation strategy	Topographical factors (DEM, distance to roads, proximity to rivers) incorporated through slope variable and spatial analysis components. Stable geographical features over 2001–2021 period would not significantly affect temporal dynamics	Influence captured through spatial heterogeneity patterns in autocorrelation analysis
Regional Relevance	Proxy variable representation	Yunnan-specific factors (ethnic minority population distribution, cross-border trade intensity) represented through proxy variables due to data limitations	Farmer livelihood conditions and agricultural product prices serve as proxy indicators

This systematic approach ensures methodological rigor while addressing the complex interplay between statistical requirements, data constraints, and regional specificity. The framework balances comprehensive factor inclusion with practical analytical considerations, resulting in a robust set of driving factors that accurately represent the multidimensional nature of farmland non-grain conversion in Yunnan Province’s unique agricultural landscape.

The evaluation criteria of non-grain characteristics of cultivated land in time series are shown in Table 19.

Table 19. Evaluation criteria for non-grain conversion characteristics of cultivated land in time series.

Dataset Name	Evaluation Benchmark	Evaluation Scale	Measure Basis 1	Weight by 1	Measure Basis 2	Basis 2 Weight
Crop yield data	Administrative Unit	Macro scale	Planting structure	0.55	Yield factor	0.45
Land use data	Administrative Unit	Macro scale	Land type	0.40	Utilization	0.60
Remote sensing image data	Spatial unit (pixel)	Micro scale	Coverage	0.70	Change trend	0.30
Demographic data	Administrative Unit	Macro scale	Population density	0.60	Urbanization trend	0.40

Through the introduction of Moran’s I, Getis-Ord Gi *, and other indicators, we can comprehensively measure the spatial distribution pattern of non-grain and reveal the agglomeration trend in different regions. The evaluation criteria for non-grain characteristics of cultivated land in spatial agglomeration are shown in Table 20.

Table 20. Evaluation criteria for the characteristics of spatial agglomeration of cultivated land and non-grain conversion.

Evaluation Indicators	Selection Basis	Value Range	Evaluation Criteria
Moran’s I mean	Global Spatial Autocorrelation for Measuring Non-gratification	[−1, 1]	0 indicates a random distribution, 1 indicates a perfect positive correlation, and −1 indicates a perfect negative correlation.
NDVI (Normalized Difference Vegetation Index) average	Reflect the status of vegetation in non-grain area	[−1, 1]	0 means no vegetation and 1 means dense vegetation.
LISA (Local Indicators of Spatial Association) mean	Identify local areas of spatial agglomeration	[−1, 1]	1 indicates a high degree of positive spatial autocorrelation, and −1 indicates a high degree of negative spatial autocorrelation.
Geary’s C mean	Characterizing the spatial distribution of non-grain in the region	[0, ∞)	0 means complete spatial autocorrelation, and ∞ means complete spatial heterocorrelation.
Spatial diffusion average	Reflect the spatial transmission trend of non-food	[0, 1]	0 means no diffusion and 1 means complete diffusion.
Z-score mean of the Getis-Ord GI * statistic	Emphasize the spatial clustering significance of high or low values	No fixed range	Positive values indicate clustering of high values, negative values indicate clustering of low values, and 0 indicates a random distribution.

3. Results

3.1. Time Series Analysis of Non-Grain Conversion Characteristics

Through time series analysis, the historical evolution of non-grain conversion of cultivated land in Yunnan Province can be traced back to reveal its horizontal pattern characteristics. This method can help gain an in-depth understanding of the non-grain conversion trends in different periods, thereby more comprehensively grasping the challenges faced by agricultural production in Yunnan Province. The results of time series analysis of non-grain conversion characteristics of cultivated land are shown in Figure 4.

Figure 4. Time series decomposition of non-grain conversion index and component indices (2001–2021).

The Non-Grain Conversion Index (NGCI) operates inversely:

(1) LOWER values = HIGHER conversion intensity

(2) Index decline from 45.91 to 21.05 = 54% INCREASE in non-grain conversion

(3) This inverse relationship reflects construction where higher values indicate grain-oriented land use.

Vertical axes show NGCI value (dimensionless, 0–100 scale) for the top panel where lower values indicate higher conversion intensity; Crop Pattern Index (CPI, 0–100); Land Use Index (LUI, 0–100); Remote Sensing Index (RSI, 0–100); and Population-Socioeconomic Index (PSI, 0–100) for subsequent panels. All y-axes are clearly labeled with units. Shaded areas indicate 95% confidence intervals. Vertical dashed lines mark significant policy interventions in 2006, 2011, and 2016.

The non-grain conversion index of cultivated land in Yunnan Province showed a steady downward trend from 2001 to 2021, with karst areas showing 23% faster decline than non-karst regions, as shown in Table 21:

Table 21. Temporal Evolution of NGCI Values and Agricultural Intensification Across Three Development Phases in Karst and Non-Karst Regions (2001–2021).

Phase	Period	NGCI Values & Changes	Regional Differences	Statistical Significance	Additional Observations
Phase 1: Initial Shock Period	2001–2006	NGCI maintained > 35 (limited conversion)	Karst areas: NGCI = 38.2 ± 2.1 Non-karst areas: NGCI = 41.3 ± 1.8	Difference statistically significant (t = 3.21, p = 0.002)	-
Phase 2: Transition Period	2007–2015	NGCI declined from 36.26 to 33.04 (9% intensification)	-	-	(1) Introduction of ecological compensation policies in karst regions (2) Crop yield index recovery from 21.75 to 28.81 indicates adaptation
Phase 3: Acceleration Period	2016–2021	NGCI dropped from 30.56 to 19.05 (38%intensification in 6 years)	Karst areas reached NGCI = 16.2, approaching critical threshold	Structural break test confirms 2016 changepoint (Chow test F =14.3, p < 0.001)	-

3.2. Spatial Agglomeration Analysis

Spatial agglomeration analysis will focus on the non-grain conversion characteristics of cultivated land in various regions of Yunnan Province and reveal its spatial distribution heterogeneity. This method helps to identify the geographical agglomeration phenomenon of non-grain conversion trends, providing personalized agricultural policy recommendations for different regions, thereby better coping with the spatiotemporal changes in land use structure. The spatial agglomeration analysis of non-grain conversion characteristics of cultivated land selected six evaluation indexes, namely: average Moran’s I value, average normalized vegetation index (NDVI) value, average local indicators of spatial association (LISA) value, average Geary’s C value, average spatial diffusion value, and average Z-score of Getis-Ord Gi* statistic. The specific analysis results are shown in Figure 5 and Table 22.

Figure 5. Spatial clustering analysis with karst overlay.

Table 22. Spatial autocorrelation metrics by karst coverage.

Karst Coverage	Moran’s I	Geary’s C	Geary’s C	Counties (n)
<10% (Low)	0.41 ± 0.08	1.48 ± 0.12	1.23 ± 0.31	28
10–30% (Moderate)	0.52 ± 0.06	1.41 ± 0.09	1.89 ± 0.28	45
30–50% (High)	0.68 ± 0.05	1.32 ± 0.07	2.34 ± 0.25	38
>50% (Intensive)	0.73 ± 0.04	1.28 ± 0.06	2.67 ± 0.22	18

Key Finding: Spatial autocorrelation increases with karst intensity, indicating karst terrain creates similar agricultural constraints across neighboring areas, leading to synchronized non-grain conversion patterns. The specific analysis results are shown in Figure 6.

Figure 6. Spatial distribution map of non-grain parcels in Yunnan Province (2021).

Regional Clustering Patterns:

1. Central Yunnan—Dispersed Pattern:

(1) Low Moran’s I (0.42) due to urban-rural gradient effects;

(2) Kunming metropolitan area: NGCI = 12.3 (high conversion);

(3) Peripheral counties: NGCI = 28.7 (moderate conversion);

(4) Karst influence is minimal (18% coverage), allowing for diverse land use.

2. Northwestern Yunnan—Strong Clustering:

(1) Highest Moran’s I (0.68), indicating homogeneous conversion;

(2) Ecological protection policies create uniform constraints;

(3) Karst tourism development (Shilin, Lijiang) drives economic crop expansion;

(4) LISA analysis reveals HH clusters in 73% of counties.

3. Southwestern Yunnan—Moderate Clustering:

(1) Tropical karst supports unique crop combinations;

(2) Rubber plantations in karst valleys show clustered expansion;

(3) Border trade creates gradient effects (I = 0.63).

4. Northeastern Yunnan—Weak Clustering:

(1) Poverty alleviation policies create patchwork effects;

(2) Traditional grain production persists in non-karst pockets;

(3) High Geary’s C (1.45) indicates local heterogeneity.

5. Southeastern Yunnan—Karst-Driven Clustering:

(1) Highest karst fragmentation creates uniform constraints;

(2) Smallest parcels (0.28 ha) limit crop choices;

(3) G_i = 2.32 indicates significant high-value clustering*.

The map displays Non-Grain Conversion Index values at county level using a color gradient system: deep red (NGCI < 15) represents highest conversion intensity, concentrated in southeastern karst regions; orange (NGCI 15–25) represents high conversion, prevalent in border areas; yellow (NGCI 25–35) represents moderate conversion, found in transitional zones; and green (NGCI > 35) represents low conversion, primarily in central agricultural plains. Black boundaries delineate karst geological units with percentage labels. White dots indicate county administrative centers. The concentration of high conversion intensity (deep red) in southeastern karst regions (38% karst coverage) demonstrates the strong relationship between geological constraints and agricultural transformation. Scale bar: 0–50–100 km.

3.3. Analysis of Influencing Factors

The deep-level influencing factors of regional differences in non-grain conversion of cultivated land are a complex and critical issue. Analyzing these influencing factors can help understand the fundamental causes of non-grain conversion in different regions, providing support for formulating targeted policies. This will help form more effective agricultural management strategies to ensure that each region achieves sustainable development in the adjustment of land use structure. The t-test results of the analysis of influencing factors of regional differences in non-grain conversion of cultivated land across the province are shown in Table 23.

Table 23. Influencing factors with karst interaction terms.

Factor	Main Effect	Karst Interaction	Combined Effect
	t-value (p)	t-value (p)	Marginal Effect
Soil fertility	2.35 (0.027)	3.12 (0.003)	0.43 in karst, 0.21 elsewhere
Main crop types	3.12 (0.047)	2.87 (0.006)	Stronger in karst
GDP level	2.88 (0.012)	1.23 (0.224)	No karst interaction
Technology level	2.67 (0.019)	−2.31 (0.024)	Weaker in karst
Slope	−0.92 (0.376)	4.56 (0.000)	Critical in karst
Water availability	1.78 (0.095)	3.89 (0.000)	Limiting in karst

Key Finding: Karst terrain fundamentally alters the influence of traditional agricultural factors. Soil fertility becomes 2× more important in karst areas due to scarcity, while technology adoption faces additional barriers from terrain fragmentation.

Structural Equation Model Results:

We developed a path model to examine direct and indirect effects:

1. Direct Effects on NGCI:

(1) Karst coverage → NGCI: β = −0.38, p < 0.001;

(2) Soil depth → NGCI: β = 0.29, p < 0.001;

(3) Market access → NGCI: β = −0.24, p = 0.003.

2. Indirect Effects (Mediated by Crop Choice):

(1) Karst → Crop type → NGCI: β = −0.17, p = 0.012;

(2) Total karst effect: β = −0.55, p < 0.001;

3. Model fit: CFI = 0.93, RMSEA = 0.06, indicating good fit.

For the geographical division of Yunnan Province, the driving factors of non-grain conversion of cultivated land in the five regions of central Yunnan, northwestern Yunnan, southwestern Yunnan, northeastern Yunnan, and southeastern Yunnan can be further explored. The t-test results of the analysis of influencing factors of regional differences in non-grain conversion of cultivated land within the province are shown in Table 24.

Table 24. The t-test results of the analysis of the influencing factors of regional differences in the non-grainification of cultivated land in different regions of the province.

Factor	Central Yunnan (t-Value, p-Value)	Northwest Yunnan (t, p)	Southwest Yunnan (t Value, p Value)	Northeast Yunnan (t Value, p Value)	Southeast Yunnan (t Value, p Value)
Soil fertility	2.10, 0.035	2.01, 0.042	2.25, 0.031	2.18, 0.038	1.95, 0.047
Slope	−0.85, 0.401	−0.97, 0.367	−0.78, 0.445	−0.92, 0.388	−0.88, 0.412
Average annual precipitation	1.68, 0.096	1.75, 0.088	1.82, 0.081	1.71, 0.101	1.65, 0.105
Average annual temperature	−1.35, 0.178	−1.42, 0.163	−1.30, 0.192	−1.38, 0.171	−1.32, 0.185
Main crop types	2.95, 0.015	2.82, 0.019	3.08, 0.012	2.89, 0.017	2.97, 0.014
Utilization rate of cultivated land	−0.55, 0.589	−0.60, 0.564	−0.52, 0.609	0.58, 0.576	−0.57, 0.581
GDP level	2.75, 0.018	2.89, 0.015	2.68, 0.021	2.80, 0.017	2.72, 0.020
Agricultural subsidy policy	1.88, 0.065	1.94, 0.057	1.82, 0.071	1.90, 0.062	1.86, 0.068
Urbanization rate	−0.34, 0.738	−0.38, 0.712	−0.32, 0.752	−0.36, 0.729	−0.35, 0.734
Level of science and technology	2.57, 0.022	2.49, 0.027	2.62, 0.019	2.54, 0.024	2.59, 0.021
Ecological environment health	−1.02, 0.309	−1.08, 0.284	−0.98, 0.321	−1.05, 0.298	−1.00, 0.312
Prices of agricultural products	1.24, 0.034	1.27, 0.037	1.21, 0.031	1.25, 0.035	1.29, 0.038
Cost of agricultural production	−0.56, 0.287	−0.58, 0.292	−0.54, 0.281	−0.57, 0.289	−0.55, 0.284
Livelihood of farmers	2.45, 0.015	2.42, 0.014	2.48, 0.017	2.43, 0.016	2.50, 0.018
Land ownership system	−1.24, 0.129	−1.26, 0.132	−1.22, 0.126	−1.25, 0.131	−1.23, 0.127
Policy support	0.78, 0.455	0.80, 0.460	0.76, 0.448	0.79, 0.457	0.77, 0.452
Model overall significance	4.45, 0.002	4.52, 0.001	4.40, 0.003	4.48, 0.002	4.43, 0.002

As can be seen from Table 7, for the soil fertility factor, in central Yunnan, its t-value is 2.10 and the significance level is 0.035, indicating that soil fertility has a significant positive impact on non-grain conversion of cultivated land; while in other regions, except for southeastern Yunnan, soil fertility also has a significant positive impact on non-grain conversion of cultivated land. Secondly, the factor of main crop types has significant t-values in all regions, indicating that the planting structure of different crop types has a significant impact on non-grain conversion of cultivated land. In central Yunnan, the t-value of GDP level on non-grain conversion of cultivated land is 2.75, and the significance level is 0.018, showing that the economic development level is positively correlated with non-grain conversion of cultivated land. The t-values in other regions also show similar trends, demonstrating the universality of economic factors’ influence on non-grain conversion of cultivated land in different regions. However, for factors such as slope, annual average precipitation, annual average temperature, agricultural subsidy policy, urbanization rate, and ecological environment health level, the t-values in various regions do not show significance, indicating that these factors may have different or uncertain impacts on non-grain conversion of cultivated land. From the calculation results of t-values and p-values, the soil fertility, main crop types, and GDP level in central Yunnan have a significant impact on non-grain conversion of cultivated land, while the southeastern Yunnan region is mainly affected by main crop types, farmer livelihood conditions, and technological level. This indicates that the non-grain conversion process in different regions is driven by different factors, reflecting the differences in land use patterns and economic development levels between regions. Secondly, the high value of overall model significance indicates that the established model has a good fitting effect in explaining the influencing factors of non-grain conversion of cultivated land.

A comprehensive analysis shows that the non-grain conversion of cultivated land in different regions of Yunnan Province is affected by a combination of multiple factors, among which factors such as soil fertility, main crop types, and economic level have a more significant impact on non-grain conversion of cultivated land, while the influence of other factors is relatively weak or uncertain. The emergence of these differences may be due to differences in natural environment, economic development level, policy support, and other factors in various regions. An in-depth analysis of the causes of these regional differences can reveal a close relationship with the natural environment and economic conditions. For example, the terrain in central Yunnan is relatively flat with higher soil fertility, while southeastern Yunnan has more mountains and limited land use, leading to differences in the selection of main crop types and the degree of dependence on the economic level in different regions. In addition, policy factors may also have an impact on land use in different regions, such as different implementations of agricultural subsidy policies and land ownership systems, may affect farmers’ land use methods to a certain extent. Therefore, when formulating regional cultivated land management and protection strategies, it is necessary to consider the characteristics and influencing factors of different regions in a refined and differentiated manner in order to achieve more precise management and sustainable utilization of non-grain conversion of cultivated land.

3.4. Future Prediction Analysis

The prediction of future non-grain conversion of cultivated land trends is the cornerstone of strategic planning. By predicting and analyzing key indicators, it is possible to foresee the possible trends of non-grain conversion and provide forward-looking scientific support for agricultural decision-making. This process is carried out in three stages: first, modeling and analysis of historical changes in key indicators of non-grain conversion from 2001 to 2021; then, using relevant indicators from 2022 to 2024 to verify the model constructed in this paper in reality; finally, predicting the results of non-grain conversion from 2025 to 2035 using models that have been verified to meet accuracy requirements.

The prediction and analysis results of key indicators of non-grain conversion from 2025 to 2035 are shown in Figure 7 and Table 25.

Figure 7. Prediction results with uncertainty bands.

Table 25. The 2025–2035 projections with uncertainty.

Year	NGCI Prediction	95% CI	Karst Areas	Non-Karst
2025	20.08	[18.34, 21.82]	17.23	22.45
2030	14.67	[12.11, 17.23]	11.34	17.89
2035	9.31	[6.23, 12.39]	5.67	12.78

1. Prediction Model Validation (2022–2024):

(1) Mean Absolute Percentage Error: 7.49%;

(2) Root Mean Square Error: 1.23 NGCI units;

(3) Directional accuracy: 94.3%;

(4) 95% Prediction intervals contain observed values: 92.1% (close to nominal).

2. The 2025–2035 Projections with Uncertainty

3. Probability Analysis:

(1) P(NGCI < 10 by 2035) = 0.67 overall;

(2) P(NGCI < 10 by 2035 | karst) = 0.89;

(3) P(NGCI < 10 by 2035 | non-karst) = 0.41.

3.5. Scenario Analysis with Policy Implications

In order to verify the applicability of the research design method, this study simulates three scenarios for prediction accuracy analysis: (1) Climate change scenario: simulate a 10% increase in precipitation in the next 11 years to analyze the potential promoting effect of sufficient water resources on non-grain conversion of cultivated land. (2) Policy intervention scenario: assume strengthening of cultivated land protection and agricultural policy intensity (such as increasing subsidies and strengthening regulation) to assess the impact of policy interventions on cultivated land use patterns. (3) Economic market scenario: consider the impact of market demand and price fluctuations on farmers’ crop planting choices and non-grain conversion trends of cultivated land. The parameter settings of each scenario are detailed in Table 26.

Table 26. Scenario impacts on karst versus non-karst areas.

Scenario	Parameter Change	Karst Impact	Non-Karst Impact	Policy Effectiveness
Climate Adaptation	+20% precipitation	NGCI + 2.3	NGCI + 0.8	Limited in karst due to drainage
Subsidy Enhancement	+10% grain subsidies	NGCI + 4.1	NGCI + 3.2	Effective if targeted to karst
Market Stabilization	±3% price volatility	NGCI ± 1.7	NGCI ± 2.1	Similar effectiveness
Technology Package	Precision agriculture	NGCI + 1.2	NGCI + 3.8	Barriers in fragmented karst
Ecological Restoration	20% land retirement	NGCI − 8.3	NGCI − 3.1	Major impact in degraded karst

Monte Carlo Simulation Results (10,000 iterations):

(1) Combined intervention success probability: 73% (NGCI > 15 by 2035).

(2) Required investment: 2.3 billion CNY for karst-specific programs.

(3) Cost-effectiveness: 1340 CNY per hectare protected in karst areas.

According to the above calculation, the prediction accuracy results of non-grain cultivated land in 2025–2035 under each scenario are shown in Figure 8.

Figure 8. Prediction accuracy results of non-grain conversion of cultivated land from 2025 to 2035 under three scenarios.

From the analysis of the three scenarios in Figure 8, it can be seen that the prediction accuracy of the model under different scenarios differs little from the normal scenario, with errors within ±0.5%, showing strong robustness and applicability: (1) In terms of the climate change scenario, the prediction accuracy before 2030 is slightly lower, mainly affected by the initial uncertainty of increased precipitation and changes in drought frequency. However, after 2031, as climate stability increases, the model’s prediction accuracy gradually surpasses the normal scenario, indicating that the model can dynamically adapt to the impact of long-term climate change. (2) In terms of the policy intervention scenario, the enhanced policy intensity significantly improved the cultivated land use pattern, with prediction accuracy rising to above 98% from 2028 onwards. Agricultural subsidies and land protection measures directly promote food crop planting, effectively curbing the non-grain conversion trend, verifying the positive effect of policy interventions on model predictions. (3) In terms of the market volatility scenario, although market demand and price fluctuations increase complexity, the model prediction error remains between 0.2 and 0.5%, showing good generalization ability, fully capturing the impact of market factors on cultivated land use.

Based on the above analysis, the following policy recommendations can be proposed to decision-makers: (1) Response to climate change: Strengthen water resource management, promote water-saving irrigation technology, and develop adaptable food crop varieties. (2) Optimization of policy interventions: Increase agricultural subsidy intensity, strictly implement cultivated land use control, promote land contracting system reform, and improve cultivated land use efficiency. (3) Regulation of market fluctuations: Improve agricultural product price support policies, strengthen supervision of grain circulation, reduce market risks, and promote supply-demand balance. The comprehensive implementation of the above measures will help curb non-grain conversion of cultivated land and ensure sustainable agricultural development and food security.

4. Discussions

4.1. Limitations and Future Research Priorities

While this study provides crucial insights into karst agricultural dynamics, several limitations warrant acknowledgment and define priorities for future research:

(1) Data and Methodological Constraints

Current 30 m resolution satellite imagery may miss productive soil pockets in small karst depressions (<900 m²), potentially underestimating localized agricultural potential. The 21-year study period (2001–2021) may not fully capture long-term karst-agriculture co-evolution cycles that operate on centennial timescales. Linear weight combinations in the NGCI calculation (α = 0.35, β = 0.25, γ = 0.20, δ = 0.20) may oversimplify nonlinear threshold responses characteristic of karst systems. Limited ground truth data in remote karst areas (n = 120 validation points) constrains model validation, particularly for extreme conversion scenarios.

(2) Geographic and Contextual Scope

This study focuses on Yunnan Province’s specific karst geology and agricultural context. Generalization to other karst regions requires consideration of differences in carbonate rock composition, dissolution rates, climate regimes, and agricultural traditions. The influence of ethnic minority agricultural practices and cross-border trade, while captured through proxy variables, deserves direct investigation.

4.2. Future Research Directions

Priority research areas to advance understanding of karst agricultural systems include the following:

(1) Technological Advances:

Development of UAV-based hyperspectral imaging for karst micro-topography mapping at sub-meter resolution; Application of deep learning architectures (particularly graph neural networks) to capture nonlinear, multi-scale karst-agriculture interactions; Integration of LiDAR technology for precise soil depth measurement in complex karst terrain.

(2) Systemic Understanding:

Development of coupled human-natural system models that integrate karst hydrology, soil formation processes, and agricultural decision-making frameworks; Expansion to comparative analysis across global karst agricultural regions (Mediterranean karst, tropical karst of Southeast Asia, and temperate karst of North America) to identify universal patterns and region-specific adaptations; Investigation of traditional ecological knowledge in karst agriculture management among indigenous communities.

(3) Policy and Management Applications:

Design of payment for ecosystem services schemes specific to karst agricultural landscapes; Development of early warning systems for agricultural transformation based on real-time monitoring of NGCI components; Creation of decision support tools for farmers operating in karst environments, incorporating geological constraints into crop selection and land management recommendations.

This research demonstrates that understanding non-grain conversion in karst landscapes requires moving beyond conventional agricultural models to incorporate geological constraints as primary drivers. As global food security increasingly depends on marginal lands, with karst regions covering 15% of Earth’s land surface and supporting over 1 billion people, our framework provides essential tools for sustainable agricultural development in these challenging yet critical environments. The methodologies and insights developed here offer a foundation for managing agricultural transformation in geologically constrained landscapes worldwide, contributing to both local food security and global sustainability goals.

5. Conclusions

This study provides the first comprehensive analysis of non-grain conversion dynamics in karst landscapes, revealing fundamentally different agricultural transformation patterns compared to conventional agricultural regions. Through integrated analysis of Yunnan Province’s cultivated land changes from 2001 to 2021, we demonstrate that karst terrain acts as a primary driver rather than a secondary modifier of agricultural land use decisions, with profound implications for food security in geologically constrained regions.

Our research advances the field through four major contributions that provide both theoretical insights and practical applications:

(1) Methodological Innovation: The Dynamic Spatial-Temporal Clustering Model (DSTCM)

We developed an innovative analytical framework that integrates karst-specific variables into agricultural land change analysis, achieving 18% higher prediction accuracy (92.51% ± 2.3%) compared to standard models. The incorporation of karst intensity layers (k-dimension) into traditional spatiotemporal analysis frameworks enables accurate capture of threshold effects unique to karst agriculture. This methodological advancement provides a replicable framework for analyzing agricultural transformation in other karst regions globally.

(2) Quantitative Evidence of Differential Karst Impact

The 54% intensification in non-grain conversion (NGCI: 45.91 → 21.05) occurred disproportionately in karst areas, which experienced 23% faster conversion rates despite covering only 28% of the province. Spatial distribution mapping reveals that 73% of high-intensity conversion clusters occur in areas with >30% karst coverage. This differential provides the first quantitative evidence that geological constraints fundamentally reshape agricultural transformation pathways, with karst areas showing stronger spatial autocorrelation (Moran’s I = 0.73) compared to non-karst regions (Moran’s I = 0.41).

(3) Critical Threshold Identification for Agricultural Sustainability

We identified NGCI = 10 as a critical threshold below which agricultural recovery becomes extremely difficult in karst areas due to irreversible soil degradation and rock desertification. With current trajectories projecting karst areas to reach NGCI = 5.67 by 2035 (95% CI: 3.89–7.45), compared to 12.78 in non-karst areas, this finding provides an early warning system for irreversible agricultural transformation. The probability of crossing this threshold by 2035 is 89% in karst areas versus 41% in non-karst regions.

(4) Evidence-Based Policy Design Framework

Standard agricultural interventions show significantly reduced effectiveness in karst terrain, with technology adoption facing 3 × higher implementation costs (1340 CNY/hectare versus 447 CNY/hectare). Our scenario analysis demonstrates that karst-specific interventions achieve 73% success probability in maintaining NGCI above critical thresholds, compared to 41% for conventional approaches. Successful strategies must address soil conservation (1 cm of soil formation requires 10 years in karst), water harvesting systems for rapid runoff capture, shallow-rooted crop varieties adapted to limited soil depth, and virtual parcel consolidation through farmer cooperatives to overcome fragmentation.