The Spatiotemporal Evolution and Scenario Prediction of Agricultural Total Factor Productivity Under Extreme Temperature: Evidence from Jiangsu Province

Abstract

With the intensification of global climate change, frequent extreme temperature events pose increasing challenges to agricultural production. The aim of this study is to characterize the spatiotemporal evolution of county-level agricultural total factor productivity (ATFP) under extreme temperature events, reveal key driving factors and crop-specific heterogeneity, and predict potential high-risk areas, which is crucial for providing scientific basis for risk management and adaptive policy formulation in globally climate-sensitive agricultural regions. This paper selects Jiangsu Province as a typical case study, uses the DEA-Malmquist model to measure agricultural total factor productivity (ATFP), systematically analyzes the spatiotemporal dynamic evolution characteristics of ATFP at the county scale, and selects the random forest and XGBoost ensemble models with optimal accuracy through model comparison for prediction, assessing the evolution trends of ATFP under different climate scenarios. The results showed that: (1) From 1993 to 2022, the average ATFP increased from 0.7460 to 1.1063 in the province, though development showed uneven distribution across counties, exhibiting a “high in the south, low in the north” gradient pattern. (2) Mechanization, agricultural film and land inputs are the core elements driving the overall ATFP increase but there are obvious crop differences: mechanization has a more prominent role in promoting the productivity of wheat and maize, while labor inputs have a greater impact on the ATFP of rice. (3) The negative impacts of extreme climate events on agricultural production will be significantly amplified under high-emission scenarios, while moderate climate change may have a promotional effect on certain crops in some regions.

1. Introduction

Global climate change is profoundly impacting agricultural production systems, with the increasing frequency, intensity, and duration of extreme temperature events emerging as major threats to agricultural sustainability. The Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC) reports a marked increase in global extreme climate events [1], which has led to substantial fluctuations in major food crop yields worldwide, ranging from 18% to 43% [2]. As the world’s largest producer of wheat and rice, China’s agricultural system exhibits high sensitivity to climate shocks, with its climate vulnerability index significantly exceeding the global average [3]. Agricultural Total Factor Productivity (ATFP) is a comprehensive indicator of agricultural production efficiency and technological progress. ATFP characterizes the agricultural production process by jointly accounting for land, labor, capital, and intermediate inputs, thereby reflecting both resource allocation efficiency and the adaptive capacity of agricultural systems [4]. Under the intensifying impacts of extreme heat and other climate events, existing regional agricultural development pathways and climate adaptation policies are increasingly challenged. Therefore, the paper focuses on the spatiotemporal evolution trend of ATFP under extreme temperature scenarios, integrating meteorological data, remote sensing information, and machine learning algorithms to deeply analyse climate risk differences within different crop regions.

Existing studies have extensively examined the impacts of climate change on agricultural production using econometric analyses based on historical data and controlled environmental experiments. Commonly applied methods include multiple linear regression [5,6], meteorological yield decomposition [7,8], difference-in-difference approaches [9], and grey relational analysis [10]. Empirical evidence suggests that rising temperatures generally exert negative impacts on crop yields [11], although moderate warming (e.g., 1.5 °C) may temporarily enhance yields for certain crops, while higher temperature increases can lead to significant yield reductions [12]. Temperature and light conditions have also been shown to directly affect rice growth and spatial yield patterns [13,14]. Through controlled environmental experiments [15,16] and crop models [17,18,19], studies have shown that climate change has a significant impact on crop yields and planting suitability [20,21,22], including yield increases in winter wheat and summer maize under warming conditions and associated shifts in planting areas [23,24,25].

Compared to the long-term trends of climate change, extreme weather events (such as high temperatures, droughts, and floods) typically have a more intense and direct impact on agriculture. Zou et al. (2024) empirically found that extreme rainfall events result in a consistent negative marginal effect on ATFP [26]. Cunningham and Zhu (2022) predicted through simulation studies that heavy rainfall events could reduce winter wheat growth and lead to excessive soil moisture, thereby inhibiting crop growth [27]. Suresh et al. (2021) and Chen et al. (2018) found that crop yields significantly decreased under the backdrop of frequent extreme weather events [28,29]. Liu et al. (2021) discovered that by exacerbating heat stress, drought, floods, and pest outbreaks, crop yields were reduced [30].

As extreme climate events intensify and global food demand continues to rise, improving agricultural production efficiency has become increasingly important. ATFP, as a core measure of agricultural efficiency, captures the interaction between multiple production inputs and outputs [31,32,33]. Previous studies have applied the Malmquist index [34], Cobb-Douglas production functions [35], and transcendental logarithmic production functions [36] to analyze relationships between climatic factors and agricultural productivity across regions and crops. Owing to substantial spatial heterogeneity in farmland conditions, high-resolution analyses are particularly important in China. For example, combined changes in temperature and precipitation have been shown to promote ATFP growth in Northwest, Northeast, and Southern China, while exerting negative effects in the Huai River region [37]. Increased sunshine duration and precipitation have also been found to positively influence agricultural productivity in western China [38]. Focusing further on the regional scale, studies on Jiangsu Province—a major grain-producing area in China—have identified changes in precipitation and temperature as key climatic drivers of fluctuations in grain yield [39,40]. For example, in southern Jiangsu, spatiotemporal variations in temperature and rainfall have been shown to significantly affect the output of major grain crops [41]. In the saline-alkali agricultural areas of northern Jiangsu, climate change poses a particular threat to agricultural production by altering hydrothermal conditions [42].

Despite extensive research on the agricultural impacts of climate change and extreme weather events, current studies still have three shortcomings. First, in terms of spatial scale, the majority of studies have conducted efficiency analyses based on municipal-level or broader perspectives, given the accessibility of indicator data. The county-level is a crucial link in the implementation of the rural revitalization strategy, and serves as an important implementation unit for achieving agricultural and rural modernization. It is evident that further in-depth research is required on the subject of refined agricultural production efficiency from a county-level perspective in typical regions; second, in terms of coverage of research subjects, existing literature is often limited to response analyses of single crops, failing to systematically compare the differential responses of different crop types to extreme weather; finally, in terms of time dimension, current research primarily focuses on retrospective analyses of historical climate impacts, with insufficient systematic assessments of future trends in extreme weather events and their impacts on agriculture.

To address these gaps, this study selects Jiangsu Province, one of China’s major grain-producing regions, as a representative case. Through county-level data, this paper analyzes the spatiotemporal evolution of ATFP under extreme temperature events and future agricultural risk patterns under multiple climate scenarios. The main contributions of this study are threefold: (1) A high-resolution spatiotemporal analysis at the county scale is conducted to uncover the regional disparities and dynamic changes in agricultural production efficiency under extreme temperature events, enabling precise identification of agricultural production risk. (2) Wheat, rice, and maize are treated as independent research subjects to examine the differential responses of these major staple crops to extreme temperature events. The sensitivity of their ATFP is systematically quantified and compared, and the key factors influencing crop production are identified. (3) This study integrates multi-scenario climate projection models (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5) with machine learning algorithms to predict the potential impacts of extreme temperature events on agricultural production efficiency, delineating future high-risk areas and providing a scientific foundation and decision-making support for the formulation of more precise climate adaptation policies and the optimization of agricultural production strategies.

2. Research Area, Research Methods, and Data Sources

2.1. Research Area

Jiangsu Province is located between 30°45′ N–35°20′ N and 116°18′ E–121°57′ E, situated in the core coastal region of eastern China. As the main province of the Yangtze River Delta, it possesses both land and sea geographical characteristics (as shown in Figure 1). Its geographical layout exhibits the typical characteristics of “one river, one canal, and two belts”. The main channel of the Yangtze River spans 433 km from east to west, the Grand Canal runs 718 km from north to south, and a 957-kilometre coastline connects to a maritime space of 37,500 square kilometers. Within its 107,200 square kilometers of land area, a unique climate transition zone exists, transitioning from a warm temperate to a subtropical climate (with an annual average temperature of 13–16 °C and annual precipitation of 800–1200 mm). The geographical feature of a north-south span of approximately 500 km forms a typical climate transition zone, with climate risks exhibiting significant spatial heterogeneity. Over the past decade, the drought index in northern Jiangsu Province has increased by 23%, while the frequency of extreme precipitation in southern Jiangsu has risen by 31%, reflecting a pronounced “north drought, south flood” trend. Currently, 43% of the province’s basic farmland is located in climate risk overlap zones (Jiangsu Province Climate Change Assessment Report, 2017). As a typical example of China’s intensive agricultural development, the province has 13 prefecture-level cities and 95 counties (cities and districts), supporting 6.05% of the country’s agricultural population on 3.18% of its arable land, with a land reclamation index of 62.3%. The fertilizer input per unit of arable land is 1.7 times the national average, producing 5.5% of the country’s grain, 7.6% of vegetables, and 3.4% of meat products, achieving self-sufficiency in staple foods and surplus exports despite being the province with the highest population density. In 2024, Jiangsu Province’s total grain production reached approximately 38.1 million tonnes, remaining stable above 35 million tonnes for 11 consecutive years, ranking among the top in the country. Jiangsu Province is located in the transitional climate zone between the warm temperate and North Subtropical zones and has a highly intensive agricultural production model. Using it as a typical case study in this paper can provide valuable insights for improving agricultural production efficiency in other climate-sensitive regions, as well as critical data support for broader regional climate adaptation research and policy formulation.

2.2. Research Methods

This paper takes agricultural total factor productivity (ATFP) as the research subject. First, ATFP is estimated through the DEA-Malmquist model, and Z-score normalization is applied to agricultural input variables and ATFP to eliminate the influence of scale differences. Second, ridge regression is employed to identify the main factors that significantly affect ATFP. In addition, Random Forest, Support Vector Machine, and Neural Network methods are used to analyze the sensitivity of ATFP to extreme climate events, thereby extracting key feature indicators with substantial impacts on ATFP and selecting the model with the best predictive performance. Finally, based on these key features, an ensemble learning framework integrating random forest, Bayesian regression, and XGBoost is constructed, from which the model with the highest accuracy is selected as the final predictive model, so as to more comprehensively reveal the impacts of extreme climate change on ATFP. In this paper, the ATFP calculations, machine learning computations, and mapping with spatial downscaling were performed using DEA FARONTIER 4.1, Python 3.12, and ArcMap 10.8, respectively. The overall research framework is illustrated in Figure 2.

2.2.1. DEA-Malmquist Model

DEA is a non-parametric frontier method for evaluating the relative efficiency of decision-making units with multiple inputs and multiple outputs, and it is not affected by differences in measurement units. In DEA, two standard specifications are the CCR and BCC models. The CCR model assumes constant returns to scale (CRS) and measures overall technical efficiency, where scale effects and technical inefficiency are mixed. The BCC model assumes variable returns to scale (VRS) and estimates pure technical efficiency by separating scale effects. Because counties differ substantially in production scale, using CCR alone may confound scale differences with technical efficiency and bias inter-county comparisons; therefore, we adopt DEA-BCC for period-by-period efficiency evaluation. For productivity dynamics, the Malmquist index is then computed from DEA distance functions following Coelli and Rao (2005) [43,44]. This method constructs a production frontier of production efficiency through the linear programming method of data envelopment analysis (DEA). Specifically, given data from N research areas within a specific period, the output-oriented DEA model is used to solve the linear programming problem for the i-th research area, as shown below:

$$\left\{\begin{array}{l}{\max }_{\varphi ,\lambda }\varphi \\ st-\varphi y_i+Y\lambda \ge 0\\ x_i-X\lambda \ge 0\\ \lambda \ge 0\end{array}\right.$$

(1)

where $y_i$ is the M × 1 vector of outputs for the i-th region, $x_i$ is the K × 1 vector of inputs for the ith region, Y is the N × M matrix of outputs for all N regions, X is the N × K matrix of inputs for all N regions, $\lambda$ is an N × 1 vector, and $\varnothing$ is a scalar.

Assuming that study region i is in each period $t=1,2,\cdots T$, $x_t$ is the total input variable in period t, $y_t$ is the total output variable, and that the technology satisfies the equations outlined in Coelli’s technique, the s-th and t-th period Malmquist productivity indices for study region i are ${\displaystyle \frac{d_i^s(y_t,x_t)}{d_i^s(y_s,x_s)}}$ and ${\displaystyle \frac{d_i^t(y_t,x_t)}{d_i^s(y_s,x_s)}}$. The Malmquist index is used to measure the ATFP changes between two research regions, calculated as the ratio of the distance of each research region relative to the general technology. The formula is as follows:

$$m_i(y_s,x_s,y_t,x_t)=[{\displaystyle \frac{d_i^s(y_t,x_t)}{d_i^s(y_s,x_s)}}\times {\displaystyle \frac{d_i^t(y_t,x_t)}{d_i^s(y_s,x_s)}}{]}^{1/2}$$

(2)

2.2.2. Machine Learning Algorithms

First, this paper uses the Z-score normalization method to standardize agricultural input and ATFP data. This involves subtracting the mean of the variable from each value and dividing by the standard deviation, transforming the data into a form with zero mean and unit variance, thereby eliminating the impact of different scales. To avoid errors caused by multicollinearity, ridge regression is used to identify the primary influencing factors [45]. Ridge regression is a regularization method for multiple linear regression that effectively addresses multicollinearity issues [46]. In multiple linear regression, when input features are highly correlated, traditional least squares regression can lead to model instability. Ridge regression solves this problem by adding a regularization term (L2 norm) to the loss function, reducing the risk of overfitting. The specific regression equation is as follows:

$$\widehat{\beta }=\underset{\beta }{argmin}(\sum _{i=1}^n(y_i-X_i\beta {)}^2+\lambda \sum _{j=1}^p{\beta }_j^2)$$

(3)

Among them, $y_i$ is the response variable, $x_i$ is the input feature, $\beta$ is the regression coefficient, and $\lambda$ is the regularization parameter, which controls the size of the regression coefficient.

Secondly, this paper uses Random Forest, Support Vector Machine, and Neural Network to analyze the sensitivity of various extreme climate indicators. Among these, Random Forest is an ensemble learning method that enhances model accuracy by constructing multiple decision trees and combining their predictions [47]. Each decision tree randomly selects a subset from the training data using Bootstrap sampling and considers only a portion of the features during each split. Ultimately, it outputs regression predictions by averaging the predictions from multiple trees or classifies results through voting. For regression problems, the prediction of Random Forest can be expressed as:

$${\widehat{y}}_{RF}(x)={\displaystyle \frac{1}{T}}\sum _{t=1}^T{f}_t(x)$$

(4)

Among them, T represents the number of trees, $f_t\left(x\right)$ is the prediction result of the t-th tree.

Support Vector Machine (Support Vector Machine, SVM) is a supervised learning method used for classification and regression. The core idea is to find a hyperplane that separates samples of different classes [48]. In regression problems, SVM constructs an optimal regression function so that most data points fall within the tolerance band while minimizing prediction errors. For a linear regression problem, the optimization objective of SVM is to minimize the following objective function:

$$\mathit{min}{\displaystyle \frac{1}{2}}{\Vert w\Vert }^2+C\sum _{i=1}^n{\xi }_i$$

(5)

Neural networks (Neural Network, NN) are nonlinear models that mimic biological neural systems, performing complex data mapping through multi-layered neuronal networks [49]. Each neuron processes input signals by weighted summation and outputs results through activation functions (such as ReLU, Sigmoid, or Tanh), widely applied in classification and regression tasks. The output of a neural network can be expressed by the following formula:

$$y=\sigma \left(\sum _{i=1}^n{w}_i{x}_i+b\right)$$

(6)

where $x_i$ is the input, $w_i$ is the weight, $b$ is the bias term, and $\sigma$ is the activation function.

Finally, this paper integrates three machine learning methods—Random Forest, Bayesian Regression, and XGBoost—to construct an ensemble model for ATFP prediction, which comprehensively captures the impact of extreme temperature events on agricultural productivity. The Bayesian approach often combines Bayesian optimization for hyperparameter tuning, with the optimization objective typically being to minimize a loss function [50]. During the Bayesian optimization process, a surrogate model (such as Gaussian process regression) is established to estimate the shape of the loss function and select the optimal parameter combination. The formula usually involves maximizing an acquisition function, such as Expected Improvement (EI):

Among them, where $x$ is the hyperparameter combination to be evaluated, $f\left(x*\right)$ is the current best predicted value, and $f\left(x\right)$ is the predicted value of the objective function.

XGBoost (Extreme Gradient Boosting) is an efficient Gradient Boosting Tree (GBDT) algorithm based on the boosting concept. It iteratively trains a series of weak learners (usually decision trees) to gradually improve the model’s predictive performance [51]. In each iteration, XGBoost optimizes the model parameters by minimizing the loss function. The objective function of XGBoost consists of two parts:

$$EI\left(x\right)=E\left[\mathit{max}\right(0,f\left(x\right)-f\left(x*\right)\left)\right]$$

(7)

where $T$ is the number of leaf nodes of the tree, $\gamma$ is the complexity penalty term of the tree, and $\lambda$ is the L2 regularization parameter of the weight of leaf nodes, ${\omega }_j$ is the weight of leaf nodes. By minimizing this objective function, XGBoost seeks a set of optimal tree structure and leaf node weights, so that the prediction error of the model on the training set is minimized and has strong generalization ability.

$$\mathcal{l}(\theta )=\sum _{i=1}^n\mathcal{l}(y_i,{\widehat{y}}_i)+\sum _{k=1}^K\Omega (f_k)$$

(8)

To more comprehensively capture the impact of extreme temperature events on ATFP, this paper combines three machine learning methods: random forests, Bayesian regression, and XGBoost, to build an ensemble model. Random forests enhance prediction accuracy by integrating multiple decision trees, offering strong robustness but being complex and difficult to interpret when dealing with high-dimensional data. Bayesian regression introduces prior information to handle uncertainty and noise, providing probabilistic explanations, but it is sensitive to prior selection and has high computational costs in high-dimensional data. XGBoost optimizes models using gradient boosting algorithms, excelling in handling nonlinear relationships and missing data, but it is prone to overfitting. The ensemble model improves predictive performance by combining the predictions of multiple models, thus overcoming the limitations of a single model and enhancing prediction stability and robustness. The optimal combination method is selected based on the performance of each model for final prediction, using a weighted average approach to comprehensively consider the predictions of each model, thereby maximizing the advantages of each model. This method captures complex patterns more comprehensively, especially in the face of nonlinear relationships or high-dimensional data, better adapting to various scenarios and improving the accuracy of ATFP predictions under future extreme climate changes.

2.3. Data Sources

To calculate agricultural total factor productivity (ATFP) in Jiangsu Province, we compiled annual county-level agricultural input-output data for 1992–2022. The data were collected primarily from the Jiangsu Agricultural Statistical Yearbook and the statistical yearbooks of the corresponding prefecture-level cities. Based on data availability and established ATFP measurement practice, this paper selected seven input indicators and one output indicator [4], and all variables are reported with their measurement units and data sources. Due to the lack of input and output indicators of each crop, the ratio of the sown area of the crop to the sown area of the crop was selected to multiply the inputs and outputs of each county as a whole to indicate the inputs and outputs of each crop [52]. The specific indicators are shown below (

Table 1. Input-Output Indicators of ATFP in Jiangsu Province.

Indicators	Definition		Remarks
Input indicators	Land investment	Total sown area of crops (thousands of hectares)	The area of all cultivated or uncultivated land sown or transplanted in that year
	Fertilizer input	Agricultural fertilizer amount (ten thousand tons)	Contains nitrogen fertilizer, phosphate fertilizer, potassium fertilizer and compound fertilizer
	Mechanical inputs	Total power of agricultural machinery (10,000 kilowatts)	It is mainly used for the power of various power machines in agriculture, forestry, animal husbandry and fishery
	Human input	Number of people employed in agriculture (10,000 persons)	Number of employees in agriculture, forestry, animal husbandry and fishery x (total agricultural output value/total output value of agriculture, forestry, animal husbandry and fishery)
	Inflow of irrigation	Effective irrigated area (thousands of hectares)	The area of arable land that can be irrigated with water and equipment support
	Agricultural chemical inputs	Pesticide use (tons)	Total amount of chemicals used for agricultural pest control and plant growth control
	Plastic film input	Usage of agricultural plastic film (tons)	Used to retain soil moisture and heat
Output indicators	agricultural output	total value of farm output (CNY 10,000)	The total amount of agricultural products and various support services for agricultural production

):

In order to further analyze the impact of extreme temperature on agricultural total factor productivity in Jiangsu Province, this paper collects daily temperature data between 1993 and 2022 from the CNO5.1 climate data released by the National Climate Center of China. For the regional characteristics of Jiangsu Province, the inverse distance weight interpolation method (IDW) was used to spatially interpolate the climate data on a national scale, resulting in county-level data applicable to this paper. To ensure the scientificity and reliability of the climate data, this paper refers to the Climate Change Detection Indicator (ETCCDI) index developed by the expert organization of the World Meteorological Organization (WMO) [53], from which climate indicators related to the extreme value, intensity and frequency are selected to effectively capture the phenomenon of extreme temperature event changes in Jiangsu Province, as shown in

Table 2. Extreme values, intensity and frequency indices of extreme temperature.

Type	Index
Extreme temperature index	TXx (maximum daily maximum temperature of the year)
	TNn (minimum daily minimum temperature of the year)
Extreme temperature intensity index	TN10p (cold night: percentage of days with minimum temperature < 10% percentile)
	TX90p (warm day: percentage of days with maximum temperature > 90% percentile)
Extreme temperature frequency index	FD (Frost days: the number of days with minimum temperature < 0 °C in a year)
	SU (summer days: the number of days with daily maximum temperature > 35 °C within a year)
	ID (Freezing days: the number of days with the highest daily temperature < 0 °C in a year)
	TR (hot night days: the number of days with minimum temperature > 20 °C in a year)

.

In addition, in order to analyze the potential impacts of future extreme temperature changes on agriculture in Jiangsu Province, this paper obtains data on future climate scenarios for the period 2025–2035 based on the Coupled Model Intercomparison Program for Phase 6 (CMIP6) supported by the International Panel on Climate Change (IPCC), and selects five representative global climate models (ACCESS- ESM1-5, BCC-CSM2-MR, IPSL-CM6A-LR, MRI-ESM2-0, and IPSL-CM6A-LR) covering four different typical Shared Socioeconomic Pathways (SSP) scenarios, which are SSP1-2.6 (low forcing scenario, with radiative forcing reaching 2.6 W/m² in 2100), SSP2-4.5 (medium forcing scenario, with radiative forcing reaching 4.5 W/m² in 2100), SSP-3.70 (medium forcing scenario, with radiative forcing reaching 7.0 W/m² in 2100) and SSP5-8.5 (high forcing scenario, with radiative forcing reaching 8.5 W/m² in 2100). In order to accurately apply the national-scale climate model data to the regional level of Jiangsu Province, this paper adopts Quantile Mapping (QM), a statistical downscaling method, to correct the model output, and combines it with Inverse Distance Weighting (IDW) to effectively interpolate the national climate data to the Jiangsu Province scale. Interpolation to the Jiangsu Province range is processed [54,55], which effectively improves the accuracy of the climate model output, ensures the reliability of the data, and provides the basis for the subsequent impact analysis.

3. Spatiotemporal Trend and Main Influencing Factors

3.1. Temporal and Spatial Variation Trend of ATFP

The spatiotemporal variation trend of the average value and its growth rate of ATFP in Jiangsu Province is shown in Figure 3. From 1993 to 2022, the average value of the ATFP increases from 0.7460 to 1.1063. However, development across different counties is uneven, showing a gradient distribution characterized by “high in the south and low in the north”. In the early period (1993–2002), ATFP in Jiangsu Province grew rapidly but was unevenly distributed, mainly concentrated in the economically more developed southern Jiangsu region and around cities with better infrastructure. Most counties, primarily those using traditional farming methods, still faced significant development bottlenecks [56]. In the mid-to-late period (2003–2022), the growth rate of ATFP in Jiangsu Province stabilized. With the application of intelligent technologies and the popularization of modern agricultural facilities, productivity gaps between regions gradually narrowed.

According to different geographical regions, Jiangsu Province is divided into three major areas: Southern, Central, and Northern. Each area has distinct characteristics in agricultural production and development levels. In Southern area, such as Pukou, Yangzhong, Liyang, and Danyang, the ATFP is generally high. The widespread application of modern agricultural technology and precision agriculture has kept these counties and cities at the forefront [57]. In Central area, like Taixing, Jingjiang, Gaoyou, and Yizheng have relatively high ATFP but generally slow growth rates. Compared to Southern Jiangsu, these counties and cities still lag behind in the popularization of modern agricultural technology and infrastructure construction [58]. In Northern area, the ATFP is lower, with significant variations in growth rates. For example, Binhai, Lianshui, Pei County, and Feng County have gradually improved their agricultural production efficiency through increased introduction of modern agricultural technology and infrastructure improvements. However, other regions, such as Dafeng, Yandu, Shuyang, Pizhou, and Sheyang, are constrained by traditional agricultural technology and resource allocation delays, resulting in negative ATFP growth rates.

From 1993 to 2022, the ATFP in Jiangsu Province for wheat, rice and maize showed an upward trajectory. The average ATFP for wheat increased from 0.9239 to 0.9804, for maize from 1.0428 to 1.0782, and for rice from 1.0004 to 1.0455. However, there were significant differences in development levels and growth rates across different crops, periods, and regions in Figure 4, Figure 5 and Figure 6. In the early stages, the ATFP of wheat was mainly concentrated in central Jiangsu and parts of southern Jiangsu, such as Jingjiang and Yangzhong. Over time, the ATFP of wheat in some counties and cities in northern, like Sihong and Suyu, also saw significant improvements. The ATFP of rice remained relatively stable overall, with higher values concentrated in the richly watered main rice-producing areas of the Li Xia River region, such as Xinghua, Hanjiang, and Ganyu. These areas have better irrigation conditions, which has improved production efficiency. In contrast, areas like Rugao, Hongze, and Qidong experienced slower growth rates, possibly due to uneven water distribution, fragmented agricultural production, and a reduction in labor force, which limited the improvement in production efficiency. The ATFP values of maize were higher mainly in Jiangdu, Hongze, and Taixing, where the promotion of agricultural technology and optimization of planting patterns effectively increased output efficiency. By comparison, the ATFP values in Donghai, Guannan, and Danyang remained low, indicating that these regions have not yet formed large-scale maize cultivation systems, and the improvement in production efficiency is relatively lagging.

3.2. Main Influencing Factors of ATFP

Through ridge regression analysis, a multiple linear regression equation was constructed between ATFP and seven agricultural input variables. The analysis results reveal significant differences in the impact of different agricultural production factors on ATFP (as shown in Figure 7). Mechanical input is the primary factor affecting overall agricultural total factor productivity. As mechanization levels increase, not only does labor productivity improve, but production costs also decrease, making mechanical input a key driver for enhancing ATFP. The second is the input of agricultural films. Large-scale greenhouse cultivation and crop protection needs make agricultural films highly effective in increasing crop yields and improving growing environments. Additionally, land input has a more pronounced promoting effect on ATFP in regions such as Fengxian, Hanjiang, Gaoyou, and Jinhu. Jiangsu Province has relatively abundant land resources, especially in the northern and central regions, where land is a fundamental element driving agricultural production growth. In contrast, the impact of labor, irrigation, pesticide, and fertilizer input is relatively weaker. With the gradual reduction in labor and the increasing degree of agricultural mechanization, the influence of labor input on ATFP is gradually diminishing. Although the use of fertilizers and pesticides has to some extent promoted yield increases, overuse has led to near saturation, and their marginal contribution to total factor productivity is declining [53].

For the three crops, the key factors affecting the ATFP in agriculture vary (as shown in Figure 8). Mechanization input is a critical factor for improving wheat ATFP, which is closely related to the high reliance on mechanization in sowing, field management, and harvesting processes. Rice production is more labor-intensive, with areas like Jiangyin, Yixing, and Xishan showing a significant positive effect of labor input on ATFP. This is mainly due to the high labor intensity in transplanting, field management, and manual harvesting during rice cultivation [59]. For maize, the main factors influencing ATFP include labor input, mechanization input, and agricultural film input, indicating that effective labor and technology inputs can not only improve the crop growth environment but also create more favorable conditions for maize planting [60]. Regional differences also play a significant role in the impact on agricultural productivity in Jiangsu Province, where southern regions (such as Jiangyin, Yixing, and Xishan) focus on precision rice cultivation, making labor input more impactful on ATFP; while northern regions (such as Fengxian and Peixian) benefit from abundant arable land, leading to widespread large-scale planting, with land input becoming the key factor. Additionally, in areas with higher mechanization levels (such as major wheat and maize production zones), the role of mechanization input in boosting agricultural productivity is particularly evident, and the process of agricultural modernization advances more rapidly in these regions.

4. Scenario Prediction

4.1. Sensitivity Analysis of Extreme Temperature Events

To analyze the impact of extreme temperature changes on the ATFP in Jiangsu Province from 1993 to 2022, this paper employed three machine learning models for sensitivity analysis: Random Forest, Support Vector Machine, and Neural Network. By comparing the values R² of each model, the random forest model (R² = 0.98) was ultimately selected. In selecting extreme temperature events, this paper screened six key climate indices based on historical climate data in Figure 9. Additionally, this paper adopted an ensemble learning approach, combining random forest, Bayesian regression, and XGBoost to predict ATFP to enhance the accuracy and stability of predictions. The optimal parameter combination was found by integrating three learning methods. The final results show that among all models, the ensemble model of random forest + XGBoost performed best, with a prediction accuracy R² = 0.9874, significantly higher than any single model. Based on the predictions from this ensemble model, the trends in ATFP under different climate scenarios were further analyzed, and the potential impact of extreme climate on agricultural productivity was evaluated in light of the characteristics of various crops.

4.2. Prediction of ATFP in Jiangsu Province

Based on the ATFP prediction results for different climate scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5) in Jiangsu Province, the impact of extreme temperature changes on total factor productivity in agriculture (ATFP) shows significant differences. The changes compared to the average ATFP from 1993–2022 are shown in Figure 10. Overall, the growth rate of ATFP in Jiangsu Province is relatively high under the SSP1-2.6 and SSP2-4.5 scenarios, indicating that moderate climate change can help optimize agricultural production conditions. However, under the SSP3-7.0 and SSP5-8.5 scenarios, the growth rate of ATFP shows a clear downward trend, especially under the SSP5-8.5 scenario, where the intensification of extreme weather events has a significant impact on agricultural production, leading to a slowdown in overall growth, and even negative growth in some areas. This is consistent with the findings of Chen and Gong (2021) [61], who concluded that future global warming will have considerable negative impacts on agriculture.

For different crops, the impact of extreme climate on ATFP varies significantly. Under the SSP2-4.5 scenario, wheat shows a higher ATFP growth rate (as shown in Figure 11), indicating that mild climate change is beneficial to wheat production. However, under the SSP3-7.0 scenario, due to extreme heat leading to shortened growing seasons and hindered grain filling, the ATFP of wheat in Fengxian and Peixian counties has significantly declined. These regions are more susceptible to extreme weather events, particularly those caused by high temperatures and droughts. The ATFP of rice shows relatively stable growth rates under the SSP1-2.6, SSP2-4.5, and SSP3-7.0 scenarios in Figure 12, showed that rice has strong adaptability, especially in areas with proper water resource management. However, in coastal regions such as Yancheng and Dongtai, extreme heat and drought can lead to water shortages and exacerbate salinization, affecting rice growth and development, resulting in a decline in growth rates. The ATFP of maize shows a higher growth rate under the SSP1-2.6 and SSP2-4.5 scenarios, but under the SSP3-7.0 and SSP5-8.5 scenarios, some areas exhibit a clear downward trend in Figure 13. maize is highly sensitive to temperature changes, and extreme heat can affect its pollination and grain setting. In particular, in northern (such as Sihong, Donghai, and Guanyun), drought conditions have worsened, impacting maize growth and leading to a decrease in ATFP growth rates.

5. Conclusions and Recommendation

This paper uses the DEA-Malmquist model to calculate ATFP based on agriculture and climate data from 1993 to 2022. Jiangsu Province is selected as a typical case study to analyze the spatiotemporal dynamics of ATFP across counties and identify the primary driving factors through ridge regression analysis. Additionally, the sensitivity to extreme temperature events is analyzed, and a random forest and XGBoost ensemble model is used to predict the trends in ATFP under different climate scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5). The research findings indicate:

Jiangsu Province saw a significant overall increase in ATFP from 1993 to 2022 (mean 0.7460–1.1063), but there was significant imbalance in development between counties, exhibiting a “south high, north low” gradient characteristic. This is consistent with the findings of Yang et al. (2020) [41] regarding productivity trends in the region. The average ATFP for wheat increased from 0.9239 to 0.9804, for maize from 1.0428 to 1.0782, and for rice from 1.0004 to 1.0455. Specifically, the overall improvement in ATFP across the province is primarily driven by mechanization, agricultural film, and land inputs. By crop type, mechanization is the core driver of efficiency gains in wheat and maize, while rice production relies more on labor inputs. The province’s ATFP responds significantly to extreme temperature changes, with production efficiency benefiting under moderate scenarios (SSP1-2.6, SSP2-4.5) but growth being significantly suppressed under extreme scenarios (SSP3-7.0, SSP5-8.5).

The research conclusions indicate that for the sustainable development of agriculture, it is imperative to advance the following important areas in a coordinated manner: First, promote the transformation of agricultural production from reliance on traditional high-input inputs (fertilizers, pesticides) to innovation and efficient elements (mechanization, land optimization), which is the inevitable path to achieving sustainable efficiency improvements. Second, the prevention of extreme climate risks must be elevated to the level of a national agricultural security strategy, particularly in response to the threats of extreme heat and drought. To this end, it is necessary to strengthen crop-specific and region-specific sensitivity assessments and risk zoning, and based on these, develop and promote targeted adaptive technologies and management measures to build a more resilient agricultural production system. Third, regional differences must be fully recognized. National policies and resource allocation should avoid a one-size-fits-all approach, instead strengthening regional and categorized guidance to support different climate zones and agricultural ecological regions in exploring optimal development paths based on their resource endowments, dominant crops, and climate risk patterns. The case study of Jiangsu Province not only provides scientific basis for its own high-quality agricultural development but also offers critical decision-making references and scientific support for the nationwide promotion of agricultural modernization and effective response to climate change risks. However, due to the limited availability of county-level data, crop-specific inputs and outputs were estimated using a sown-area share approach. Future studies could improve measurement accuracy by incorporating higher-resolution data and by integrating causal econometric or process-based crop models to better assess agricultural system resilience under climate change.