Short-Term Climate Prediction over China Mainland: An Attempt Using Machine Learning, Considering Natural and Anthropic Factors

Abstract

Climate extremes pose significant natural threats to socioeconomic activities. Accurate prediction of short-term climate (STC) can provide relevant departments with warnings to effectively reduce this threat. To accurately predict STC in China, this study utilizes machine learning algorithms, particularly the random forest (RF) model, to evaluate the role of both natural and anthropogenic factors. Monthly temperature and precipitation data from 160 meteorological stations spanning China, as well as natural climate factors and an economic activity index, were obtained to perform a seasonal hindcast of air temperature and precipitation observed from 1979 to 2018. Our focus was to predict the seasonal mean temperature and precipitation, specifically the summer (June, July, and August (JJA)) and winter (December, January, and February (DJF)) air temperature and precipitation anomalies using forecast factors from the preceding season. Results show that a comprehensive consideration of both natural and anthropogenic effects provides a more accurate fit to the observed climate trends compared to using only one factor. When both factors were integrated, the model scores (coefficient of determination) exceeded 0.95, close to 1.00, which is significantly higher than those of natural (0.86 for temperature, 0.85 for precipitation) or anthropogenic (0.90 for temperature and 0.50 for precipitation) factors alone. Furthermore, we also attempted to predict similar components for 2019 and 2020. The average relative error between predictions and observations was less than 10%, indicating that this integrated model’s performance exhibited a significant improvement in predicting the STC. The findings of this study underscore the importance of accounting for both natural and anthropogenic factors in predicting climate trends to inform sustainable decision-making in China.

1. Introduction

Climate change has emerged as one of the most pressing global challenges of our time [1], with a wide range of climate extremes, including heatwaves, droughts, floods, and storms, posing severe threats to human lives, ecosystems, and economies worldwide. Among the countries affected, China has already felt the impact of climate change on various sectors, including agriculture, water resources [1], public health [2], and infrastructure. In recent decades, the country has experienced an increase in climate extremes, such as flash floods [3], city storms, landslides [4], and drought events [5], which have resulted in significant human and economic losses. For instance, between 2011 and 2015, China experienced over 780 flash flood events, leading to an average death toll of 400 people per year [6,7]. Moreover, landslides, which are directly related to extreme precipitation events, have become a significant concern in China [8], given the region’s dynamic rainfall patterns and susceptibility to geological hazards. Notably, the catastrophic soil and rock landslide in the city of Chongqing [9] in southwestern China and the relatively slow landslide in Sichuan Province in 2017 [10] were attributed to extreme rainfall and an increase in soil water content due to anthropogenic activities. Further, given China’s complex topography and large population, it remains particularly vulnerable to climate extremes. Thus, accurately predicting short-term climate trends is crucial to mitigate and adapt to the impacts of climate change in the country. As such, it is crucial to develop proactive measures to mitigate the impacts of climate change, and accurate short-term climate trend prediction is critical in achieving this goal.

Short-term climate (STC) prediction involves the prediction of climate variables such as temperature and precipitation on a timescale ranging from a season to an inter-annual period [11]. This method involves hindcasting and forecasting climatological features based on atmospheric science principles and other physical bases such as large-scale atmospheric circulation patterns. Short-term climate prediction has numerous applications in various sectors [12,13], including water management, irrigation management, transport, and wildfire control. For example, it can provide early warnings and mitigation schemes for floods, droughts, heatwaves, and heavy snow conditions, among others.

In China, accurate short-term climate prediction is particularly important due to the country’s large population, diverse economy, and geographical diversity. Reliable short-term climate prediction can help decision-makers and stakeholders prepare for potential climate-related risks and implement proactive measures to adapt to changing climate conditions. For instance, accurate short-term climate prediction can help stakeholders develop effective measures to mitigate and adapt to climate-related risks and improve the resilience of social–economic systems [14]. Furthermore, short-term climate prediction can assist in resource allocation and planning [15,16], which is critical for the long-term sustainability of social–economic activities.

Climate prediction models face challenges due to the dynamic changes in precipitation and temperature. The resulting variability poses difficulties in constructing stable and reliable models for early warnings of weather or climate extremes. Therefore, the development of various models for weather and climate predictions continues.

In the past, the STC relied heavily on numerical weather prediction (NWP), which has been shown to retrieve vital information embedded behind severe weather [17]. However, the physical parameterizations of models [18], fine-scale weather prediction, and high computational costs pose significant limitations. These limitations have opened a new thought of adoption of machine learning (ML) algorithms, which are considered ideal ways to achieve reliable predictions based on complex features [19,20]. ML methods are successful in reducing computational costs and improving the accuracy and reliability of predictive models [21].

Moreover, ML methods, including artificial neural network (ANN), support vector machine (SVM), and random forest (RF), have grown in attention in different fields [22], including hydrology and atmospheric sciences. These methods are expected to boost and substitute physically based numerical models for accurate prediction. Machine learning also has the capability to deal with big data, making it an empowering tool for remote sensing observation and analyzing tremendous data from GCM experiments [23].

Among the ML algorithms, the RF model has shown exceptional performance in establishing regression relationships by utilizing multiple datasets [24]. Its non-parametric approach and ability to retrieve variable importance make it a reliable and preferred choice in various scientific research applications. The integration of the RF algorithm with numerical models has resulted in remarkable improvements and uncertainty reductions, particularly in weather research and forecasting [25,26].

Temperature and precipitation in a given region are typically influenced by a complex interplay of various climate factors. Developing an accurate model for short-term climate prediction requires careful consideration of predictors that are closely associated with the predictand. Previous studies have identified a range of climate indices, including the El Niño/La Niña phenomenon, Indian Ocean Dipole (IOD) [27], Pacific Decadal Oscillation (PDO) [28], and others, as significant drivers of temperature and precipitation changes. In addition, anthropogenic activities have been identified as a significant contributing factor to extreme precipitation [29], particularly in southern China [30,31]. However, the dynamic nature of climate indices remains a source of uncertainty in prediction results. While previous studies have primarily focused on natural factors to forecast future climate, short-term climate prediction is determined by the interaction of both natural and anthropogenic factors [32,33,34]. Therefore, relying solely on natural factors for short-term climate prediction is inadequate, and the introduction of human-induced effects into the predictive model requires further evaluation to improve the accuracy of the forecast.

In this study, we aim to focus on integrating anthropic and natural factors in machine learning methods to minimize uncertainties in predicted STC of air temperature and precipitation across Mainland China. Our main objective is to predict short-term climate in China using machine learning algorithms, mainly the random forest (RF) model, while considering both natural and anthropic factors. We aim to evaluate the role of both factors in short-term climate prediction and to compare the performance of the RF model using either factor alone and both combined. This study’s findings can provide insights for sustainable decision-making in China and contribute to global efforts to mitigate climate change impacts. To provide a concise understanding of the study, the article is structured as follows: In Section 2, we describe the methodology used in this study. This section includes a detailed explanation of the data collection and processing techniques employed, as well as an overview of the random forest model and evaluation metrics utilized. Section 3 presents the study’s results, starting with the hindcasting of air temperature and precipitation distributions, followed by short-term climate prediction for 2019 and 2020. Additionally, this section also presents the evaluation of the role of natural and anthropic factors in the climate predictions. In Section 4, we provide a discussion of the study’s findings. This includes a discussion of the significance of the findings, a comparison with previous studies, and the limitations and future directions of the research. Finally, in Section 5, we summarize the key findings of this study and discuss their implications for sustainable decision-making in China. Furthermore, we provide recommendations for future research that can help to further develop and improve the methodology used in this study.

2. Data and Methods

2.1. Data

2.1.1. Meteorological Data

In this study, we used a comprehensive dataset of monthly mean precipitation and air temperature collected from 160 weather stations located across mainland China. The spatial distribution of these weather stations can be seen in Figure 1. These datasets are maintained and supported by the National Climate Centre (NCC) and are considered highly accurate and reliable. Access to these datasets can be obtained through the NCC website at https://cmdp.ncc-cma.net/nccdownload/index.php?ChannelID=1, accessed on 12 January 2021.

As shown in Figure 1, the weather stations are densely and homogeneously distributed across Southern, Central, and Eastern China, but sparsely in Western China. Of particular concern is the Tibetan Plateau, where the number of weather stations is the lowest, potentially affecting the accuracy of our model in that region. Nevertheless, these datasets can still reflect the climate distribution and changes in air temperature and precipitation over Mainland China.

2.1.2. Natural Factors

In addition to the comprehensive dataset of monthly mean precipitation and air temperature obtained from 160 weather stations across Mainland China, we also incorporated several natural factors that have been shown to play a significant role in the spatial and temporal trends of climatic extremes in China. These factors include the Arctic Oscillation (AO), Pacific Decadal Oscillation (PDO), North Atlantic Oscillation (NAO), and El Niño Southern Oscillation (ENSO). We obtained monthly datasets of the aforementioned indices from the National Centers for Environmental Prediction/National Weather Service/NOAA/US Department of Commerce. These datasets can be accessed through the following link: https://psl.noaa.gov/data/climateindices/list, accessed on 12 January 2021.

As shown in Figure 2, the variability in climate indices indicates that the AO and NAO have recently dominated with positive phases, while the ENSO and PDO indices were prevailed on by the negative phase. Moreover, our analysis shows that the period before 1990 was characterized by the dominance of the positive phase on the ENSO and PDO indices. By considering these natural factors alongside the station observations, our model aims to capture the most relevant and effective predictors of short-term climate trends in China.

2.1.3. Anthropic Factors

We included anthropic factors in our analysis to examine their impact on short-term climate, as economic activities can represent some human influences [35]. Specifically, we examined the following 5 socioeconomic indicators: Gross Domestic Product (GDP), GDP of primary industry (GDPP), GDP of the secondary industry (GDPS), GDP of the tertiary industry (GDPT), and per capita gross domestic product (GDPPC). These datasets were obtained from the National Bureau of Statistics (NBS) and can be accessed at https://data.stats.gov.cn/index.htm, accessed on 15 January 2021. The temporal trends of these selected factors are presented in Figure 2e.

Based on Figure 2, the natural indices exhibit oscillations of 2–3 years and 7–10 years, and studies have shown that the phase and variations of these indices have important predictive implications for STC prediction over China mainland [36]. Meanwhile, all the anthropic factors showed a significant increasing trend from 1979 to 2018. GDP, GDPT, and GDPS exhibited an exponential increasing trend, with GDP showing a rapid increase in the index around 1993, while GDPS and GDPT exhibited a rapid increase around 2005. On the other hand, GDPP and GPPPC showed a smoother, linear-like increasing trend. These anthropic factors reflect human activities, which also have an impact on the STC changes in China mainland [35].

2.2. Methods

2.2.1. Random Forest Model

The random forest (RF) method is a useful machine learning algorithm developed by Breiman (2001) [37] to address both classification and regression problems. It is a non-parametric and ensemble learning method that combines multiple decision trees to achieve highly accurate predictions. RF is well-suited for modeling complex relationships and interactions among variables, as well as handling high-dimensional datasets with many predictors.

In this study, we adopted the RF method to investigate the impact of integrating natural and anthropic factors on short-term climate predictions over China. The RF algorithm is highly versatile, allowing for the inclusion of a variety of variables in the model to identify their relative importance and contribution to the final prediction. Additionally, RF provides a robust measure of variable importance, enabling us to identify the most influential variables in the prediction process.

We employed RF algorithms to perform hindcasting and forecasting of seasonal temperature and precipitation across China. The aim was to evaluate the effect of natural and anthropic factors as predictors of climate changes.

We established a separate model for each station when training the models. For the prediction of climate variables for a specific season, we utilized the data from the preceding season as input values. For instance, to predict the temperature and precipitation for the JJA (June, July, and August) season, we used the data from the MAM (March, April, and May) season. Similarly, to predict the climate variables for the DJF (December, January, and February) season, we used the data from the SON (September, October, and November) season. When considering only natural factors, we used the previous season’s temperature, precipitation from all 160 stations, and other climate indices as input values to establish the relationship with the specific station’s temperature and precipitation of the current season. When only considering anthropic factors, we used the previous season’s economic indices to establish the relationship between the temperature and precipitation of the current season. Finally, when considering both natural and anthropic factors, we used the data from the previous season, including temperature, precipitation, other climate indices, and economic indices, to establish the relationship between temperature, precipitation, and these factors.

The entire training process was performed using the data from 40 years (1979–2018) of the natural, anthropic, or combined datasets of the preceding season to train the models for the DJF or JJA station’s temperature or precipitation. Subsequently, we used each station’s model for hindcasting the temperature and precipitation and forecasting for 2019 and 2020.

In this study, we utilized the scikit-learn 1.2.2 package in Python (https://scikit-learn.org/stable/modules/ensemble.html#forests-of-randomized-trees, accessed on 16 January 2021) to establish a random forest regression model for prediction. Seasonal datasets for the natural factor, anthropogenic factor, and combined factor were created using the predictive factors outlined in Section 2.1.2 and Section 2.1.3. These datasets were used to correspondingly predict air temperature and precipitation for each meteorological station. Subsequently, models were constructed for each station. Hindcasting and forecasting were performed using these individual models for each station.

2.2.2. Statistical Metrics

The STC of air temperature and precipitation in a specific year is generally measured using temperature anomaly (TA) and percentage of precipitation anomaly (PPA). These measures are used because the anomalies in a specific year are often large, and TA and PPA can more intuitively show whether the STC is warmer (wetter) or cooler (drier) in that year. The expressions for TA and PPA are as follows:

$$TA=T-T_{avg}$$

(1)

$$PPA=\frac{\left(P-P_{avg}\right)}{P_{avg}}\times 100\%$$

(2)

where T is the average seasonal (JJA or DJF) air temperature in a specific year, T_avg is the average seasonal air temperature over a certain period of time (in this study, it is 1979 to 2018), P is the average seasonal precipitation in a specific year, and P_avg is the average JJA or DJF precipitation over that same period of time. The units of TA, T, and T_avg are degrees Celsius (ºC), while P and P_avg are in units of millimeters (mm). PPA is a dimensionless variable.

To evaluate the difference between predictions and actual observations, statistical metrics are used to evaluate model performance in the hindcasting and predicting process. The metrics adopted include mean bias error (MBE), percentage error (PE), and the coefficient of determination of the prediction (R²). Their formulations are shown below:

$$MBE=\frac{1}{N}{\displaystyle \sum }_{t=1}^n\left[y_{s,t}^{pred}-y_{s,t}^{obs}\right]$$

(3)

$$PE=\frac{1}{N}{\displaystyle \sum }_{t=1}^n\frac{y_{s,t}^{pred}-y_{s,t}^{obs}}{y_{s,t}^{obs}}\times 100\%$$

(4)

$$R^2=1-\left[\frac{{\sum }_{t=1}^n{\left(y_{s,t}^{obs}-y_{s,t}^{pred}\right)}^2}{{\sum }_{t=1}^n{\left(y_{s,t}^{obs}-{\overline{y}}_s^{obs}\right)}^2}\right]$$

(5)

$${\overline{y}}_s^{obs}=\frac{1}{N}{{\displaystyle \sum }}_{t=1}^n{y}_{s,t}^{obs}$$

(6)

Herein, $y_{s,t}^{pred}$ and $y_{s,t}^{obs}$ express the predicted and observed values (temperature or precipitation), respectively, at station $s$ and time $t$. ${\overline{y}}_s^{obs}$ stands for mean observed values at a given meteorological station $s$.

The best possible score for $R^2$ is 1.0, and it can be negative as the model can perform arbitrarily worse than predicting the mean of the target variable. If a constant model that ignores the input features and always predicts the expected value of $y_{s,t}^{obs}$ were used, it would obtain a constant score of $R^2=0.0$ [38,39].

In summary, we utilized the RF algorithm to explore the contribution of natural and anthropic factors to short-term climate in China. By combining various predictor variables, we aimed to improve the predictability of climate extremes and provide insights into the underlying drivers of climate variability in the region.

3. Results

3.1. Hindcasting of STC

3.1.1. Model Performance Evaluation

After training the data using the RF algorithm, we obtained the model scores of precipitation and temperature prediction for each site under different scenarios. Figure 3 displays the model performance scores of temperature prediction under three different scenarios, based on the selection of predictors, where the horizontal axis represents the station IDs (from 1 to 160), and the vertical axis shows the performance scores of each station’s prediction model. Our findings indicate that both natural and anthropic factors are strong predictors of temperature, with performance scores ranging between 0.86 and 0.98. Interestingly, anthropic factors exhibit a stronger influence on the RF model performance than natural factors. The use of anthropic factors for seasonal temperature prediction improved the RF model performance at several stations, with higher performance scores exceeding 0.95. However, a few stations exhibited lower performance scores ranging between 0.90 and 0.95. These results indicate that anthropic factors have a significant relationship with temperature variations.

Moreover, the combination of anthropic and natural factors showed better performance than using a single type of predictor. The scenario of combined factors yielded a higher RF model performance, with scores above 0.98. Therefore, the consideration of both anthropic and natural factors in the RF model can significantly predict and explain over 98% of the temperature datasets.

In contrast to the temperature predictions, the prediction of precipitation demonstrated distinct behavior, displaying dynamic variability against anthropogenic factors and revealing significant uncertainty (Figure 3). The RF model’s performance in predicting precipitation using only anthropogenic factors ranged between 0.30 and 0.70, averaged about 0.50, with some stations exhibiting weak scores of less than 0.20. Natural factors, on the other hand, were found to be more effective inputs for predicting precipitation (about 0.85), particularly in areas with relatively lower levels of human activity, such as the Qinghai-Tibet Plateau and northwest region.

Similar to temperature prediction, the incorporation of both natural and anthropogenic factors has been shown to improve the predictability of precipitation (blue line in Figure 3). The RF model demonstrated higher performance scores, averaging at 0.95, compared to the use of natural factors alone, indicating that the integration of socioeconomic factors in models is crucial for predicting and explaining the complex variations in precipitation. Furthermore, the consideration of both natural and anthropogenic factors can enhance the accuracy of precipitation predictions and offer better insights into the impact of human activities on local climate dynamics [29,40].

The performance scores of the short-term climate models for temperature and precipitation indicate that including both natural and anthropic factors can lead to enhanced model accuracy. Therefore, it is important to consider both types of factors when utilizing machine learning techniques to predict short-term climate in Mainland China. In the following section, we will present the hindcasting results of the RF model for STC over Mainland China.

3.1.2. Seasonal Hindcasting Performance

Based on the previous analysis, it is evident that the model that incorporates both natural and anthropic factors outperforms the others. To further evaluate the model’s performance, we employed the combined model to hindcast historical temperature and precipitation data from 1979 to 2018 and compared the predicted results with the observed values.

The hindcasting results and comparison with observed values for monthly mean air temperature are shown in Figure 4. Panels (a) and (d) present the observed results for DJF and JJA, respectively; panels (b) and (e) show the hindcasting results; panels (c) and (f) depict the mean bias error (MBE) between observed and hindcasted values.

The hindcasting results for the winter season indicated that the combined model accurately predicted the mean temperature at each weather station, with both hindcasted and observed values conforming to the normal distribution of DJF temperature ranges ranging between −5 °C and 20 °C. The mean temperature gradually decreased northward, with the northern regions exhibiting a temperature range of 0 °C to 10 °C, while southern China exhibited a warmer temperature range of 10 °C to 20 °C. The MBE values between the observed and hindcasted temperature values were low, ranging from −1 °C to 1 °C.

For the summer season, the combined model predictions were consistent with the observed values, exhibiting similar spatial patterns that featured higher temperatures in the southern regions and lower temperatures in the northern regions, with a gradual decrease from southeast to northwest. Moreover, the MBE values for the JJA temperature were also low, ranging between 0 and 1 °C, indicating a high level of accuracy in the hindcasted results.

Overall, the hindcasting results for the winter season tended to overestimate the observed temperatures, while those for the summer season tended to underestimate them, but with similar spatial distribution patterns to the observed values and with errors ranging below 1 °C, which is acceptable for STC prediction. These findings suggest that the combined model performs well in predicting temperature variability.

Compared to temperature, the results for monthly mean precipitation (Figure 5) were more complex. Both the observed and hindcasted values showed a decreasing trend from southeast to northwest, ranging from 250 mm to less than 50 mm. At the same time, the southern regions displayed a homogenous pattern of abundant seasonal precipitation, ranging between 100 mm and 300 mm, whereas the northern regions received relatively low precipitation, less than 50 mm. These findings highlight the spatial heterogeneity of precipitation distribution across the study region.

For DJF precipitation, the MBE was mainly negative (Figure 5c), indicating that the combined model slightly overestimated DJF precipitation values, particularly in central and northern China, with the MBE ranging around 10mm. However, positive MBE values were found in the Yangtze–Huai River Basin and northwest of Xinjiang, indicating that the combined model exhibited a weak underestimation of less than 5 mm in parts of the southeast and northwest regions of China. Further, the MBE values for JJA precipitation (Figure 5f) presented a contrasting pattern to the DJF, with the combined model showing a slight overestimation of less than 5 mm in the southeast and underestimation ranging from 1 mm to 20 mm in the remaining regions.

In summary, the hindcasting results for both winter and summer seasons’ temperature and precipitation have shown the model’s high predictive skill and its ability to reproduce the spatial patterns observed in the real-world data. The specific impacts of natural and anthropogenic factors will be discussed in detail in the following sections.

3.2. Analyzing the Role of Natural and Anthropogenic Factors on STC

To examine the individual impacts of natural and anthropogenic factors on STC, we conducted two sensitivity experiments using the RF model separately. The first test utilized only natural factors as input variables, while the second test incorporated only anthropogenic factors as input variables. Subsequently, to evaluate the respective roles of these factors in STC prediction, we compared the hindcasting results from these tests with the observed data.

The hindcasted results of DJF temperature using either natural or anthropogenic factors are displayed in Figure 6a and Figure 6b, respectively. Both results exhibit spatial patterns similar to the actual observed data (Figure 4a), with negative MBE values. Specifically, the results reveal that natural factors present a lower MBE of less than 1 °C (Figure 6c) in comparison to the anthropogenic factors. In contrast, the anthropogenic factors result in a relatively higher MBE, particularly in northern China, with the MBE close to 2 °C (Figure 6d). This indicates that predicting temperature solely based on anthropogenic factors can lead to overestimation, particularly in northern China.

The hindcasted results of JJA temperature distribution (Figure 6e,f) exhibit spatial patterns similar to the actual observed data (Figure 4d). However, the MBE shows a completely opposite pattern. The hindcasted result using only natural factors overestimates the temperature in the north and eastern plateau, while underestimating the temperature in the Huang–Huai River Basin and the Yangtze–Huai River Basin. On the other hand, the result using only anthropogenic factors exhibits the opposite pattern, overestimating the temperature in the southeastern region with some stations’ MBE exceeding 2 °C, while underestimating the temperature in the north with an MBE of around 1 °C.

In addition to temperature, we assessed the predictability of precipitation using the RF model. Figure 7 displays the hindcasted precipitation and the MBE with respect to the observations. Similar to the temperature results, the hindcasted precipitation distribution using either natural factors (Figure 7a) or anthropogenic factors (Figure 7b) exhibits patterns consistent with the observations (Figure 5a). The MBE results of the hindcasted precipitation and observations indicate that the hindcasted precipitation using natural factors has a slight overestimation in the northwestern and eastern plateau regions, while overall, it underestimates precipitation, particularly in the southwestern region, where the MBE reaches 20 mm (Figure 7c). In contrast, the result using anthropogenic factors exhibits the opposite pattern, overestimating the DJF precipitation overall, with an MBE of around 5–20 mm, but with an MBE of less than 3 mm in the southeastern region where the natural factors do not perform well (Figure 7d). These findings suggest that natural factors alone cannot fully account for the variability of DJF precipitation in this region, and that anthropogenic factors have a greater impact on DJF precipitation in this area.

It is noteworthy that the hindcasted JJA precipitation using only natural or anthropic factors also exhibit opposite patterns in terms of MBE distribution. The MBE of the natural factors result performs relatively well in the Huang–Huai and Yangtze–Huai River Basins, with an MBE of less than 10 mm. The result shows a slight underestimation in the Huang–Huai River Basin and a slight overestimation in the Yangtze–Huai River Basin, while other regions exhibit an underestimation of precipitation ranging from 10–30 mm (Figure 7g). On the other hand, the anthropogenic factors result exhibits the opposite pattern, with most regions showing slight overestimation of precipitation, while the entire southeastern region shows a precipitation underestimation exceeding 30 mm. Notably, both methods exhibit an MBE of less than 10 mm in the important rainfall belt of the Huang–Huai River Basin in China.

Overall, these findings suggest that solely considering either natural or anthropogenic factors may result in inaccurate temperature predictions, with the degree of inaccuracy varying depending on the region and the type of factor considered. However, as shown in

Table 1. The seasonal range of the MBE estimated between hindcasted values under different input features and observed meteorological variables (temperature and precipitation) from 1979 to 2018.

Season	Feature	Temperature (°C)		Precipitation (mm)
		min	max	min	max
DJF	Natural	−1	0	10	20
	Anthropic	−2	0	−20	0
	Combined	−1	0	−10	10
JJA	Natural	−1	1	−10	30
	Anthropic	−2	2	−30	30
	Combined	0	1	0	20

, the hindcasted result using only natural factors slightly outperforms the result using only anthropogenic factors. This is because human influence can indirectly affect the variations in the selected natural factors. The combined model improves the accuracy of predicting STC by integrating both factors. The next section will use the combined model to conduct prediction experiments for specific years to demonstrate the advantages of the combined model.

3.3. STC Prediction in 2019 and 2020

Based on the analysis above, it is evident that the combined model performs reliably and outperforms the use of either natural or anthropic factors alone. In the following section, we use the integrated model to predict the temperature and precipitation in the DJF and JJA seasons for 2019 and 2020. Note that due to the abundance of extreme temperature and precipitation values in specific years, temperature anomaly (TA, Equation (1)) and percentage of precipitation anomaly (PPA, Equation (2)) are commonly used to represent the short-term climate (STC) characteristic of that year. For the prediction of STC, the critical evaluation criteria are as follows: (1) whether the pattern of temperature/precipitation anomaly is correctly predicted, and (2) whether the anomalies of the rain belt are accurately forecasted [41].

Figure 8 displays the observed (first row) and predicted (second row) TA, as well as the respective percentage errors (PE, third row) for the 2019 (left column) and 2020 (right column) DJF seasons. The combined model demonstrated good agreement with observations for DJF temperature anomalies in 2019 and 2020, with warmer anomalies observed in coastal regions and cooler seasons in central and western China. Further, the model exhibited a slight underestimation and overestimation of DJF temperature in northern and southern China, respectively, and the respective PE ranged between 0 and 20%. However, the model showed higher and mixed percentage errors (|PE| > 20%) north of the Huang–Huai River Basin.

As shown in Figure 9, the TA for JJA in 2019 (Figure 9a) and 2020 (Figure 9b) was negative in the north and positive in the south. In 2019, the temperature anomaly was higher in the Jianghuai and Huang–Huai River Basins and the eastern plateau, and lower in the northeast and Inner Mongolia. Regarding these patterns, the combined model (Figure 9c) showed good agreement with the observations, with PE < 5% at most stations. Similarly, in 2020, positive TA appeared mainly in the southeast coastal region, while a negative anomaly was observed in the northwest and central China. The output trend of the combined model (Figure 9d) was consistent with the observations, but with a slight underestimation in the numerical values. In summary, in terms of JJA TA, the combined model predicted cooler temperatures in the northeast and warmer conditions in the rest of China, with a PE range between −10% and 10%. The forecast of temperature anomalies was nearly accurate compared to the observed JJA temperature, with a slight underestimation in the north and overestimation in the south, and lower PPA of less than 5% at most stations.

As shown in Figure 10a,b, abundant precipitation was observed in the Yangtze–Huaihe River Basin and southern China in 2019 and 2020, indicating a southward shift of the rain belt. The combined model (Figure 10c,d) managed to reproduce such patterns accurately, with a PPA exceeding 100% in the Yangtze–Huaihe River Basin, indicating a southward shift of the rain belt. However, the model may have failed to retrieve accurate rainfall intensities, which may have led to higher percentage errors (|PE| ≥ 20%). Specifically, the model underestimated DJF precipitation, particularly in northern China, with a PE ≥ 10%, and overestimated precipitation in southern China, with a PE ≥ −10% (Figure 10e,f). Overall, the model tended to overestimate the PPA in 2020.

Furthermore, the comparison between the DJF of 2019 and 2020 showed that the combined model overestimated the DJF precipitation of 2020 to a greater extent than that predicted for 2019. Both 2019 and 2020 were El Niño years, and previous research has indicated that the precipitation distribution in China tends to be biased towards the east and less towards the west, with flood-prone conditions in the south and drought-prone conditions in the north. The observed results also exhibited these characteristics, with a high PPA exceeding 300% in the Huai River basin and southern China in 2019 (Figure 11a). The predicted results of the combined model (Figure 11c) also reflected these patterns, with the high values of PPA concentrated in the above regions.

In 2020, extreme PPA was observed in the Huang–Huai River Basin (Figure 11b), which was also indicated by the model prediction (Figure 11d). Combining the features of these two years, the combined model predicted a northward migration of precipitation and significant patterns of higher, medium, and lower rainfall distribution in the south, central (extending northeast), and northwest, respectively. The model also captured the heterogeneity patterns of precipitation, especially for the lower and medium rainfall in northwest and northeast China, which was consistent with the observed results. However, the model encountered relatively higher precipitation amounts in the south, and the predicted intensity was lower than observed, with PE varying between 0 and 30% (Figure 11e,f). Additionally, the model underestimated JJA precipitation in the south and slightly overestimated the rainfall intensity in the north.

In conclusion, the combined model exhibited promising performance in predicting seasonal spatiotemporal climate trends in China mainland, which can serve as a valuable tool for informing policy- and decision-making aimed at mitigating the impacts of climate change. However, further improvements are needed to enhance the accuracy and precision of extreme value predictions in the combined model for STC forecasting.

4. Discussion

This study aims to assess the impact of natural and anthropic factors on short-term climate in China using machine learning algorithms, particularly the random forest (RF) model. Our results suggest that combining natural and anthropic factors improves climate trend predictions significantly, with the model scores increasing by 0.20 and 0.40 when the two factors were combined compared to using only one factor.

Previous studies have also emphasized the importance of considering both natural and anthropic factors when assessing climate predictions [42]. We found that regions with a significant increase in anthropic factors also showed a higher frequency of extreme precipitation. Our sensitivity analysis showed that the use of anthropic factors led to a lower performance of the RF model at some stations than the use of natural factors. This may be due to the use of anthropic datasets averaged at the national scale, while China has diverse social–economic development and climate characteristics.

Our results also showed that the RF model exhibited better performance in predicting extreme temperatures than extreme precipitation. While the model demonstrated significant retrieval of precipitation patterns across sub-regions of China, further optimization is required to improve the predictive skills of precipitation intensity.

5. Conclusions

In conclusion, this study evaluated the impact of natural and anthropic factors on short-term climate in China using the random forest model. The findings suggest that a comprehensive consideration of both natural and anthropogenic factors significantly improves the accuracy of climate trend predictions compared to using only one factor. The model scores (the coefficient of determination) exceeded 0.95 when both factors were combined.

The study also attempted to predict similar components for 2019 and 2020, demonstrating that the combined model’s performance exhibited a significant improvement in predicting short-term climate. These results underscore the importance of accounting for both natural and anthropic factors in predicting climate trends and informing sustainable decision-making in China. Future studies could improve the accuracy of precipitation intensity prediction by optimizing the RF model.