Characteristic Identification of Heat Exposure Based on Disaster Events for Single-Season Rice along the Middle and Lower Reaches of the Yangtze River, China

Abstract

As global warming continues, heat stress events are expected to increase and negatively affect rice production. Spatiotemporal changes in single-season rice exposure to heat stress were explored along the middle and lower reaches of the Yangtze River (MLRYR) in China during 1971–2020 based on created heat thresholds in different phenological stages, derived from comparison of historical heat records for single-season rice and agro-meteorological data. The feature importance (IF) provided by the Random Forest model was used to modulate the relationship between threshold accumulated temperature and yield reduction rate caused by heat stress. In addition, critical temperature thresholds at different phenological stages were determined by combining Overall Accuracy and the Receiver Operating Characteristic (ROC) curve. According to historical disaster records, the heat stress occurred before the reproductive phase (i.e., the tillering–jointing stage) and ended in the filling stage. Critical temperature thresholds of T_max at tillering–jointing, booting, flowering and filling were quantified as 36, 35, 35 and 38 °C, with higher IF values of 13.14, 10.93, 17.15 and 13.15, respectively. The respective values of Overall Accuracy and the areas under the ROC curve were greater than 0.85 and 0.930, implying that each threshold performed excellently in identifying heat occurrence. Based on the determined critical thresholds, accumulated harmful temperature (Tcum), number of heat days (HD), first heat date (FHD) and last heat date (LHD) were presented to characterize heat exposure. It was clear that Tcum and HD exhibited a north-to-south increasing trend from 1971 to 2020, with the obvious increasing occurrence in most parts of the study region through the period of 2010 to 2020. FHD occurred earlier in most stations except the northeast parts, while LHD ended later in southern MLRYR. Exploring heat critical thresholds at different phenological stages highlighted in this study can help decision-makers monitor and evaluate heat exposure to single-season rice in MLRYR and further develop mitigation strategies to ensure rice production security.

1. Introduction

Rice (Oryza sativa L.) is a crucial cereal grain worldwide, serving as a primary source of sustenance for over 50% of the world’s population. Around 160 million hectares of agricultural land are dedicated to rice cultivation worldwide [1]. Experts estimate that the production of rice and other crops must increase by over 70% to meet the dietary demands of the world’s rapidly growing population by the year 2050. This increase in production is necessary to prevent food shortages and crises [2,3,4]. However, rice grown in the tropics and subtropics may be potentially at a greater risk of heat stress. Heat stress is recognized as an essential threat to rice cultivation in China [5], Sudan, Laos, India [6,7], etc., with increased frequency and severity in extreme temperature events induced by global warming across the major rice planting areas [8,9]. In China, there have been numerous instances of extreme heat events for rice from June to August after the annual plum rain season, specifically around the middle and lower reaches of the Yangtze River (MLRYR). The severe economic loss and grain reduction caused by heat disasters have occurred multiple times in MLRYR. For instance, heat events in 2003 contributed to a 5.18 million tons reduction in the predicted grain harvest and more than USD 1.5 billion in economic loss [10,11]. Following global warming, the frequency and intensity of extreme heat events have substantially increased in major parts of the world [12], posing challenges to rice production [13]. In MLRYR, the occurrence of rice heat stress has fluctuated markedly from year to year, with an overall increasing trend in the occurrence frequency and intensity [14]. Hence, it is essential to understand how rice responds to heat exposure in MLRYR, which can guide the implementation of targeted measures to reduce economic and production losses.

Heat stress is generally acknowledged as irreversible damage to spikelet fertility and yields during the growing season when elevated temperatures are beyond the expected threshold [15], as a consequence of spikelet sterility, reduced kernel weights, and a shorter grain-filling stage [16,17,18]. Under the constant temperature of 38 to 40 °C, the number of unfertilized grains is significantly increased and eventually depressed total grain yield [4,19,20]. Based on high-temperature controlling experiments, empirical statistical approaches or crop simulation models, several studies [21,22,23,24] have explored rice’s response to heat stress during the reproductive phases, e.g., the booting, flowering and filling stages [25]. Sanchez et al. [26] reviewed previous studies and concluded that rice yield was most sensitive to high temperature before flowering throughout the whole reproductive phase, followed by the grain-filling stage, which was contradictory with the result derived from the WOFOST crop model [27]. The current research on the effects of rice heat stress has generated some controversy regarding the varying sensitivities of rice growth stages. This lack of a precise explanation makes it challenging to assess the degrees of damage caused by unique heat events on rice during different growth stages. Such an assessment is significant in evaluating the intensity of rice heat stress exposure.

To accurately gauge the intensity of heat stress on rice, two basic aspects must be considered. First is the heat stress event, including the intensity, duration, and degree of elevated temperature [28]. Second is the rice’s resistance to heat stress, which varies depending on its growth stage [25]. It is widely accepted that the heat stress process is defined according to the threshold of average temperature (T_ave) at 30 °C or maximum temperature (T_max) at 35 °C for more than 3 consecutive days throughout the reproductive phase [5,9,14,24,25]. To some extent, it is difficult to accurately evaluate the state of each unique heat process, neglecting the difference in the disaster-causing degree. Although these studies can better explain the yield variation corresponding to the increase in temperature, the intensity of heat stress is often exaggerated or underestimated due to the lack of consideration of differences in the heat tolerance of rice at different growth stages. Rapid climate warming causes earlier heat stress occurrences, such as identifying heat stress at the vegetative phase (e.g., the seeding, tillering, and jointing stage), which was generally not cited previously [5,22,25,26]. Subjected to heat stress in the vegetative phase, rice is likely to be exposed to scorching senescence of leaves followed by abscission, and reduced plant height, tiller number, and biomass [4,28], especially tiller number per plant determines panicle number [17]. To effectively manage heat stress on rice crops, it is essential to integrate the identification of specific phenological stages with severe heat hits. The same temperature threshold was generally utilized only to reflect a certain degree of impact on rice growth but not necessarily involve different impacts in different phenological stages. Therefore, the altered heat resistance of rice should be identified during the dynamic growth.

Agro-disaster representations integrating disaster records and meteorological and crop phenology data [25] can be explored to reveal the relationship between meteorological factors and crop yields due to the disaster process and to redefine optimal triggering thresholds at different physiological stages. Moreover, whether rice exposure to heat stress only occurs at the reproductive phase and subsequent results from sensitivity to heat stress are still controversial and need to be deeply explored. In this paper, we used historical rice heat-disaster representations in MLRYR to determine the triggering threshold of heat stress occurrence and characterize single-season rice exposure to heat stress. The main objectives are: (1) to investigate the key phenological stages of rice exposure to heat stress based on the historical disaster data; (2) to quantify the critical temperature threshold for heat stress to single-season rice at different phenological stages; and (3) to delineate the characteristics of heat exposure, which can help us to develop strategies for adaptation to heat stress in MLRYR.

2. Material and Methods

2.1. Study Region

The study region spans across Jiangsu (JS), Zhejiang (ZJ), Anhui (AH), Jiangxi (JX), Hubei (HB) and Hunan (HN) Province, situated in MLRYR ranging between 24–35° N latitude and 109–123° E longitude (Figure 1). Under the control of a subtropical monsoon climate, the MLRYR is characterized by hot and rainy summers and cold and dry winters. Annual mean temperature ranges from 14 to 20 °C. Average annual precipitation is approximately 1000–1400 mm, with 40–60% occurring from June to August. The dominant rice cropping systems are double-cropped and single-cropped systems, including three types of rice (double-season early, double-season late and single-season rice). With the adjustment of planting structure and the shortage of workforce, the planting area of single-season rice in MLRYR has significantly increased in the past decades [29], accounting for approximately 73% of the rice cultivated area by 2020 [30]. The transplant stage of single-season rice in this region is typically in May and harvest in mid-to-late September. The phenological stages of single-season rice are consistent with the temporal distribution of hot weather in MLRYR, and overwhelming heat stress events occur in the reproductive phase, as has been demonstrated previously [20,31,32].

2.2. Data Description

Datasets including meteorological, phenology, yield, and disaster data were used to construct heat-disaster samples (

Table 1. Information on meteorological, phenology and disaster data used in this study.

Classification	Data Index	Source	Period	Area Scope
Meteorological	Daily maximum temperature	National Meteorological Information Center	1971–2020	501 stations
Phenology	Seeding, tillering, jointing, booting, flowering and filling date of single-season rice	Agro-meteorological experimental station	1981–2010	501 meteorological stations interpolated from phenology data in 34 stations
Yield	Yield reduction rate (YRR) of single-season rice	National Bureau of Statistics	1971–2020	Six provinces (JS, ZJ, AH, JX, HB and HN)
Disaster	Records of heat stress process of single-season rice	《China Meteorological Disasters Book》, 《Yearbook of Meteorological Disasters in China》	1971–2000, 2004–2019	Six provinces (JS, ZJ, AH, JX, HB and HN)

). Meteorological data at 501 meteorological stations, including daily maximum temperature (T_max) from 1971 to 2020, were used to explore the critical temperature threshold for characterizing the occurrence of heat stress to single-season rice, downloaded from the National Meteorological Information Center, China Meteorological Administration (Figure 1). Several missing values of meteorological data were replaced by the average value in the adjacent two days. Phenology data include average seeding, tillering, jointing, booting, flowering, and filling dates of single-season rice during 1981–2010, and were obtained from 34 agro-meteorological experimental stations of six provinces. The Kriging method was applied to interpolate phenology data to 501 stations. Single-season rice begins to flower in the last ten days of July in eastern MLRYR, while rice beginning to flower is delayed until early August in the western areas. The yield data of single-season rice at the provincial level has been archived by the National Bureau of Statistics (https://data.stats.gov.cn, accessed on 21 August 2022 ) from 1971 to 2020. The yield reduction rates (YRRs) were calculated by separating the trend yield fitted to adopting the method of the linear moving average. Historical heat events of single-season rice, including the time, location, and disaster degree, can be recorded in the Yearbook of Meteorological Disasters in China [33] and the China Meteorological Disasters Book (JS, ZJ, AH, JX, HB and HN) [34]. The aforementioned data were used to construct heat-disaster samples, facilitating the construction of the optimal caused-disaster thresholds for different growth phases. Meteorological data and phenology data were used for calculating accumulated harmful temperature (Tcum), number of heat days (HD), first heat date (FHD) and last heat date (LHD) to evaluate single-season rice heat stress exposure in MLRYR.

2.3. The Construction of Rice Heat Stress-Disaster Samples

Documentary evidence of historical agro-disaster events can provide disaster information for relevant studies [25,35,36]. Destruction evidence and methodology data can be coupled to explore the disaster weather conditions that can trigger crop damage [25,37]. The critical values of catastrophe characteristics can be obtained by comparing the difference in weather conditions between the disaster process and a historical non-disaster process. Using historical data on rice heat stress events and meteorological conditions, the exact phenological stages and thermal conditions were screened out to trigger heat occurrence in single-season rice in MLRYR. Then, T_max was extracted to construct heat (H) and non-heat (NH) databases according to the historical disaster representations and exact phenological stages, integrating growth stage, location, occurrence time, and final time to identify optimally the critical threshold of heat stress for single-season rice.

2.3.1. Target Phenological Stages of Rice Heat Stress

Due to the unique combination of crop phenology and local climate, the thermal-sensitive period to heat stress likely differed between regions [9,38]. The national standard of the People’s Republic of China [39] stipulated that booting, flowering, and filling were the crucial stages in the heat stress process for single-season rice, as extensively studied in the literature [5,24]. Out of our historical heat representations of single-season rice, the earliest heat record was set on June 11 in JS province (early tillering stage), while the last instance of rice heat stress at the latest was recorded on 9 September in ZJ province (late filling stage). Furthermore, regardless of the rice characteristics and climate warming trends, the investigated phenological stages focused on the tillering–jointing, booting, flowering and filling stages, broadly grouped into one vegetative and three reproductive stages.

2.3.2. Construction of Heat (H) and Non-Heat (NH) Databases

Variations in the shape of the temperature distribution played a more crucial role, with notable damage from extreme temperature rather than the mean values. For instance, even a few hours of peak high temperature during the critical flowering or reproductive phase could drastically cause physiological damage and lead to vital crop failure, independently of any substantial changes in the mean state [14,22,40]. Given extensive cognition and high sensitivity to extreme temperatures on rice plants, T_max was better developed as a critical temperature threshold for triggering single-season rice heat stress rather than the mean temperature in this study.

According to historical heat records for single-season rice in MLRYR, disaster samples, including the T_max, growth stage, location, occurrence time and final time for each heat process, were integrated into the preliminary databases separately. Since our research hypothesis for this paper was that there were different heat stress thresholds in each phenological stage, historical heat-disaster samples screened from the preliminary database that only heat stress affects a single growth stage were reintegrated into the new databases, denoted as the H database here. All T_max values from each heat sample were considered thermal conditions for single-season rice to demonstrate the altered status during the temporal span of a heat process/event. To better identify rice heat stress, the NH database was constructed by a collection of T_max values over the previous 5 days based on disaster records, which represented the temperature conditions that did not trigger rice heat stress. Additionally, we used clear heat-disaster records covering multiple phenological stages as application validation samples to verify the availability of critical temperature thresholds for single-season rice in MLRYR.

For instance, consider the following disaster record: “20 July–2 August 2016, ear differentiation and inflorescence emergence of single-season rice in Hefei (AH province) were affected by high-temperature weather, and rice was damaged with yield decreased”. Following the representation of meteorological data, the T_max values of the corresponding meteorological station and temporal span at Hefei were 33, 34.3, 36, 37.4, 37.7, 38.2, 39.1, 38.5, 37.8, 38.2, 38.7, 37.7, 35.7, and 35 °C, also labeled as a flowering sample in H database. Meanwhile, the T_max values 15–19 July (the prior 5 days) and 3–7 August (the previous 5 days) of 30.8, 29.7, 32.8, 33.5, 28.5 °C and 31.5, 32.5, 31.9, 33.6, 32.2 °C, respectively, were also retrieved and were labeled as two flowering samples of single-season rice in the NH database. Finally, the H database collated 122 tillering–jointing samples, 108 booting samples, 148 flowering samples, and 237 filling samples, as well as the NH database including 144 tillering–jointing samples, 216 booting samples, 296 flowering samples, and 474 filling samples.

2.4. Critical Threshold Identification for Rice Heat Stress

There were complex nonlinear relationships between crop growth and meteorological conditions that were not straightforward. While traditional Pearson correlation analysis typically determined the relationship between two variables, it fell short in measuring nonlinear relationships [41]. As a widely used machine learning decision tree model, Random Forest can handle the nonlinear relationship among the variables by utilizing a non-parametric algorithm. The IF is the importance score obtained from model training that can rank the relative importance of each predictor variable in controlling the response variable [41,42]. To determine different threshold temperatures that damage the yield under heat exposure, the Random Forest model was applied in this study to explore the first three key threshold temperatures highly correlated with YRR, which were preliminary screened out by the higher IF value. Then, the Overall Accuracy and Receiver Operating Characteristic (ROC) curves were often used to provide accuracy between the constructed H and NH databases in classification algorithms and were combined to identify optimum critical thresholds for different growth stages on the principle of high IF values and high accuracy.

2.4.1. The Importance Score of Random Forest (IF)

Random Forest was used to establish the model between accumulated temperature calculated by different threshold values and YRR of single-season rice. The IF is based on the out-of-bag (OOB) regression prediction error. The variable importance is computed by mean square error ($MS{E}_{OOB}$):

$$MS{E}_{OOB}=\frac{1}{n}{{\displaystyle \sum }}_{k=1}^{nn}{\left(D_k-D_{OOB,k}\right)}^2$$

(1)

where $nn$ is the number of observations, $D_{OOB,k}$ is the average of all OOB predictions across all trees in Random Forest. IF is estimated using the “%IncMSE” metric in the Random Forest model. This process was implemented using the package “random forest” in Python 3.9.

2.4.2. Overall Accuracy

Overall Accuracy is measured as how accurately T_max can trigger rice heat stress and indicates the proportion of correct retrievals for both heat occurrences and non-occurrences.

$$OverallAccuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

(2)

Here, true positives ($TP$) are the number of rice heat stress instances in the H database that are retrieved correctly; false positives ($FP$) are the number of non-heat samples in the NH database that are retrieved incorrectly; true negatives ($TN$) are the number of non-heat samples in the NH database that are retrieved correctly and false negatives ($FN$) are the number of rice heat stress instances in the H database that are retrieved incorrectly.

2.4.3. ROC Curve

The ROC curve is a graphical representation used to display the diagnostic ability of binary classifiers and obtain a more comprehensive accuracy with variation in the diagnostic threshold, as well as the change in true and false positive fractions. It is widely used in various fields of study, including medicine, radiology, natural hazards, and machine learning [43]. The ROC curve is constructed by plotting multiple pairs of true positive rate ($TPR$) values against false positive rate ($FPR$) values. $TPR$ represents the proportion of observations that are correctly (truly) retrieved to be positive out of all positive observations ($TP+FN$). Similarly, $FPR$ is the proportion of observations that are incorrectly (falsely) retrieved to be positive out of all negative observations ($TN+FP$).

$$TPR\left(sensitivity\right)=\frac{TP}{TP+FN}$$

(3)

$$FPR\left(1-sensitivity\right)=\frac{FP}{TN+FP}$$

(4)

With $FPR$ as the x-coordinate and $TPR$ as the y-coordinate, the ROC curve is drawn and shows the trade-off between sensitivity ($TPR$) and specificity ($1-FPR$) by changing the diagnostic threshold. The ROC curve was used to evaluate how well the predictor (T_max) could identify whether rice heat stress would occur or not. The AUC value, which represents the area under the ROC curve, provides an estimate of the overall performance of T_max. The evaluation standard is AUC < 0.6 (Fail), 0.6 ≤ AUC < 0.7 (Poor), 0.7 ≤ AUC < 0.8 (Fair), 0.8 ≤ AUC < 0.9 (Good), 0.9 ≤ AUC ≤ 1 (Excellent).

2.5. Characteristics of Rice Heat Stress

2.5.1. Accumulated Harmful Temperature (Tcum) and Number of Heat Days (HD)

Tcum is the sum of temperature differences that exceed the critical threshold, while HD is the number of rice heat stress days at different phenology per station from 1971 to 2020. Tcum and HD can effectively assess the frequency and intensity of heat exposure to single-season rice under observed contemporary climate conditions in MLRYR. The Mann–Kendall (MK) test, a rank-based nonparametric method, detected statistically significant trends and mutation time of heat exposure in rice over a 50-year period. Its advantage is that a nonparametric test is not disturbed by outliers, and sample data does not need to follow a certain distribution. In this paper, the null hypothesis (H₀) was that there has been no trend in heat exposure over time; the alternate hypothesis (H₁) was that there has been a trend (increasing or decreasing) over time. The mathematical equations for calculating MK Statistics $S$, $E\left(S\right)$, $var\left(S\right)$ are as follows:

$$S={{\displaystyle \sum }}_{i=1}^{n-1}{{\displaystyle \sum }}_{j=i+1}^{n}sig(X_j-X_i)$$

(5)

$$sig\left(X_j-X_i\right)=\left\{\begin{array}{c}+1 if \left(X_j-X_i\right)>0\\ 0 if \left(X_j-X_i\right)\le 0\end{array}\right.$$

(6)

Under the assumption of random independence of the time series, the standardized test statistic $UF$ is defined as follows:

$$UF=\frac{\left[S-E\left(S\right)\right]}{\sqrt{var\left(S\right)}}$$

(7)

$$\left\{\begin{array}{c}E\left(S\right)=\frac{n\left(n-1\right)}{4} \\ var\left(S\right)=\frac{n\left(n-1\right)\left(2n+5\right)}{72}\end{array}\right.$$

(8)

Then, repeat the $S$ and $sig\left(X_j-X_i\right)$ procedures with inverse time series to define the statistic $UB$:

$$UB=-\frac{\left[S-E\left(S\right)\right]}{\sqrt{var\left(S\right)}}$$

(9)

In these equations, $X_j$ and $X_i$ are the time series observations in chronological order, $n$ is the length of the time series, $S$ is the cumulative number of time $i$ values greater than time $j$ values, and $E\left(S\right)$ and $var\left(S\right)$ are, respectively, the mean and variance of $S$. Positive $UF$ values indicate an upward trend in the time series; negative $UB$ values indicate a negative trend. If $\left|UF\right|>U{F}_{1-\alpha /2}$, (H₀) is rejected, and a statistically significant trend exists in the time series of heat exposure. The critical value of $U{F}_{1-\alpha /2}$ for a significance value of 0.05 from the standard normal table is 1.96. Similarly, if $\left|UB\right|$>1.96, there is a statistically significant trend in the inverse time series. The intersection of $UF$ and $UB$ curves occurs in the range [−1.96, 1.96], then the intersection is when the mutation starts, the mutation point.

2.5.2. First Heat Date (FHD) and Last Heat Date (LHD)

The formation of rice heat stress is a dynamic process of gradual accumulation [44], from the temperature increasing gradually to the critical threshold of normal crop growth and the continuous high temperature until rice is damaged to varying degrees. Generally speaking, most domestic studies consider a heat process where the maximum air temperature exceeds the critical threshold for at least 3 days. So, the date when the first heat process occurred was regarded as FHD. Similarly, the end date of the last heat process was considered to be LHD.

3. Result

3.1. Characteristics of the Heat (H) and Non-Heat (NH) Databases

Clear differences in T_max values showed significantly higher T_max values in the H database than in the NH database at similar phenological stages (

Table 2. Statistics of T_max values in H and NH databases for single-season rice.

Phenology	Database Type	Minimum	Maximum	Mean	Standard Deviation
Tillering–jointing	H	23	40.8	33.69	3.65
	NH	22.2	37.4	30.30	3.08
Booting,	H	28.1	41.7	36.31	1.98
	NH	21.9	38.3	31.62	3.32
Flowering	H	24.8	42.4	36.43	2.25
	NH	21.4	36.8	30.68	3.41
Filling	H	23.3	43.2	36.82	2.68
	NH	19.7	39.9	33.15	3.60

). Take the mean value as an example; the difference between H and NH databases at tillering–jointing, booting, flowering and filling stages were 3.39, 4.69, 5.75 and 3.67 °C, respectively. The mean values of T_max in the H database showed an increasing trend with the advancement of the growth period and were 33.69, 36.31, 36.43 and 36.82 °C, respectively. The violin plots in Figure 2 showed the detailed distribution of T_max values in H and NH databases for single-season rice. It could be seen that most values in the H database were distributed ranging from 30 °C to 40 °C, accounting for nearly 90% of all T_max values. From the H databases of different phenological stages, T_max from 36 to 37 °C was detected as having the highest frequency of single-season rice heat stress in the tillering–jointing database, accounting for 13.23% of T_max values detected in this range. The highest frequencies at booting, flowering and filling were located in the range 35~36 °C, 37~38 °C and 37~38 °C, with 20.40%, 19.78%, and 20.81% of T_max values in the H database, respectively. From the perspective of the NH−5 and NH+5 databases, the distributed core range of sample values in NH−5 databases was higher than that of NH+5 databases. The study determined the critical threshold temperature for triggering heat occurrence to single-season rice as 30~40 °C.

3.2. Critical Threshold Identification for Heat Occurrence to Single-Season Rice

3.2.1. The First Three Key Thresholds Explored Preliminary

The important results of 11 different test thresholds from 30 to 40 °C with 1 step are present in Figure 3. It can be seen that the IF values at different phenological stages had a certain degree of volatility, showing a trend of initially decreasing, then increasing, and then decreasing with the increase of test thresholds. The first three IF values at tillering–jointing are 30, 32, and 36 °C, with scores of 17.8, 14.02, and 13.14, respectively. The top three IF values at booting and flowering were homogeneously 30, 31, and 35 °C. The test threshold of 35 °C at flowering had the highest importance score with a value of 17.15, while 30 °C at booting ranked first with a score of 18.82, followed by 31 and 35 °C at booting, with scores of 12.47 and 10.93. Relatively speaking, test thresholds at filling ranked first three IF values were higher than those in other phenological stages, with the first three higher scores being 11.62 at 31 °C, 12.95 at 37 °C, and 13.15 at 38 °C.

3.2.2. Optimal Critical Threshold for Heat Occurrence at Each Phenological Stage

Overall Accuracy values from 30 to 40 °C with 1 step for heat occurrence were calculated and demonstrated in Figure 4, while the detailed statistics for TP, TN, FP, FN, Overall Accuracy, TPR, and FPR corresponding to the first three key thresholds were listed in

Table 3. Detailed statistics of Overall Accuracy and ROC curves in the H and NH databases. (TP and FP represent the number of rice heat stress in the H database that are retrieved correctly and incorrectly, respectively; TN and FN represent the number of non-heat samples in the NH database that are retrieved correctly and incorrectly, respectively; TPR and FPR represent the proportion of TP to all positive observations (TP + FN) and FP to all negative observations (TN + FP), respectively.).

Phenology	Test Threshold (°C)	TP	TN	FN	FP	Overall Accuracy	TPR	FPR
Tillering–jointing	30	122	3	0	241	0.342	1	0.988
	32	122	57	0	187	0.489	1	0.766
	36	122	235	0	9	0.975	1	0.037
Booting,	30	108	6	0	210	0.352	1	0.972
	31	108	9	0	207	0.361	1	0.958
	35	108	171	0	45	0.861	1	0.208
Flowering	30	148	46	0	250	0.437	1	0.845
	31	148	65	0	231	0.480	1	0.780
	35	148	253	0	43	0.903	1	0.145
Filling	31	237	7	0	467	0.343	1	0.992
	37	237	356	0	118	0.834	1	0.249
	38	234	373	3	101	0.854	1	0.213

. Lower thresholds like 30–32°C correctly identified heat samples (with all TPR values of 1) but falsely identified non-heat samples (with FPR values from 0.776 to 0.988), resulting in lower Overall Accuracy (without the values above 0.5). For tillering–jointing, the peak of Overall Accuracy was set at the test threshold of 36 °C (exactly 0.975) with higher TPR and lower FPR. As for booting and flowering, despite being below the maximum value, 35 °C still had a high level of Overall Accuracy (exactly 0.861 and 0.903) with a high rate of correctly recognizing heat occurrence (TPR = 1) and a low rate of falsely recognizing non-heat occurrence (FPR, 0.208 at booting and 0.145 at flowering). Given that the optimal critical threshold must concurrently have a high IF value, high Overall Accuracy, high TPR and low FPR, test thresholds of 36 °C at booting and 37 °C at flowering were discarded due to low IF values below 10. The optimal critical threshold at filling was 38 °C, which had the highest Overall Accuracy and IF values.

The ROC curves for each phenological stage were drawn in Figure 5, illustrating a trade-off between achieving a high true-positive rate and a low false-positive rate. Each test point on the ROC plot referred to a sensitivity-specificity pair corresponding to a particular decision threshold. AUC values were obtained from TPRs and FPRs at each test point, reaching 0.993 at tillering–jointing, 0.959 at booting, 0.989 at flowering, and 0.930 at filling, corresponding to ‘excellent’ performances of T_max as the trigger factor to discriminate whether heat occurred or no heat. Measured comprehensively by the method above, the T_max values of 36, 35, 35. and 38 °C were sequentially confirmed as the optimal critical threshold for identifying the occurrence of heat exposure to single-season rice in MLRYR as the phenological stage advanced.

3.2.3. Verification and Application of a Multi-Stage Heat Event

The effects of long-time series of high-temperature events on crops are more complex and severe. According to the heat-disaster record, the southeast HB province suffered from the heat weather process for the first time during 2~6 July 2013 (day of the year (DOY) 183~187, tillering–jointing). Later, the single-season rice in this region was heavily affected by heat stress from 21 to 30 July (DOY 202~211, late booting and early flowering). After that, the grain filling of single-season rice was damaged again by the third heatwave process during the period from 13 to 18 August (DOY 225~230, filling). T_max values were analyzed from 21 stations in southeast HB from 2 July to 18 August (DOY 183~230), including a three-stage actual heat process indicated by the grey-shaded area (Figure 6). Similar to the previously described construction of heat and non-heat samples, the T_max values of 21 locations corresponding to 27 June~1 July and 19~23 August (DOY 178~182 and DOY 231~235, respectively) were checked as the previous 5 days of non-heat process (Figure 6). Temperature values for the verification process of this whole heat event fluctuated significantly, and the daily T_max values reached the peaks in three grey shadow areas, respectively.

During the first heat process (DOY 183~187), 28.5% of stations with T_max values above 36 °C showed heat stress one day ahead of the actual heatwave occurred. The air temperature was continuously heated until the T_max of all stations was greater than 36 °C after 2 days (DOY 184). The heat gradually decreased but remained high until all stations were judged to be non-heat on DOY 188, consistent with the actual record ending of the first heat wave. In terms of the second heatwave on record (DOY 202~211), only 1~2 stations from DOY 197 to 201 exceeded the determined threshold of 35 °C, with a difference of about 0.2 to 0.6 °C. Rice heat stress occurred on DOY 202, with 12 locations identifying heat development, and air temperatures continued to rise through DOY 203 to 205. On DOY 206, all stations experienced heat stress with up to 1.8~5.1 °C difference from the threshold. Severe high temperatures persisted until DOY 211, with the proportion of heat stations decreasing to 71.4%. The southeast HB was judged to be the first day that non-heat occurred on DOY 215, a delay of 3 days compared to the actual heat record. The third heat process occurred during DOY 225~230, and heatwave start-up determined by the 38 °C threshold was delayed by one day with five stations on DOY 226. 94.5% of the stations were identified as heat stations on DOY 228, reaching the peak ratio during this third process. The third heatwave ended on DOY 230, exactly consistent with disaster records in MLRYR, as only one station had T_max above 38 °C on DOY 231. It was evident that various temperature thresholds were robust and reliable indicators in identifying heat stress on single-season rice at different phenological stages.

3.3. Characteristics of Heat Exposure for Single-Season Rice in MLRYR

3.3.1. Accumulated Harmful Temperature (Tcum) and Number of Heat Days (HD)

The spatiotemporal distribution characteristics were explored from 1971 to 2020. From a spatial perspective (Figure 7a–d), the regional mean of Tcum from tillering–jointing to filling followed a response of first increasing and then decreasing, and the corresponding mean Tcum reached its maximum value of 8.7 °C at flowering. High Tcums were mainly found in mid-western ZJ, mid-eastern JX, and a small part of areas in HN at booting, as well as the southern region of MLRYR at flowering (e.g., mid-eastern HN, central JX, western ZJ and the southeast corner of HB), with Tcum values for 19.7% stations at booting and 30.1% stations at flowering greater than 12 °C. The Tcums at tillering–jointing and filling were comparatively lower, with the observed Tcum for most stations being less than 6 °C. According to the spatial distribution of HD (Figure 7e–h), the average annual HD of most locations at tillering–jointing and filling were below 4 d, while there were generally greater than 4 days of high-temperature weather at booting in southern MLRYR. The region with high HD values above 6 d at booting was the same as the high Tcum-dominated region, accounting for 32.5% of heat stations. Compared to the HD at flowering, the range of heat stress with HD values above 4 d for single-season rice was wider than that at booting, and HD gradually decreased from south to north in this region. In addition, the number of non-heat stations reached 64 for single-season rice at filling, accounting for nearly 12.8% of the 501 meteorological stations in MLRYR, followed by tillering–jointing with 16 locations. Overall, the middle and northeast parts of the study region were considered regions with a high frequency of rice heat stress.

In the past 50 years (Figure 8), high occurrence times of rice heat stress were mainly concentrated in 1971, 1978, 1988, 2003, 2007, 2013 and 2017, with the regional HD for each station being greater than 16 d. Among them, Tcum in 2003, 2013 and 2017 assumed a leading position, with values of 49.9, 54.9 and 41.3 °C, respectively. The MK test showed that the temporal variation characteristics of HD were consistent with that of Tcum (Figure 9), with both showing a decreasing trend first since 1971 and then gradually increasing after 2000, which was also considered a mutation point. It is worth noting that Tcum and HD increased significantly since 2010 (p < 0.05). Therefore, the period was divided into three slices, i.e., 1971–2000, 2001–2009, and 2010–2020, to distill the difference in mean trends of Tcum and HD (Figure 10). It was clear that Tcum and HD were relatively low during 1971–2000 but increased during 2001–2009 and 2010–2020.

3.3.2. First Heat Date (FHD) and Last Heat Date (LHD)

It was obvious that FHD of 59.9% stations occurred earlier on DOY 162–168 (early tillering–jointing, Figure 11a), as heat stress for single-season rice occurred later in the northeast corner of MLRYR on DOY 179–188 (late tillering–jointing and early booting). On the contrary, the heat process ended later in the earlier FHD region with a high LHD of DOY 224–243 (filling, Figure 11b), and the northeast corner of MLRYR and central HB region showed the non-heat state earlier with the lower LHD of DOY 204–213 (flowering). Only four stations did not recognize the heat for single-season rice throughout the investigated phenological stages.

4. Discussion

4.1. Determined Thresholds for Different Phenological Stages

Over the past years, the advancement in crop simulation models, climatic models and machine learning have made it possible to explore the impacts of extreme climate events on crop growth and yield on a large scale, over a long sequence as well as from multi-dimensions [5,9,22,24]. Identification indicators for agro-meteorological disasters and the criteria to distinguish the occurrence of crop disasters are the basis for the research on hazard damage mechanisms, spatiotemporal distribution characteristics, monitoring and early warning, as well as risk assessment. In this paper, critical temperature thresholds (T_max) for various phenological stages were established to reflect the ability to identify heat stress for single-season rice in MLRYR. Based on meteorological data combined with historical disaster records, regional thermal hazard indicators were constructed as 36 °C at tillering–jointing, 35 °C at booting–flowering and 38 °C at filling, respectively. As exhibited in a series of heat control experiments, exposure to 5d above 35 °C at flowering could result in up to 90% spikelet sterility [17,23]. In contrast, the negative impacts of high temperatures (over 35 °C) at tillering were considerably less, meaning that rice could tolerate relatively high temperatures at the vegetative phase [7,45,46]. High temperature (38 °C day/30 °C night) in the Aus type N22 for 6 h increased 29% spikelet sterility at flowering [47], while identical treatment for 20 consecutive days at filling reduced 24.6% grain weight [48]. Rice plants suffering from 40 °C for 10~15 d at tillering and filling showed about 86% and 13% lower total yield per plant, respectively [49,50]. The differences observed between temperature thresholds and the results mentioned above collectively indicate that rice’s tolerance and sensitivity to high temperatures vary depending on the growth stage, with the booting and flowering stages being the most sensitive stages, followed by tillering–jointing and filling. As confirmed in previous studies [4,25,51], hot weather before and during flowering tended to lead to more severe heat stress for rice. Heat stress occurring during the vegetative phase could affect tiller formation, and its extended heat exposure during the booting stage would impact spikelet meristem differentiation [52]. The influence of hot weather on rice filling is controversial. It was reported that a certain degree of heat stress could compensate for reduced grain-filling duration by increased filling rates [23,48,53], while some researchers believed it affected the photosynthesis and material migration of rice, thereby altering grain size, increasing grain chalkiness and then reducing grain quality and weight [4,46,53,54,55].

The critical temperature threshold for rice heat stress varies due to the complexity of disaster occurrence, the roles of agronomic management practices, and crop genetic backgrounds [4,14]. Given the adaptation of rice plants to local climate and environment, even the same variety of rice in different regions may have different thresholds, and it implies that heat critical threshold could be of regional practicality and variety limitations. For instance, experts have reported that high day temperature (38 °C) during anthesis can reduce the number of pollen grains on the stigma in the Aus type N22 by 55% and the japonica type Moroberekan by 86%, but not in the indica type IR64 [47]. This difference in susceptibility to heat stress is attributed to different genetic backgrounds. In this research, the importance analysis, overall accuracy and ROC curve were jointly used to gain the optimal threshold for single-season rice in MLRYR, comprehensively considering the proportions of correctly identified heat (TPR) events and incorrectly identified non-heat (FPR) events. The model for identifying heat or non-heat occurrence to single-season rice based on daily T_max performed excellently at each phenological stage, with relative IF values and AUC values greater than 10 and 0.930, respectively. This kind of historical disaster-based indicators construction method has been widely used in coping with disaster indexes for agro-meteorological disasters, such as heat [25], drought [37,56] and frost [57,58], avoiding the effects of subjective factors and the limitations of experimental regions.

Furthermore, a multi-stage heat event with long time series was checked for conformity to the disaster record based on the confirmed critical thresholds at different phenological stages (Figure 6), and the results indicated that the single-season heat indicators can reasonably reflect actual heat exposure in regional scales. As rice heat stress exposure in MLRYR was always highlighted and more severe than in other regions, the typical variety in MLRYR was considered in the study. However, it was definite that the characteristics of heat exposure might be different as regional conditions and rice varieties might vary. Although refined thresholds for different growth stages might not play the best application in some specific regions, these thresholds could be of practicality in regions with the same latitude and with similar rice varieties. Of course, the application of the proposed method should be further explored in other climatic regions and varieties.

4.2. Potential Adaption Strategies to Mitigate Rice Heat Stress

It is an indisputable fact that global warming has a negative impact on rice yield by accelerating the growth process of rice and shortening its biomass accumulation time [41,59]. The relationship between rice yields and temperature varied with the extent of rising temperature and region baseline temperature [60]. The higher baseline temperature leads to greater heat stress for rice under future global warming.

Decision-makers should highlight the need to develop adaptive rice security mechanisms to mitigate the negative effects of warmer conditions. In our present work, extreme rice heat stress exposure in the southern MLRYR was more severe than that in the northern part (Figure 7), characterized by strong intensity and prolonged duration. The indica-type rice should be chosen as the main cultivation in the southern region instead of the Aus type and the japonica type. As concluded by previous reporters [47,61], the indica type was more tolerant to heat exposure than the remaining two. Derived from Figure 7 and Figure 11, rice plants at booting and flowering were greatly susceptible to extremely high temperatures, as well as having become increasingly fierce over the past decade (2010–2020, Figure 9 and Figure 10). Hence, appropriate adjusting of the rice’s sowing and transplanting dates [62], giving rise to variation in phenology length, can be an effective strategy to escape or avoid in the time of a crucial phenological phase subjected to heat exposure in the study region. Assuming that rice plants inevitably encounter high temperatures, several agronomic management practices can be taken into account by artificial control, e.g., altered flowering pattern, proper application of growth regulators such as CTK, SA, BR, and ethylene precursors or increased transpiration cooling of the canopy [4,7]. Obviously, most of the current adaption strategies are targeted at exposure to heat stress that occurred at the reproductive phase, but we advanced explored phase affected by heat stress to the tillering–jointing stage in MLRYR. In theory, the tiller number and panicle number have shown a strong and positive correlation in rice in both favorable or stress conditions [45]. Heat stress for the entire region, FHD of most sites occurred at early tillering (Figure 11a, high temperature during tiller formation should be paid equal attention, and some measures should be used to alleviate initial rice heat stress. Commonly, the tiller number is determined by a hormonal network of abscisic acid (ABA) and strigolactones [63,64]. Recent research [65] demonstrated that ABA could repress the outgrowth of unproductive tillers by regulating the non-participation of strigolactones in the formation of high-order tillers. Applying exogenous ABA or drying the field can promote the endogenous level of ABA so as to enhance the stress resistance of rice vegetative phase to heat stress, which has been presented in different abiotic stress tolerances, such as drought, high salinity, and cold [65,66,67].

4.3. Uncertainties and Limitations

According to extensive research and practice [5,9,14,24], 35 degrees Celsius has been identified as a critical temperature threshold for different development stages of rice plants. This study emphasized that heat thresholds were strongly influenced by local climate conditions, crop genetics, and evolving phenological stages, ultimately redefining the critical threshold library for single-season rice heat stress exposure in MLRYR. There were two relatively higher thresholds (36 °C and 38 °C) in this library established by the historical disaster-based indicators construction method. Compared to the threshold of 35 °C, higher thresholds might raise the risk of lack of ability to capture slight heat stress and lead to an underestimation of rice heat stress risk assessment. Even then, research results from MK test analysis of heat exposure via the investigated phenological phase still suggested that rice plants were increasingly being exposed to extreme heat with rising global mean temperatures since the 2000s, particularly with obvious increased trends in regional exposure since 2010. Moreover, such a trend would be more severe with future warming, which was consistent with the findings of previous studies [9,14,68]. In the near future, we could conduct field experiments under relevant critical temperatures at different phenological stages to verify the reliability of a critical threshold database or improve the precision of the threshold to decimal places at the region scale.

The phenology of single-season rice has varied over the years due to the influences of climate change and production demands. Consequently, a fixed threshold sowing date might not be appropriate for prediction in ongoing years, but we still lack proper ways to predict it currently. Meanwhile, warming temperatures increase the speed of crop development, which could potentially shift the timing of the reproductive phase in future climates [22]. Rough estimates are that 1 °C warming reduces the time to flowering by ~4 days [69], but this will vary by crop, variety and baseline temperatures. Apart from exposure to extreme reproductive heat, further factors could help to determine the overall crop response to environmental change (e.g., increasing atmospheric CO₂ and O₃ concentrations, changes in soil moisture and vapor pressure deficit, increases in mean temperatures, and the interactions between these changes). According to experts, elevated night temperatures have a determinant impact on yield and grain quality [52], and this trend seems to be accelerating more rapidly than day temperature. An important caveat concerns our inability to account for CO₂ concentrations. Plants use CO₂ as an input in the photosynthesis process; therefore, increasing CO₂ levels may spur plant growth and yields. Yield declines stemming from extreme heat stress may be offset by CO₂ fertilization [24,70]. Field experiments found that high temperature (35~38 °C) and high CO₂ level (660 µmolmol⁻¹) only had a positive impact on rice biomass but a negative impact on rice yields by reduced utilization of additional non-structural carbohydrates in the sink under heat exposure [70,71,72]. Rice plants grown under elevated CO₂ would be more sensitive to heat stress [70,71]. To date, the magnitude of interaction on rice from both elevated temperature and increased CO₂ is still debated and requires further investigation. Given the basis of previous studies, a comprehensive index for regional heat evaluation should be explored and constructed in future studies in order to jointly reflect the influence of the interaction between high temperature and its meteorological factors on rice plants and yield formation. The findings in our study can better assist the construction of a comprehensive index as fundamental information for the index exploration process.

5. Conclusions

Based on historical disaster records and daily maximum temperature (T_max), a critical threshold database of rice heat stress at different growth stages was constructed to evaluate single-season rice heat stress exposure along MLRYR between 1971–2020. T_max of 36 °C at tillering–jointing, 35 °C at booting, 35 °C at flowering and 38 °C at grain filling were identified as the critical thresholds for heat occurrence to single-season at different phenological stages, with relatively IF values and AUC values greater than 10, 0.930, respectively. A multi-stage heat event checked for conformity also showed that the critical threshold at each phenological stage could be applied to accurately identify heat events for rice plants in MLRYR. The spatiotemporal characteristics of heat stress for single-season rice in MLRYR were analyzed, and it was found that Tcum and HD increased from north to south, with most parts of the study region experiencing significant increases from 2010 to 2020 by the MK test. FHD occurred earlier for single-season rice across the region, except in the northeast corner, where it was later. Conversely, the LHD occurred later in the same area where FHD appeared earlier. Given the potential for future climate extremes, research on the critical temperature thresholds for heat stress to rice is of great significance for improving the rice heat stress meteorological service and for reducing yield and economic losses caused by heat stress.