Southern China (18°10′–35°07′N; 97°21′–122°50′E) is selected as the study region in this research. The geographic location and terrains are shown in Figure 1. Most of southern China has a subtropical monsoon climate, with a mean annual precipitation and an annual pan evaporation of 1,320 mm and 1,545 mm, respectively, and an annual average maximum and minimum air temperature of approximately 23 and 14°C, respectively (Fan et al., 2016). Owing to these climate characteristics, this region is prone to HHTH, especially in the summer. Greenhouse-grown tomatoes in southern China are mainly grown in the winter half year; seedlings are raised in July to August and fruits can ripen in November. Therefore, HTHH mainly occurred in the seedling stage of greenhouse-grown tomatoes. Data from 305 meteorological stations located in the study region during 1990–2019 are collected. The collected parameters included daily outdoor maximum air temperature, minimum air temperature, average RH, minimum RH, sunshine duration, and average wind speed at 1.5 m above ground surface. These meteorological data are maintained and rigorously quality controlled by the China Meteorological Administration (http://data.cma.cn/).

For tomato production, the optimal range of RH is between 50% and 70%, and temperature exceeding 32°C will bring losses. Hence, in this study, we defined a daily maximum temperature of >32°C inside the plastic greenhouse as high temperature, and an average RH of >70% inside the plastic greenhouse is considered as high humidity. When high temperature and high humidity co-occurred and lasted for 3 days or more, it can be defined as a plastic greenhouse HTHH event.

There are no long-term historical microclimate data inside plastic greenhouses in southern China. It is practical and valuable to simulate microclimate inside greenhouses by using observations from outdoor meteorological stations for studying spatiotemporal characteristics of plastic greenhouse HTHH events at the regional scale. Machine learning methods are employed to simulate long-term daily maximum temperature and average RH in typical plastic greenhouses. Based on the existing literatures, we select six machine learning methods that are most commonly used in the field of meteorology. The selected models include multiple linear regression (MLR) (Lee et al., 2019); support vector machine (SVM) (Bayat et al., 2020; Fan et al., 2021), which is a kernel-based algorithm; two neural networks algorithms, back-propagation neural network (BP) (Sun, 2017) and extreme learning machine (ELM) (Feng et al., 2020); and two tree-based ensemble models, random forest (RF) (Jin et al., 2020) and extreme gradient boosting models (XGBoost) (Fan et al., 2021).

In order to build the simulation models, daily maximum temperature and average RH in three typical plastic greenhouses were observed. The locations of the greenhouses and the observation periods are shown in Table 1. After abandoning the abnormal observations, which accounted for 1.06% of the total observation, there are a total of 2,230 daily maximum temperature records and 1,502 daily average RH records (the observations from Liancheng were not used to build the RH simulation model, because of RH inside this greenhouse is mainly controlled artificially). The daily observations of nearby meteorological stations are used as input independent variables to simulate daily maximum temperature and average RH inside greenhouses. The sample data are divided into two subsets, the odd days are used to train model and the remaining are used to evaluate the model.

After several variable combination attempts, eight variables are selected to simulate daily maximum temperature inside the plastic greenhouse: outdoor maximum temperature of the day, outdoor average RH of the day, outdoor minimum RH of the day, outdoor sunshine duration of the day, outdoor maximum temperature of the previous day, outdoor maximum temperature of the next day, outdoor average RH of the next day, and outdoor minimum RH of the next day. Seven variables are selected to simulate the daily average RH inside the plastic greenhouse: day sequence, outdoor average RH of the day, outdoor minimum RH of the day, outdoor sunshine duration of the day, outdoor minimum RH of the previous day, outdoor average wind speed of the previous day, and outdoor minimum RH of the next day.

The performances of the six machine learning methods were evaluated using the coefficient of determination ( R 2 ) and root mean square error ( R M S E ). The mathematical expressions of the two statistical indicators can be found in a previous study (Dou and Yang, 2018). R 2 measures how well the model fits the data as a proportion of total variation. R M S E measures the stability of the model. Considering the two statistics together can provide a more comprehensive evaluation of the model’s performance. The lowest value of R M S E and R 2 value close to one represent the best model performance.

Response of tomato to HTHH stress is related to the temperature, humidity, and duration of the event. A comprehensive intensity index ( C S I ) of HTHH is built as: C S I = w t I t + w h I h + w d I d (1) where t ,   h ,   and   d are the element of temperature, humidity, and duration days of the HTHH event, respectively; w are the weights corresponding to each element, which are determined by the AHP method; and I are the normalized values of each element, calculated as: I i = V i − V m i n V m a x − V m i n (2) where V i is the original observation value of maximum temperature, average RH, or duration days of a certain HTHH event. V m a x and V m i n are the maximum and minimum value of this element among all the HTHH events in southern China during the past 30 years. I i is between 0 and 1. The purpose of normalization is to eliminate the unit and dimensional differences between elements, so as to make the calculation of C S I much more scientific.

The AHP method was first introduced by Saaty (1977; Saaty 1994). It is an effective decision-making tool to solve multiple-criteria problem, and can provide the weights of the parameters (Halil et al., 2013; Pramanik, 2016; Ren et al., 2019). In this study, the AHP method is used to determine the weights of the three elements describing the intensity of HTHH. The basic step of AHP is to establish the judgment matrices by a pairwise comparison method. The weights can be obtained by solving the judgement matrix. The fundamental scale for the pairwise comparison was suggested by Saaty (1994), as shown in Table 2. The pairwise matrix and weights of the three elements of HTHH are shown in Effects of HTHH Stress on the Physiological Indexes of Tomato Based on the Experiment section.

Seedlings of tomato variety, Lycopersicon esculentum Mill. Shouhe Fenguan, were planted in a Venlo-type greenhouse in Nanjing University of Information Science and Technology from April to July 2020. At the stage of fourth true leaf, seedlings were transferred to an artificial climate box (TPG1260, Australia) and exposed to the HTHH stress. A three-factor orthogonal experimental design was conducted. The factor of high temperature (daily maximum/minimum temperature) was set as four levels: 32/22°C ( T 1 ), 35/25°C ( T 2 ), 38/28°C ( T 3 ), and 41/31°C ( T 4 ); average RH was set as follows: 50% ( R H 1 ), 70% ( R H 2 ), and 90% ( R H 3 ); the factor of stress duration days was set as follows: 2 days ( D 1 ), 4 days ( D 2 ), 6 days ( D 3 ), and 8 days ( D 4 ). The control treatment (CK) was set as 28/18°C and 50 ± 5%. From all the treatment combinations, 16 treatments were selected by an orthogonal table as shown in Table 3.

Days of flower bud differentiation (DD), Chl content of leaves, ABA content of terminal buds, and seedling quality index (SQI) were measured for each treatment to show the responses of plants to HTHH stress. The date when the apical meristem became slightly visible and small flat bumps began to form is recorded as the beginning of flower bud differentiation, and the date when flower organs became fully developed is regarded as the end of the flower bud differentiation. An Olympus electron microscope was used to observe the flower bud differentiation process. SQI, Chl, and ABA content were measured when flower bud differentiation ended. SQI was calculated as: (diameter of the main stem/plant height) × dry mass of whole plant (Li et al., 2020). Duncan’s test (p ≤ 0.05) was used to assess differences in physiological indices under different treatments.

To make the four physiological indexes comparable, change rates were calculated as Eqs. 3, 4. The physiological indexes that increased under abiotic stress were calculated with Eq. 3, such as ABA and DD. The physiological indexes that decreased under abiotic stress were calculated with Eq. 4, such as SQI and Chl. C r P I i = ( P I i − P I c k ) / P I c k (3) C r P I i = ( P I c k − P I i ) / P I c k (4) where P I i is the observed value of physiological index P I under i treatment, and P I c k is the value of P I in the CK; C r P I i represents change rate of P I i .

Six machine learning methods were used to simulate daily maximum temperature and average RH in typical plastic greenhouses. The performances of the machine learning methods are shown in Table 4. The results showed that for both daily average RH and maximum temperature, the ELM model had the greatest R 2 and smallest R S M E in the testing stage, indicating the best model performance among all methods. The observed and simulated values of daily average RH and daily maximum temperature by using the ELM model are further shown in Figure 2, and the values of regression coefficient were close to the 1:1 line. Therefore, in this study, the ELM method was selected to simulate the daily average RH and maximum temperature in plastic greenhouses in southern China during 1990 and 2019 by using outdoor meteorological station observations.

HTHH events in plastic greenhouses were extracted in southern China during 1990 and 2019 according to the definition in the Definition of Plastic Greenhouse HTHH Events. The spatial distribution of multi-year average occurrence times, average duration days, and annual accumulative days of HTHH events were analyzed. As shown in Figure 3A, the annual occurrence times of HTHH events showed the largest value in the southernmost areas of the study region (often annually >10 times). Average duration days of HTHH events was high along the eastern and southern coasts with the events often lasting over 6 days (Figure 3B). Annual accumulative HTHH days were highest in the southern area (Figure 3C), where the HTHH events occurred frequently and lasted for longer days. However, in the western area of the study region, HTHH events seldom happened.

The variation of plastic greenhouse HTHH events in different months is shown in Figure 4A. The HTHH events mainly occurred from June to September in southern China, accounting for 86.93% of the total occurrence times of HTHH events, and 89.27% of the total accumulative days. From June to September, the average monthly occurrence times was ∼1.2 and the average duration days was ∼6 days. The HTHH events rarely occurred in October to May of the next year. The interannual variation of the HTHH events during 1990 and 2019 is shown in Figure 4B. The annual occurrence times and the annual accumulative days of the HTHH events showed a significant downward trend with 0.37 times/decade (p < 0.01) and 2.63 days/decade (p < 0.05), respectively, during this period. In order to find out which months caused this decreasing trend, the interannual changes of HTHH events in different months were analyzed. We found that the occurrence times and accumulative days of the HTHH events in July had a significant downward trend, while there was no obvious interannual trend in other months. As shown in Figure 4B, the occurrence times and the accumulative days of the HTHH events in July decreased significantly (p < 0.05) by 0.1 times/decade and 1.3 days/decade, respectively.

The physiological indexes of tomatoes after flower bud differentiation were measured for each treatment, as shown in Figure 5. The physiological indexes had significant (p < 0.05) differences under different high-temperature treatments. High temperature significantly increased the values of D D and A B A , but significantly decreased C h l and S Q I . Compared with the CK, the HTHH stress with any level of RH or duration days significantly changed the physiological indexes. However, among different levels of RH, there is no significant difference for D D , C h l , and A B A . S Q I of treatments under high RH (90%) stress is significantly smaller than 50% RH treatments. Among different duration day treatments, there is no significant difference for D D and A B A . The HTHH stress with longer duration days can significantly decrease S Q I and C h l compared with shorter duration days. In general, high temperature plays a more important role than RH and duration days in the effects of HTHH stress on tomato growth.

According to the responses of physiological indexes of tomato under different HTHH stress treatments and after consulting some experts, we build the pairwise comparison matrix. Then, the weight of each element of HTHH event based on the AHP method can be calculated. The results are shown in Table 5. The weight of high temperature, average RH, and duration days are 0.579, 0.234, and 0.187, respectively. Based on Eq. 1, the C S I of each HTHH treatment can be calculated. Logistical functions are used to fit the changes of physiological indexes with the C S I , as shown in Figure 6. The goodness of fit of all the curves are at the 0.01 significance level, and the curve of ABA performs better than other indexes with an R ² of 0.88. With the increased C S I , the change rate of DD is higher than other physiological indexes. When the C S I value is ∼0.2, the change rate exceeds 1.0 for DD, 0.4 for C h l , but no more than 0.2 for S Q I , indicating that the DD is the most sensitive physiological index to HTHH stress, followed by C h l .

HTHH events are recognized at every station from 1990 to 2019, and C S I of these HTHH events was calculated by Eq. 1. The exceeding probability of C S I for each station is fitted so as to obtain the C S I of HTHH events at different return periods. Fitting equations in Figure 6 can be used to calculate potential loss rates of physiological indexes of tomatoes caused by the HTHH events at different return periods. Figure 7 takes ABA and Chl as examples to show the spatial distributions of potential loss rates of tomato under the HTHH events for the 0.33-year, 1-year, and 5-year return periods. As the return period increased, the corresponding loss rates also increased. The loss rates of Chl are significantly higher than ABA at the same return period except for the western region. This can be attributed to the fact that Chl is more sensitive to HTHH than ABA (Figure 6). In these areas, a once-a-year HTHH event can cause Chl losses of more than 40%. However, the western region is a mountainous area where the HTHH events rarely happened (Figure 3). Hence, the intensity of the HTHH events did not change much even at higher return periods; the corresponding loss rates were all below 10%.

Outdoor daily maximum and minimum temperature, RH, sunshine duration, and wind speed are the essential elements to simulate the maximum temperature and RH inside plastic greenhouses. Against the backdrop of global warming, there were significant upward trends for the daily maximum temperature and daily minimum temperature across various climatic zones of China (Fan et al., 2016). A downward trend for daily RH was found during 1956–2015 in China; this downward trend reached significance at −0.124%/year during 1986–2015, and southern China showed the greatest decline magnitude (Fan et al., 2016). However, Yan et al. (2019) found that the high-temperature and high-humidity events have increased since the mid-1980s in the middle and lower reaches of Yangtze River, in the northern part of southern China. Wind speed decreased in the 1990s and increased after 2000 in southern China (Fan et al., 2016). Since the 1990s, the sunshine duration have increased in southern China (Fan et al., 2016; He et al., 2018). Previous studies were more concerned about univariate meteorological elements; high-temperature and high-humidity events were less researched. Western north Pacific subtropical high westward extension and intensification during the past decades plays an important role in regulating China’s climate in summer (He et al., 2015; Wang et al., 2014). In this study, it was found that there was a decreasing trend of the interannual variation of the HTHH events in plastic greenhouses, especially in summer. This is different from the result of increased outdoor HTHH events obtained by Yan et al. (2019). This may have something to do with increased sunshine duration and wind speeds, which prevents the RH in plastic greenhouse from being too high.

Previous studies on the environmental constraints for facility crops mainly focus on a single factor, such as high temperature, low temperature, scant light, and high humidity. However, these environmental constraints often co-occurred in actual production, such as low temperature and scant light, high temperature, and high humidity. However, these composite stresses are seldom studied. Heat stress research trials were carried out in many previous studies, and showed that high temperature had seriously adverse impacts on the growth, fruit yield, and quality of tomatoes (Mulholland et al., 2003; Shamshiri et al., 2018; Ro et al., 2021). Hence, identification of tomato genotypes with higher resistance at high temperatures continues to draw the attention of researchers (Xu et al., 2017; Akhoundnejad et al., 2020; Vijayakumar et al., 2021). These studies found that the effects of high temperature on heat-resistant tomato genotypes were commonly alleviated by more than 10% compared to that on heat-sensitive genotypes. Therefore, selection of heat-resistant tomato varieties can be an effective method to mitigate the adverse effect of high temperature. Compared with high temperature, studies on high humidity are relatively less reported. RH affects plants mainly though regulating stomata. Arve and Torre (2015) reported that high RH slightly decreased ABA, which is consistent with the results in this study. In addition, with regard to the effect of HTHH stress on tomato growth, we found that high temperature plays a more important role than high RH and duration days. Since high temperature and high humidity often occur simultaneously and will cause more serious effects, the performance of high-temperature-resistant genotypes under high-humidity conditions should be further assessed in a future study.

Here, we found that the law of physiological indexes of tomato changed with comprehensive intensity of HTHH, and the goodness of fit of all the curves was at the 0.01 significance level (Figure 6). In many previous studies, this kind of relationship was commonly called vulnerability curve, which can be used to assess the sensitivity and adaptability of a hazard-bearing body to adverse effects (Yue et al., 2015; Wang et al., 2016; Cui et al., 2019; Li et al., 2021). It is also an important part of disaster risk assessment. Most of these vulnerability curve studies were for field crops, such as soybean, wheat, and maize. However, facility agriculture was rarely involved. Hence, the curves obtained in this study can be potentially used for disaster risk assessment in the facility agriculture, and provide essential information for guiding the layout of plastic greenhouses and early prediction of tomato losses when suffering from HTHH stress. However, due to the differences in variety, irrigation, and fertilizer usage, among others, the response of tomatoes to HTHH stress in actual greenhouse production may differ, to a certain extent, from the estimated vulnerability curves derived in this study. In addition, plastic greenhouses in practical production have different qualities and structures, so the relationship between the microclimate inside plastic greenhouses and external meteorological conditions may be different for different plastic greenhouses. This increased the uncertainty of our results. In future research, collecting microclimate observations from a larger number of and more dispersed plastic greenhouses can reduce the uncertainty.

Adverse microclimate conditions in greenhouses can lead to crop production losses and increase regulation costs. In this study, southern China was chosen as the case study area to investigate the spatiotemporal pattern of HTHH events in plastic greenhouses. In addition, physiological responses of tomato to HTHH stress were explored, so as to quantitatively identify the spatial distribution of tomato potential losses under HTHH stress in southern China. Major conclusions are as follows:

1) Among the machine learning methods, the ELM method performed best to simulate the daily average RH and maximum temperature in typical plastic greenhouses by using outdoor meteorological observations.

The results of this study are significant in guiding the layout of plastic greenhouses based on local climate resources. Although this research focused on greenhouse-grown tomato and HTHH, the approaches and technical means in this study can also be further applied to other greenhouse meteorological disasters and crops.