Prediction and Urban Adaptivity Evaluation Model Based on Carbon Emissions: A Case Study of Six Coastal City Clusters in China

Abstract

Aiming at predicting the issues of social economics, environmental pollution, climate change, and marine disasters influenced by carbon emissions, a predicting model based on carbon emissions with the Random Forest (RF) model was constructed. Meanwhile, a novel urban adaptivity evaluation model is put forward considering the above four domains of indicators; hence, the predicting and evaluation models are integrated. Six coastal city clusters of China are selected as study areas and the result of the RF model with carbon emissions shows that northern city clusters suffer more pollutant loads due to their heavy industry layout; southern cities generally have higher GDP, while they are more vulnerable toward extreme weather and marine disasters. The result of the evaluation system indicates that northern city clusters have higher urban adaptivity (0.49–0.50) due to their balance between economics and pollution as well as less vulnerability to climate change because of their relatively high latitude. On the contrary, southern cities should focus on environmental pollution and tropical storms to pursue superior compatibility.

1. Introduction

Global warming has been a globally prominent issue due to the massive emissions of carbon dioxide [1] and other Greenhouse Gases (GHG), which attract substantial awareness with their effect in recent years. Sharply increasing carbon emissions in the past decades have become the spotlight of all countries around the world [2,3]. Therefore, the extensive use of fossil energy sources has led several countries to search for the correlation between carbon emissions and other key issues [4,5,6,7], such as social economics, pollution control, climate change, and marine disasters, etc. Thus, numerous studies carried out in these years focused on the relationship between the aforementioned four domains and carbon emissions [8,9], of which carbon emissions has always been the output result of other driving factors, while it can also be treated as an input variable for predicting other indicators by contrary; hence, we carried out this research.

As for social economics, China has imposed strict policies in dealing with the complex problem between carbon emissions and social progress, working on boosting its economic prosperity as well as reducing carbon emissions and environmental pollution by improving energy utilization efficiency. These works mainly concentrate on the relationship between social economics, environmental pollution, and energy consumption, while neglecting the influence of carbon emissions [10,11]. Nevertheless, the situation of aquatic pollution and air pollution is severe in many cities of China, and the amount of urban environmental pollutants and carbon emissions varies in a similar tendency. Hence, studies should also be carried out to quantify the relationship between them [12]. Additionally, climate change is another concern closely related to carbon emissions, and can even be regarded as the most predominant issue that influences people’s daily life directly. This directly leads to the increasing frequency of extreme temperature and precipitation due to massive carbon emissions, thus the relationship among them is also to be explored [13]. Furthermore, coastal cities are facing unprecedented risks of marine disasters due to carbon emissions, such as the increasing possibility of cyclones or storms, and the steady rising of the sea level, while only a few studies focused on carbon flows that integrated sea and land through natural and anthropogenic processes [14,15]. Therefore, for quantifying the relationship between carbon emissions and marine disasters, it is essential to incorporate this factor into the comprehensive evaluation system.

With the above four domains related to carbon emissions, it is necessary to adopt appropriate methods to comprehensively predict all the above results related to carbon emissions. Numerous models and algorithms have been utilized in this area, such as gray prediction theory [16,17], time series prediction methods [18,19], linear regression prediction [20,21], and non-linear regression prediction [22,23]. These models mainly use historical statistical data as the basis for prediction, and thus have high requirements for the quality of information and materials [24]. Compared with these conventional methods, Machine Learning (ML) algorithms are supposed to solve statistical prediction problems more efficiently and perform better while dealing with complex and tedious data. Random Forest (RF) has gained remarkable popularity for its ability to handle multi-dimensional classification and regression problems with excellent accuracy and limited possibility of overfitting [25]. Many studies have proved that RF outperforms other popular ML algorithms in various application scenarios [26] and is supposed to be superior in model interpretation and faster to train; thus, it is chosen to be applied in this research.

The existing research mainly focuses on predicting a single factor related to carbon emissions, while lacking a comprehensive system that incorporates multiple important domains; thus, such a study may not be persuasive enough. For solving such a problem and proposing a more reasonable method, we construct a novel urban adaptivity evaluation system involved in the area of social economics, environmental pollution, climate change, and marine disasters, which consists of numerous indicators. The annual carbon emission amount of each city cluster is treated as an input dataset for this model; by training the input data and exporting the output results with RF, we assess the prediction results and calculate urban adaptivity therein. Six city clusters around coastal areas of the Chinese mainland are selected as the study area due to their prosperous economy, high population density, and developed industry. With the predicted and evaluated results, we further explore the relationship between carbon emissions and other factors therein.

2. Materials and Methods

2.1. Study Area

Coastal provinces of east China are regarded as the most prosperous areas within the country, which mainly ranges from 20° N to 40° N in latitude and from 110° E to 130° E in longitude. With mild temperatures and substantial precipitation, these provinces host nearly 44.52% of the population and 42.53% of the GDP for the country. Developed industry boosts the progress of local urbanization and modernization while also producing large quantities of GHG emissions as well as environmental pollutants. The problems related to carbon emissions are involved in the area of environmental protection, climate change, and marine disasters, etc., which causes great pressure upon the sustainable development of coastal cities. Therefore, the Chinese government is working on improving energy utilization efficiency and cutting off carbon emissions, while boosting the prosperity of the economy.

The aforementioned six city clusters are distributed in the prosperous areas of China (as shown in Figure 1 and

Table 1. Composition of six coastal city clusters.

Name of City Clusters	Cities Contained
Mid-southern Liaoning City Cluster (MLCC)	Shenyang, Dalian, Anshan
Beijing-Tianjin-Hebei City Cluster (BTHCC)	Beijing, Tianjin, Shijiazhuang, Tangshan
Shandong Peninsula City Cluster (SPCC)	Jinan, Qingdao, Yantai
Yangtze River Delta’s City Cluster (YRDCC)	Shanghai, Suzhou, Hangzhou, Nanjing, Ningbo, Wuxi
Southeastern Fujian City Cluster (SFCC)	Quanzhou, Fuzhou, Xiamen
Pearl River Delta’s City Cluster (PRDCC)	Guangzhou, Shenzhen, Foshan, Dongguan, Zhuhai

): Mid-southern Liaoning City Cluster (MLCC), Beijing-Tianjin-Hebei City Cluster (BTHCC), Shandong Peninsula City Cluster (SPCC), Yangtze River Delta’s City Cluster (YRDCC), Southeastern Fujian City Cluster (SFCC), and Pearl River Delta’s City Cluster (PRDCC). Required data is obtained from Urban Economy Development & Environmental Pollution Panel Data of 295 cities from 1996 to 2017 (https://github.com/kyzheng196/Calibration-algorithm-for-degradation-coefficient-K/blob/main/Data%20source.zip, accessed on 10 October 2022). Among the city clusters, each city cluster is regarded as a research unit, and the mean value of all the cities within each cluster is treated as input data for our prediction model. With the prediction and analysis of various aspects of these city clusters, we may conclude certain suggestions for futural development.

2.2. Random Forest

Machine learning has been an ideal tool in data analysis, especially for uncovering the relationship among large datasets. Due to its inherent characteristics and superior classification performance, RF is regarded as the best ML algorithm for dealing with the inner relationship among data. This model was developed by Breiman as the improvement of the classification and regression tree (CART) to improve its predicting accuracy. This method realizes better predictive performance without obviously extending computational expense. It has a random selection process consisting of inputting variables and data, and generating certain classification and regression trees [27]. With the above mechanism, the RF algorithm integrates all the decision trees in which each tree depends on a random vector that is independent and identically distributed (as shown in Figure 2).

According to the detailed diagram of the RF algorithm and the regression classification principle of CART, we defined several following probabilities based on the decision tree principles, of which N(t) is regarded as the sample number of the decision tree training sample data set L at node t, N_j(t) represents the number of sample data at node t that belong to the jth category w_j, and n represents the total number of training sample data.

(1) The sample probability p at node t based on L is estimated as

$$p\left(t\right)=\frac{N\left(t\right)}{n}$$

(1)

(2) With the dataset L, the sample probability p at the corresponding node t contained in the category w_j can be expressed as

$$p(w_j\mid t)=\frac{N_j\left(t\right)}{N\left(t\right)}$$

(2)

(3) We assume that t_L and t_R represent the left and right sub-nodes of t, respectively, thus the corresponding sample probabilities based on data set L (p_L and p_R) can be demonstrated as

$$p_L=\frac{p\left(t_L\right)}{p\left(t\right)}$$

(3)

$$p_R=\frac{p\left(t_R\right)}{p\left(t\right)}$$

(4)

Based on the proportion of the sample data belonging to each category at node t, the corresponding category at t can finally be determined. Considering the relationship between input variables and output variables, we set carbon emissions as the input data and treat data involved with social economy, environmental pollution, climate change, and marine disasters as output data. As for the input dataset, the amount of annual carbon emissions of each city cluster is regarded as the input variable. Then, to train the RF model with the above variables and predict the corresponding results, an evaluation system was constructed.

Furthermore, the RandomForestRegressor of the sklearn (Scikit-learn) package is utilized to realize this model, and the Bayesian optimization algorithm is incorporated for hyperparameter tuning, which uses TreeBagger to train the Random Forest model, specifies the parameters to be tuned, and returns that to the out-out-bag index. Among all parameters, n_estimators is set as 100, max_depth is set as 8, max_features is set as ‘auto’, and crirerion is set as ‘mse’ (https://www.scikitlearn.com.cn/). During the training process, K-fold cross-validation is applied to split the data into the training set and testing set, thus to evaluate the accuracy of this model.

2.3. Urban Adaptivity Evaluation System

As discussed above, we classify the input data into four categories, which involve the area of social economics, environmental pollution, climate change, and marine disasters; much research has focused on these fields before [4,5,12] and has obtained meaningful results. Due to the aforementioned four categories having an inner relationship with carbon emission and the data being more available, they’re chosen as representative indicators of this work. With each category, we set some indicators and assign corresponding weights attached to them, thus building an urban adaptivity evaluation system (as shown in Figure 3). The determination of weights is based on the above works, in which Han et al. [28] constructed a calculation model with energy consumption and pollution emission, Wang et al. [29] focused on the relationship between pollution and green finance, and Wu et al. [30] worked out the impact of energy structure and environmental regulation on air pollution. The four categories are all related to carbon emissions, where social economics and climate change have higher weights as they are supposed to be more significant. Such an evaluation system comprehensively considers the aforementioned fields around carbon emissions, and thus it is supposed as a novel element of this work.

In this evaluation system, the percentage of all domains and indicators represents their relative weight compared with others. Considering their relative relation within this system, the domain of social economics and climate change are set to 30% weight, while the domain of environmental pollution and marine disasters are set to 20% weight, which may reflect their relative importance and mutual connections to some extent; for coastal city cluster residents, they’re inclined to care more about economic life and climate change, while cyclones and environmental problems also matter, thus the relative weight is determined eventually.

The urban adaptivity evaluation system comprehensively considers the influence of all aforementioned aspects, representing the harmony and balance among social development, environmental protection, and other areas. Therefore, it is supposed to be the objective function that reflects the overall performance of city clusters under a certain time scale.

For the area of social economics, the GDP of each city is selected as the single representative indicator; the indicator of environmental pollution contains the emissions of sewage, sulfide dioxide (SO₂), and smoke dust; climate change covers six subareas: annual mean temperature, annual highest temperature, annual lowest temperature, annual relative humidity, annual total sunshine duration, and annual total precipitation; as for marine disasters, it is mainly concerned with the increment of sea level and the influence of tropical storms.

In view of the difference of city numbers within each city cluster, we utilize the average value of each indicator as the input data to offset the influence of the city number. The relative weight [28,29,30] of each indicator within the four domains is obtained based on the consideration of their relative importance compared with other indicators. All the data was normalized with Equation (5) to be analyzed in the same order of magnitude, and then was imported into the risk evaluation system to calculate the eventual urban adaptivity value with Equation (6).

$$y_i=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}$$

(5)

where x_i is the ith value of dataset, x_min is the minimum value of dataset, x_max is the maximum value of dataset, and y_i is the normalized value of x_i.

$$A_i=\sum _{j=1}^n{w}_{1j}\left(\sum _{k=1}^m{w}_{2k}{y}_k\right)$$

(6)

where A_i is the adaptivity of city cluster i, i is the index of city cluster, j is the domain index of evaluation system, n is the domain number of evaluation system, k is the indicator index of domain j, m is the indicator number of domain j, A_i denotes the urban adaptivity of city cluster i, w_1j denotes the relative weight of domain j, w_2k denotes the relative weight of indicator k, and y_k denotes the normalized value of indicator k.

3. Results and Discussion

3.1. Analysis of Prediction Results

In terms of the training sets and prediction sets of social economics, we classify them into two categories in the following charts (as shown in Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8): the training sets are contained in the light green region and the prediction sets are shown in the darker green region. By collecting the carbon emissions data from 1997 to 2017 and predicting the following five years (as shown in Figure 4), we regard them as the input data of the RF algorithm. According to various input datasets, we set corresponding training years and the following five years as predicting years so as to work out the variation tendency and mechanism of them.

3.1.1. Social Economics

The GDP data acquired from 2010 to 2017 is supposed as the training set and that of the following five years is to be predicted (as shown in Figure 5). The overall developing trend of the six city clusters keeps increasing, while some city clusters show a different performance. Among which, the GDP of the MLCC and SFCC always remains at a relatively low level with certain variations, while an uncommon point lies in the boosting trend of SFCC that appears to drop a lot in the prediction sets. Meanwhile, although BTHCC and YRDCC lead the increasing trend and show a solid economic foundation, they both lowered development pace after 2017, which is quite different from that of the actual data as well. Considering the lack of input data during the process of training and the decrease of carbon emissions (https://figshare.com/collections/County-level_CO2_emissionss_and_sequestration_in_China, accessed on 21 November 2021) of prediction datasets, RF can solely output a relatively low level of the GDP. Such deviation originated from the shortcoming of input data, but can also be understood.

3.1.2. Environmental Pollution

The area of environmental protection incorporated three indicators: annual wastewater discharge, annual sulfur dioxide discharge, and annual smoke dust discharge (as shown in Figure 6). Overall, the discharge amount of pollutants tends to be more stable during the prediction period, with different city clusters performing variously towards each indicator.

As for wastewater discharge (as shown in Figure 6a), YRDCC always ranks first compared with its counterpart due to the developed industry in Yangtze River Delta that produces large quantities of sewage; on the contrary, wastewater discharge in the Shandong Peninsula stays as the lowest all the time.

All city clusters behave more complexly under the condition of sulfur dioxide discharge, which is shown in Figure 6b. The discharge amount of BTHCC tops all city clusters because of its substantial heavy industry plants, with YRDCC and MLCC following behind. What’s more, the developing trend of all clusters becomes more stable in the prediction datasets as well.

BTHCC and MLCC have the most discharge amount of smoke and dust compared to other city clusters (as shown in Figure 6c), which may be relative to the climate and energy structure of north China. Oppositely, the emissions amount in south China is nearly half of that in BTHCC and MLCC, which is completely different from the analysis result of wastewater discharge.

3.1.3. Climate Change

Indicators in climate change are concerned with temperature, precipitation, sunshine duration, and humidity. All the factors are selected to comprehensively represent a variation of the six city clusters in climate change and integrated to assess futural performance therein.

The annual mean temperature varies greatly among the city clusters, showing the geographical variability from north to south across China clearly (as shown in Figure 7a). Nevertheless, the variation of the highest and lowest temperature of the six clusters can be more obvious, especially for the aspect of the highest temperature. During the winter period, the lowest temperature (as shown in Figure 7d) keeps decreasing with the increase of latitude both in the training set and predicting set, and this indicator is always below zero in MLCC, BTHCC, SPCC, and YRDCC, with that of SFCC and PRDCC, remains higher than zero. While the highest temperature (as shown in Figure 7c) can be more similar among city clusters due to the overall high temperature across China when it comes to summer, nearly all city clusters have reached 40 °C in the recent 10 years except MLCC, and the highest value appears in various cities during the time series; also, the developing trend hasn’t shown apparent statistical logic, which is to be discussed.

Annual total precipitation (as shown in Figure 7b) has a similar developing trend with the annual highest temperature, which changes randomly both in the training set and predicting set, making it hard to predict accurately. However, this indicator has a distinct characteristic in view of the value itself, increasing as the latitude descends, and south China (mainly over 1500 mm) owns much more substantial rainfall than that of north China (mainly below 900 mm).

Sunshine duration is another indicator that ranges spatially because of the geographical variety among the six city clusters (as shown in Figure 7e). Generally speaking, north cities have longer sunshine duration hours that may last for 2400~2600 h yearly, while south cities mainly retain no more than 1600 h each year. The data in the prediction period fits with this trend well; thus, it is considered to be an acceptable result.

Correspondingly, relative humidity also relates to geographical characteristics (as shown in Figure 7f). This indicator decreases from the south (mainly over 75%) to north (mainly below 55%), which matches the distribution of total precipitation well because of the inherent connection between the two variables.

3.1.4. Marine Disasters

Tropical storms and sea-level increases are incorporated in the area of marine disasters. The frequency and intensity of tropical storms are shown in Figure 8a,b, which are mainly concerned with cities of south China, especially for the city clusters in Fujian Province (SFCC) and Guangdong Province (PRDCC). These city clusters nearly have three times of tropical storms per year with the average annual intensity between 4 and 5, while northern cities seldom experience tropical storms and the intensity is always limited within 1.

In terms of the annual variation of sea level, the random distribution of variation also appears both in the training set and predicting set; each city cluster has its positive and negative value under this condition. With the progress of time series, all city clusters tend to face apparent variations of sea levels, which imposes certain pressures upon city development.

3.1.5. Geographical Visualization of Prediction Results

Based on our statistical prediction results and analysis above, the aforementioned four evaluation domains that related with the development of city clusters are supposed to describe their comprehensive development and represent their characteristics from temporal and spatial dimensions. Figure 9 shows the geographical analyzing results for the four domains, and classifies all city clusters into six levels according to their performance on the basis of the training set and predicting set, in which the higher level (darker color) represents better performance of city clusters therein. Therefore, we’ll discuss the overall condition of the six city clusters with their geographical features so as to work out the coherent bond among them.

Figure 9a demonstrates the social economics layout of six city clusters, in which YRDCC, BTHCC, and PRDCC show their advantage over other regions. Considering the economic foundation of Shanghai, Jiangsu, Zhejiang, Beijing, and Guangdong, these provinces nearly lead Chinese development all the time, so this ranking result has a certain reference for regional planning.

Compared with other areas, SPCC shows the best performance in terms of pollutant indicators for environmental protection (as shown in Figure 9b); the discharge amount of wastewater, sulfur dioxide, and smoke dust in Shandong Peninsula are all significantly less than that of other city clusters (as shown in Figure 6). On the contrary, environmental pollutants in BTHCC and YRDCC are listed as the last two items of the ranking system, which is closely bound with the energy industry layout and economical level of these two city clusters. BTHCC hosts numerous steel plants, electricity power plants, and non-ferrous metal metallurgy plants with developed coal energy around Beijing, Tianjin, and Hebei; these plants produce substantial gaseous pollutants such as sulfur dioxide, nitrogen dioxide, nitrogen monoxide, carbon monoxide, and hydrogen sulfide, etc., which directly causes the prominent emissions of adverse gases in north China. YRDCC owns quantities of electronics plants, textile plants, and printing plants; these factories will discharge quite a lot of wastewater to treatment plants yearly. Meanwhile, the municipal sewage pipeline network is well-developed in east China due to the prosperous economy there, which assists to collect wastewater effectively, so YRDCC ranks highest in terms of sewage discharge (as shown in Figure 6a).

Among the four categories involved in climate change, i.e., temperature, precipitation, sunshine duration, and relative humidity, all city clusters show various performances that are related to their spatial location. PRDCC, located in the lowest latitude, and MLCC, located in the highest latitude, tend to react sharply in response to carbon emissions therein (as shown in Figure 9c). Considering their relative coastal location, the variation of temperature and precipitation plays a significant role during the process of climate change. For those cities in between, the situation of climate change can be milder especially for SPCC. The limited pollutants discharge obviously helps Shandong Province alleviate the climate change effect caused by carbon emissions.

Owing to their relatively high latitude, MLCC, BTHCC, and SPCC are less inclined to be affected by tropical storms, thus their marine disasters are apparently less frequent than that of southern city clusters (as shown in Figure 9d). Nevertheless, those cities located in the sub-tropical monsoon climate and tropical monsoon climate are more vulnerable and exposed under the risk of tropical storms and sea level, which leads to the high level of their eventual marine disasters.

3.2. Analysis of Evaluation Results

3.2.1. Urban Adaptivity of Predicting Set

Based on our description of the urban adaptivity evaluation system, the adaptivity of all the city clusters is calculated in the predicting period according to their indicators of the four domains, which is also visualized in Figure 10 to display their urban adaptivity level, respectively.

It is easy to draw the conclusion that MLCC and BTHCC show the best performance from 2018 to 2021 and SPCC keeps the lowest urban adaptivity consistently. We assume that the difference among northern city clusters is bound by their industry layout and economic level. Although SPCC discharges the least amount of pollutants, it is still not considered to be an ideal representative that balances the economy and environment due to its shortage in GDP. Meanwhile, southern city clusters commonly behave poorly which may be related to their exposure to marine disasters.

The urban adaptivity evaluation system aims to quantitatively describe the corresponding relation between a city’s prosperity and their pressure under the framework of climate change and environmental pollution; thus, it is regarded as a standard that represents this balance. With this system, we can further analyze the comprehensive performance of six city clusters based on their statistical data [31,32].

3.2.2. Urban Adaptivity in Average

The urban adaptivity of six city clusters is presented in

Table A1. Urban adaptivity of six city clusters from 2010 to 2021 (based on our demonstration of Urban adaptivity in Section 2.3 and Equation (6), the Urban adaptivity A_i can be better with a larger value, and these listed values reflect the city cluster’s adaptivity under a certain year).

Year	MLCC	BTHCC	SPCC	YRDCC	SFCC	PRDCC
2010	0.195	0.246	0.016	0.532	0.148	0.323
2011	0.520	0.582	−0.010	0.597	0.356	0.387
2012	0.570	0.519	−0.028	0.508	0.413	0.515
2013	0.704	0.537	0.243	0.359	0.299	0.397
2014	0.610	0.459	0.027	0.518	0.585	0.534
2015	0.504	0.374	0.119	0.401	0.651	0.378
2016	0.278	0.554	0.621	0.414	0.433	0.439
2017	0.339	0.621	−0.012	0.248	0.383	0.565
2018	0.417	0.632	0.013	0.546	0.342	0.427
2019	0.655	0.403	−0.012	0.453	0.603	0.435
2020	0.581	0.490	0.180	0.477	0.412	0.405
2021	0.523	0.632	0.041	0.538	0.519	0.503
Average	0.491	0.504	0.100	0.466	0.429	0.442

and their normalized values are showed in Figure 11. Meanwhile, Figure 12 describes the average of urban adaptivity and its ranking result, which contains the dataset between 2010 and 2021, so it reveals the comprehensive evaluation result of the six city clusters, among which BTHCC (0.504) and MLCC (0.491) lead the ranking result, which fits the reality to a certain extent. The complex problem between social development and environmental protection seems to find its ideal solution there.

Compared with the stated conclusions of this work and above studies, the methodology is supposed to be more significant, for we have constructed a comprehensive system of prediction and evaluation. Such a model has been applied in these city clusters of China, and it could also be studied in other regions. The method of utilizing the ML model and calculating urban adaptivity is to be modified and improved in further research, and it is expected to perform better.

3.3. Error Analysis

Due to lack of monitoring and statistical data, only GDP is selected as the validation indicator to analyze the error of prediction results for it may reflect the comprehensive performance of social and economic conditions of all city clusters. We use the data set between 2018 and 2020 as a validation set, and the Mean Relative Error (MRE), Root Mean Squared Error (RMSE), Mean Squared Error (MSE), and Mean Absolute Error (MAE) are chosen to demonstrate the performance of RF (as shown in

Table 2. Mean Relative Error of prediction results within six city clusters.

City Cluster	MLCC	BTHCC	SPCC	YRDCC	SFCC	PRDCC
MRE/%	8.20	13.08	40.29	26.89	73.29	42.64
RMSE/ trillion	0.04	0.18	0.34	0.39	0.33	0.43
MSE/ trillion	200,497.13	3,324,990.75	11,836,480.95	15,311,555.29	10,724,153.18	18,247,838.76
MAE/ trillion	0.04	0.18	0.34	0.37	0.33	0.43

). Due to the order of magnitude of GDP reaching several trillion (CNY ¥) for all city clusters of this work, the error analysis indicators can be prominent, while its results still provide references. Meanwhile, since the amount of GDP can reach several trillion for all city clusters, the error analysis results are thus still satisfactory.

The overall error analysis result is not as ideal as expected (economic boosting while carbon emissions cutdown, thus predicted error can be limited to certain level), especially for city clusters in south China, such as SFCC and PRDCC. For the predicting mechanism of the RF algorithm, it can only calculate based on statistical data, while the boosting trend of GDP with decreasing carbon emissions is opposite with its historical data, so the prediction result has an apparent relative error.

4. Conclusions

With consideration of the effect of carbon emissions on social economics, environmental pollution, climate change, and marine disasters, we put forward a novel predicting model with indicators of the above four domains based on carbon emissions and the Random Forest algorithm. Six coastal city clusters of China (MLCC, BTHCC, SPCC, YRDCC, SFCC, and PRDCC) are selected as the study area. By inputting the training set under certain time series, we obtain the prediction set of the following 5 years, thus evaluating the urban adaptivity of the six city clusters. With the above modeling and evaluation process, this paper draws the following conclusions:

(1) BTHCC and YRDCC show an obvious advantage over other city clusters in terms of social economics, while they also suffer heavy pollution in gaseous pollutants emissions and wastewater discharge, respectively. MLCC receives substantial annual pollutant loads as well, while it is less vulnerable to climate change and marine disasters due to its relatively high latitude. Similarly, SPCC also varies limitedly in temperature, precipitation, and sea level. SFCC and PRDCC are more likely to be influenced by climate change, which can be directly presented in the annual highest temperature and frequency of tropical storms, etc. Therefore, each city cluster has its own problem to be solved, which is necessary to attach more importance to local policy-makers.

(2) The characteristics of six city clusters are closely related to their geographical location and industry foundation, which is closely related with carbon emissions. The traditional economic dominant provinces within Yangtze River Delta and Pearl River Delta show their advantage over other areas. Meanwhile, quantities of heavy industry plants distributed in north China also cause substantial pollutants discharge therein, with city clusters near the South China Sea significantly affected by the monsoon yearly.

(3) Urban adaptivity considers the compatibility between social development and environmental protection. Therefore, we construct an urban adaptivity evaluation system that comprehensively considers all aspects of the above four domains. Northern city clusters perform better due to their lesser exposure to climate change and marine disasters, and southern cities are meant to focus more on the balance between economics and the environment. The assessment result may assist local policy-makers to design cities more properly in terms of industry structure and layout; also, it could help design circular economies for cities.