The Effect of Climate Change on Important Climate Variables in Taiwan and Its Potential Impact on Crop Production

Abstract

Alterations in reference evapotranspiration (ET₀) and precipitation (PP) resulting from global warming substantially affect water resources and agriculture. This study analyzed trends in ET₀, PP, and key climate variables—including air temperature (T), vapor pressure deficit (VPD), wind speed, and solar radiation (R_s)—across Taiwan from 1995 to 2022. Trends were assessed using the modified Mann–Kendall test and the multivariate Man–Kendall test at both station-wise and multi-station scales. Results indicated that ET₀ was primarily influenced by R_s, followed by T, wind speed, and VPD. Station-wise analysis revealed increasing trends in annual and seasonal T, R_s, and ET₀, while over 50% of wind speed series showed significant declines. Multi-station analysis confirmed an overall rise in ET₀. In eastern Taiwan, rising T and declining VPD and wind speed may increase the risk of pest and disease outbreaks. The arid index exhibited a general downward trend, particularly in summer, with 75% of the stations in eastern Taiwan exhibiting significant declines, suggesting a shift toward drier conditions. These findings imply that fewer crop options may be suitable for cultivation in eastern Taiwan due to water resource constraints. Additionally, seasonal and annual PP showed slight decreases, with a more uneven distribution observed in central Taiwan. Therefore, improving hydraulic facilities and irrigation systems will become important. Furthermore, comparisons between the multivariate Mann–Kendall test and the traditional univariate approach revealed some different results, indicating the need for further research to identify a more reliable approach.

1. Introduction

According to the United Nations Intergovernmental Panel on Climate Change (IPCC), human activities have gradually increased the surface temperature of the atmosphere, ocean, and land. Since the early 20th century, the global average surface temperature has risen by approximately 1.1 °C, primarily due to anthropogenic greenhouse gas emissions [1]. Projections from various climate models indicate that the 2 °C warming threshold could be exceeded between 2050 and 2060 [2]. Moreover, even under the most aggressive greenhouse gas emission reduction scenario, global warming is expected to persist until at least 2050 [3]. Most phenomena in the biosphere are affected by climate change due to global warming, and the impact of climate change on the hydrological cycle receives much attention [4]. Changes in the hydrological cycle can significantly affect ecosystems, agricultural production, and water resources [5,6]. Therefore, quantifying the impact of climate change on the hydrological cycles, is essential for developing timely adaptation measures against global warming.

In addition to evaluating the impact of climate change on the environment and water resources from basic climate parameters such as temperature, it is more appropriate to consider complex parameters such as evapotranspiration [7]. Evapotranspiration not only represents the atmospheric evaporative demand but also reflects plant water consumption [8]. It is one of the key variables for revealing the temporal-spatial patterns of eco-hydrological processes [9]. If evapotranspiration is high and close to precipitation, it may lead to water shortages for crops [10]. Furthermore, increased evapotranspiration will decrease surface runoff and groundwater, thereby reducing the amount of available water for human activities and natural processes [9]. Investigating the impact of climate change on evapotranspiration can aid in developing appropriate solutions to mitigate potential water resource challenges [11]. However, evapotranspiration is a highly complex process influenced by multiple climate variables and crop characteristics. To improve the effectiveness of evapotranspiration research, the Food and Agriculture Organization (FAO) of the United Nations recommended the reference evapotranspiration (ET₀) approach [12].

ET₀ is defined as the evapotranspiration of a hypothetical grass with a height of 0.12 m, a fixed surface resistance of 70 s/m, and an albedo of 0.23 [13]. It quantifies the atmospheric evaporative demand through climate variables regardless of the crop type, growth stage, and management method [12,14]. The water demand of a specific crop can be obtained by the product of ET₀ and the crop coefficient (K_c) [13], which is called the K_c-ET₀ procedure. Therefore, ET₀ can be used as an indicator of crop water demand and is closely related to food production [15]. Moreover, ET₀ can be used to characterize the local climate by calculating the drought index, and anomalies in ET₀ can also be used to indicate the onset of drought [16]. Given these reasons, ET₀ plays a crucial role in irrigation planning and water resource management.

There are four main climate variables that affect evapotranspiration: air temperature, solar radiation, vapor pressure deficit (VPD), and wind speed. Generally, these variables show a positive contribution to evapotranspiration; that is, higher values of these variables typically lead to increased evapotranspiration [12,17,18,19]. Intuitively, one of the expected effects of global warming on evapotranspiration is that increasingly higher surface temperatures will lead to increased evapotranspiration. However, many studies have found that evapotranspiration decreases in some regions, a phenomenon known as the “evaporation paradox” [12,17]. The evaporation paradox occurs because not all changes in climate variables that affect evapotranspiration are in the same direction of increasing evapotranspiration [20,21,22]. Additionally, these climate variables may influence each other [17], and their interactions can also impact evapotranspiration [18]. Changes in evapotranspiration are the integrated consequences of several influencing factors, and focusing on a few factors can lead to biased conclusions. Therefore, it is important to understand the trends in ET₀ and how the related climate variables affect ET₀. Furthermore, regarding available water resources, evapotranspiration can be treated as consumption, and precipitation is treated as supply, so it is necessary to consider their changes together, such as by calculating the arid index [11,17].

There are many parametric methods for trend analysis of time series, mostly based on the linearity and normality assumptions [23]. However, the time series of hydrology and other environmental variables often do not meet these assumptions. Therefore, nonparametric trend analysis methods, which require fewer assumptions, are often used to detect and estimate trends over time [24]. The most commonly used method is the Mann–Kendall (MK) test, proposed by Kendall [25], and it is recommended by the World Meteorological Organization for the analysis of hydrological or meteorological time series [12,19]. Although the MK test is a good analytical method, it can only handle one variable and one station at a time. For regional data, not only the temporal but also the spatial dimensions of the observations are taken into account when testing trends [26,27]. In this case, because multiple locations should be considered simultaneously, it constitutes a multivariate setting. Therefore, a multivariate approach should be used to investigate a common trend across multiple stations. Chebana et al. [24] summarized three types of multivariate methods extended from the MK test: (i) covariance inversion, (ii) covariance sum, and (iii) covariance eigenvalue (CE). They concluded that the statistical power of these statistical methods was not the same, and the CE method was recommended. However, multivariate MK tests have rarely been applied to trend analyses for ET₀ and important climate variables. Most studies still employ the univariate approach to trend analysis involving multiple stations [6,17,19]. These studies average the data from all stations in the same region and then apply the MK test to the arithmetic mean series. This approach is named the “univariate mean (UM) method” in this paper. To the best of our knowledge, few studies have compared the performance of the multivariate method and the UM method in trend analysis involving multiple stations.

Taiwan is an island located in East Asia on the northwest side of the Pacific Ocean, with an area of about 36,000 km². Approximately 70% of the island comprises mountains and hills, with the plains mainly concentrated along the western coast. The climate ranges from tropical to subtropical and is governed by the East Asian Monsoon. The annual rainfall is over 2000 mm, approximately twice the global average. Although Taiwan has abundant rainfall, the amount of available water per person is only one-sixth of the global average because the rivers are short and steep, causing rainfall to quickly discharge into the sea, and only a small portion is retained. Additionally, the distribution of rainfall across regions and seasons is uneven, with year-round rain in the northern region but primarily summer and autumn rain in the central and southern regions. This variability makes the utilization and management of water resources more difficult. Rice (Oryza sativa) is the most cultivated crop in Taiwan, followed by maize (Zea mays), peanut (Arachis hypogaea), mango (Mangifera indica), tea (Camellia sinensis), banana (Musa sapientum), and Azuki bean (Vigna angularis). According to government statistics, more than 60% of Taiwan’s water resources are used for agricultural irrigation, and half of the irrigation water is used for rice cultivation [28]. With the growing population and unstable water supply, how to effectively manage limited water resources and maintain the present crop productivity is a global problem [29]. There are few studies that have evaluated the hydrological impact of climate change in Taiwan, and most of them were published many years ago. Hsu and Chen [30] found that the temperature in Taiwan had risen significantly. Yu et al. [31] concluded that the evapotranspiration of paddy fields in southern Taiwan would increase between 3.1% and 5.5%, indicating that rice cultivation would have a higher water demand. Recently, many studies have examined the effects of climate change in Asia [11,12,17,19], but the data regarding Taiwan were not included. Therefore, these results may not reflect the situation in Taiwan.

This study analyzed the temporal trends of key climate variables in Taiwan from 1995 to 2022, including ET₀, precipitation, arid index, and climate variables influencing ET₀ (i.e., air temperature, VPD, wind speed, and solar radiation). Trend analysis was conducted at two spatial scales—station-wise (each station was analyzed individually) and multi-station (all stations within the same geographic region were analyzed simultaneously)—and across five temporal units: spring, summer, autumn, winter, and annual. Additionally, this study aimed to compare two different methods, including the CE and UM approaches, to identify common regional trends. The findings offer valuable insights for researchers and policymakers in assessing climate change and its implications for dry and wet conditions across different geographic regions in Taiwan. Furthermore, the results provide a comprehensive perspective on the potential impacts of climate variability on crop production.

2. Materials and Methods

2.1. Data

Taiwan Island was divided into four geographic regions—northern, central, southern, and eastern—following the classification method used by the agricultural administration departments. Daily meteorological data from 1995 to 2022 were collected from 18 meteorological stations across these regions (Figure 1). These stations are located within the primary crop cultivation zones of each region and can provide long-term observation data. The 28-year daily meteorological data on mean air temperature (T, °C), precipitation (PP, mm), mean relative humidity (RH, %), mean atmospheric pressure (P, kPa), solar radiation (R_s, MJ/m²), and mean wind speed at 2 m height (u₂, m/s) were obtained from each station. These data were provided by the Central Weather Administration (https://agr.cwa.gov.tw/NAGR/, accessed on 18 January 2024). The total missing data rate did not exceed 10%, which is within the acceptable limit for environmental research and can be imputed using appropriate imputation methods [32]. Missing values were estimated using the same day-of-the-year average method [33], where missing data for a specific date were replaced by the average of the same date in other years with complete records.

After missing value imputation, the Penman–Monteith equation (Equation (1)) was used to calculate daily ET₀. This equation is recommended by the FAO to estimate ET₀ and is considered the most reliable method to estimate ET₀ [8,16,19] under various climatic conditions:

$${\mathrm{E}\mathrm{T}}_0={\displaystyle \frac{0.408\Delta ({\mathrm{R}}_{\mathrm{s}}-\mathrm{G})+\mathsf{\gamma }\frac{900}{\mathrm{T}+273}{\mathrm{u}}_2({\mathrm{e}}_{\mathrm{s}}-{\mathrm{e}}_{\mathrm{a}})}{\Delta +\mathsf{\gamma }(1+0.34{\mathrm{u}}_2)}}$$

(1)

where R_s is the net radiation (MJ/m²/day); G is the soil heat flux density (MJ/m²/day); T is the mean daily air temperature (°C); u₂ is the wind speed at 2 m height (m/s); e_s is the saturation vapor pressure (kPa), e_s = 0.6108 * exp (${\displaystyle \frac{17.27\mathrm{T}}{\mathrm{T}+273.3}}$); e_a is the actual vapor pressure (kPa), e_a = e_s * ${\displaystyle \frac{\mathrm{R}\mathrm{H}}{100}}$; Δ is the slope of the vapor pressure curve (kPa/°C), Δ = 4098 * ${\displaystyle \frac{{\mathrm{e}}_{\mathrm{s}}}{{(\mathrm{T}+273.3)}^2}}$; γ is the psychrometric constant (kPa/°C), γ = 0.00163 * ${\displaystyle \frac{\mathrm{P}}{2.45}}$. The soil heat flux is usually small compared with R_s, and it is assumed to be zero on the daily scale. The $({\mathrm{e}}_{\mathrm{s}}-{\mathrm{e}}_{\mathrm{a}})$ is the difference between the saturation vapor pressure and the actual vapor pressure, that is, VPD.

The arid index (Equation (2)) is the ratio between PP and potential evapotranspiration (ET_P), which integrates PP and evapotranspiration information to classify locations based on aridity. According to the arid index, a location can be classified as humid (arid index > 1), semi-humid/semi-arid (0.25 < arid index ≤ 1), or arid (arid index ≤ 0.25) [17]. The definition of the arid index is on the basis of ET_P instead of ET₀. Although the concept of ET₀ is slightly different from ET_P, many studies have used ET₀ as a substitute for ET_P [34,35,36,37], and ET₀ is actually more popular in agriculture and irrigation [8]. Therefore, this study used ET₀ to calculate the arid index. A summary of the geographic and annual meteorological characteristics of the studied stations is presented in

Table 1. Summary of the geographic conditions and meteorological characteristics of each station. The meteorological characteristics are averaged over the data from 1995 to 2022.

Region	Station	Latitude (N)	Longitude (E)	Altitude (m.a.s.l.)	T (°C)	VPD (kPa)	u2 (m/s)	Rs (MJ/m2/yr)	PP (mm/yr)	ET0 (mm/yr)	Arid Index
Northern	72C440	24.95	121.03	70	22.27	0.55	3.94	4736.69	1451.56	1530.90	0.96
	72D080	24.61	121.16	1048	17.32	0.25	0.50	3589.81	2711.03	999.20	2.91
	82A750	24.96	121.63	401	20.06	0.37	1.45	3810.00	3547.91	1136.49	3.50
	82C160	24.91	121.19	195	21.56	0.50	2.89	4295.31	2098.84	1354.76	1.58
	K2E360	24.50	120.83	100	22.23	0.51	2.36	4812.51	1748.37	1476.00	1.20
Central	72G600	24.00	120.53	19	23.39	0.58	1.99	4236.69	1380.61	1359.50	1.05
	72K220	23.63	120.48	60	23.38	0.56	1.87	4510.79	1787.60	1416.42	1.29
	72M360	23.36	120.28	6	23.59	0.52	3.02	5491.08	1501.06	1648.99	0.93
	82H840	23.76	120.74	390	21.52	0.34	0.91	4633.75	2279.39	1330.67	1.73
	G2F820	24.03	120.69	90	23.52	0.55	2.21	4575.15	1667.58	1440.28	1.19
	U2H480	23.67	120.80	1150	17.05	0.21	1.06	3373.81	2473.58	909.84	2.76
Southern	BSQ810	21.95	120.80	20	25.34	0.73	4.12	5989.74	2022.48	1997.92	1.04
	B2N890	23.06	120.34	31	23.71	0.53	1.82	4452.66	2035.04	1386.30	1.53
	72Q010	22.71	120.53	45	24.98	0.69	1.45	4509.98	2360.49	1468.69	1.70
Eastern	72S200	22.83	121.08	240	22.50	0.53	1.51	3184.27	1980.31	1053.10	1.96
	72S590	22.81	121.07	290	22.43	0.39	1.30	2862.04	1897.40	905.51	2.35
	72T250	23.98	121.56	36	22.92	0.54	1.13	3857.17	2050.43	1218.36	1.75
	72U480	24.69	121.72	27	22.45	0.48	1.82	4250.42	2946.22	1325.94	2.24

m.a.s.l.: meters above sea level; T: mean air temperature; VPD: mean vapor pressure deficit; u₂: mean wind speed at 2 m height; R_s: accumulated solar radiation; PP: accumulated precipitation; ET₀: accumulated reference evapotranspiration.

.

$$\mathrm{A}\mathrm{r}\mathrm{i}\mathrm{d} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}={\displaystyle \frac{\mathrm{P}\mathrm{P}}{{\mathrm{E}\mathrm{T}}_{\mathrm{P}}}}\cong {\displaystyle \frac{\mathrm{P}\mathrm{P}}{{\mathrm{E}\mathrm{T}}_0}}$$

(2)

2.2. Analysis of Correlation and Relative Impact of Climate Variables

Before conducting the trend analysis, we used the 28-year daily data of all stations (a total of 184,086 observations) to investigate the correlation between T, VPD, u₂, R_s, and ET₀. The impact of T, VPD, u₂, and R_s on ET₀ was also investigated. Because a nonlinear relationship between ET₀ and its related climate variables exists [38,39], the Spearman correlation coefficient (ρ), which is applicable to this situation, was used as a measure of correlation. Random forest (RF) model was used with T, VPD, u₂, and R_s as predictors, and ET₀ was used as the response variable to evaluate the relative impacts of climate variables on ET₀.

RF is a nonparametric method that does not require the assumptions of a linear relationship and data distribution [40]. Furthermore, RF is an ensemble learning model formed by aggregating a large number of base models. The base model of RF is the classification and regression tree (CART), which can deal with complex interactions and nonlinearities in data [41]. Each time a CART in the RF splits a node, it randomly selects a subset of candidate predictors, and then the best splitter is determined from these randomly selected predictors based on the decrease in node impurity. In the case where the response variable is quantitative, node impurity is measured by the residual sum of squares (Equation (3)). One of the important capabilities of RF is to evaluate the importance of predictors for predicting the response [42,43]. The importance of each predictor is measured by the total decrease in node impurities from the splitting on the variable, averaged over all CARTs in the RF. An important predictor is often selected for splitting and results in a significant decrease in node impurities.

$$i(t)={\sum }_{i\in S\left(t\right)}{(y_i-{\widehat{y}}_t)}^2={\sum }_{i\in S\left(t\right)}{(y_i-{\overline{y}}_t)}^2$$

(3)

where i(t) is the impurity of node t; S(t) is the subset of training data at node t; $y_i$ is the i-th observation value in node t; and ${\widehat{y}}_t$ is the fitted value calculated by CART at node t.

2.3. Trend Analysis

Trend analysis was implemented at two spatial scales: station-wise and multi-station. In the station-wise trend analysis, the year was divided into spring (March to May), summer (June to August), autumn (September to November), and winter (December to February) according to the meteorological seasons. Five datasets (one annual and four seasonal) were considered for each station, with each dataset containing the observations of the yearly series of T, VPD, u₂, R_s, PP, ET₀, and the arid index. The modified MK test was used to investigate the trend in each variable, and Sen’s slope was used to estimate the magnitude of the trend.

The null hypothesis (H₀) of the original MK test is that there is no trend in the time series, while the alternative hypothesis (H₁) is that a monotonic trend exists. The related equations for calculating the MK test statistic S and the standardized test statistic Z_MK are as follows:

$$S={\sum }_{i=1}^{n-1}{\sum }_{j=i+1}^n\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\left(X_j-X_i\right)$$

(4)

$$\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\left(X_j-X_i\right)=\left\{\begin{array}{c}1, \mathrm{if} \left(X_j-X_i\right)>0\\ 0, \mathrm{if} \left(X_j-X_i\right)=0\\ -1, \mathrm{if} \left(X_j-X_i\right)<0\end{array}\right.$$

(5)

$$\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)={\displaystyle \frac{1}{18}}\left[n\left(n-1\right)\left(2n+5\right)-{\sum }_{p=1}^q{t}_p\left(t_p-1\right)\left(2t_p+5\right)\right]$$

(6)

$$Z_{MK}=\left\{\begin{array}{c}{\displaystyle \frac{S-1}{\sqrt{\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)}}}, \mathrm{if} S>0\\ 0 , \mathrm{if} S=0\\ {\displaystyle \frac{S+1}{\sqrt{\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)}}} , \mathrm{if} S<0\end{array}\right.$$

(7)

where X_i and X_j are the observations at two time steps in the series, j$>$i; n is the total number of observations in the time series, i.e., the length of the time series; Var(S) is the variance of S, where t_p is the number of ties for the p-th value, and q is the number of tied values. The standardized test statistic Z_MK follows a standard normal distribution.

To estimate the magnitude of the trend, Sen’s slope ($\widehat{\beta }$) (Equation (8)) proposed by Sen [44] is used as an estimator, and the sign of $\widehat{\beta }$ indicates whether the trend is increasing ($+$) or decreasing ($-$):

$$\widehat{\beta }=\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\left({\displaystyle \frac{X_j-X_i}{j-i}}\right),1\le i<j\le n$$

(8)

where X_i and X_j are the observations at two time steps in the series; n is the length of the time series.

In many real situations, the observed time series data are autocorrelated. Autocorrelation in data can result in misinterpretation of the MK test’s results. In other words, the presence of autocorrelation increases the probability of detecting trends when they do not actually exist [45,46]. In order to avoid this misinterpretation, this study used the modified MK test to conduct the station-wise trend analysis. The modified MK test was proposed by Hamed and Rao [45], which corrected the value of the variance of the MK test statistic to deal with autocorrelated data. When autocorrelation exists, the modified MK test has an empirical significance much closer to the nominal significance level (α) than the MK test. Additionally, the statistical power of the modified MK test is comparable to that of the original MK test when autocorrelation does not actually exist.

In multi-station analysis, the year was also divided into spring, summer, autumn, and winter, resulting in one annual dataset and four seasonal datasets considered for each geographic region (i.e., northern, central, southern, and eastern Taiwan). Each dataset contains a yearly series of T, VPD, u₂, R_s, PP, ET₀, and the arid index. To investigate the common trend across multiple stations in the same geographic region, an MK-type multivariate approach, the CE method, was employed. This method has the highest statistical power among the main types of multivariate MK tests [24]. The CE method was proposed by Lettenmaier [47] to avoid the problem of inverting a sample covariance matrix. Let $M^{\left(g\right)}$ be the MK test statistic for the observed time series of station g (g = 1, …, m); the CE test statistic is given by:

$$T={\sum }_{g=1}^m{\left(M^{\left(g\right)}\right)}^2$$

(9)

The distribution of T is equal to the distribution of $T^{\ast }=\sum _{g=1}^m{\lambda }_g{Z}_g^2$, where λ_g is the eigenvalue of the covariance matrix and Z_g is an independent standard normal random variable. If all the eigenvalues λ_g are equal to a constant λ, then $T^{\ast }$ is λχ²(m)-distributed; otherwise, the distribution of $T^{\ast }$ is approximated by a three-parameter Gamma distribution. More details regarding the CE method can be found in Chebana et al. [24] and Lettenmaier [47].

In addition, to explore the differences in multi-station trend analysis when using different methods, we employed a univariate approach, i.e., the UM method, to detect common trends. Specifically, the data from all stations within the same geographical region were averaged, and then the resulting arithmetic mean series was used for the MK test. The significance and direction of the trends identified using the UM method were compared with those obtained from the CE method by evaluating their z-values. The absolute value of the z-statistic indicates the statistical significance of each trend, and the sign of the z-statistic indicates that the trend is increasing ($+$) or decreasing ($-$).

2.4. Statistical Software

The statistical analyses were implemented using the R software (version 4.1.3; R Foundation for Statistical Computing, Vienna, Austria). Spearman correlation coefficients were calculated using the ‘cor’ function. The RF was executed using the ‘randomForest’ package (version 4.7–1.1). The number of CARTs grown in RF (i.e., the ‘ntree’ argument) was set to 10,000, and the number of candidate predictors when establishing the CARTs (i.e., the ‘mtry’ argument) was set to 2. The tuning principle of mtry is to minimize the similarity between CARTs within the model and ensure that each CART retains a certain degree of predictive ability [48]. Regarding another hyperparameter, ntree, in theory, increasing its value generally enhances RF model performance [49]. In this study, the values of these two hyperparameters were determined based on a balance between overall model performance and computational efficiency. The importance of each climate variable (measured as the mean decrease in node impurity) was calculated using the ‘importance’ function. The station-wise trend analysis was implemented using the ‘mmkh’ function in the ‘modifiedmk’ package (version 1.6). For multi-station trend analysis, the CE method was executed using the ‘mult.mk.test’ function in the ‘trend’ package (version 1.1.4).

3. Results

3.1. Relationship Between ET₀ and the Related Climate Variables

The heatmap of correlation coefficients (Figure 2) indicates that T, VPD, u₂, and R_s are all positively correlated with ET₀, a result that aligns with expectations based on the Penman-Monteith equation (Equation (1)). These findings are congruent with previous studies, indicating that higher values of these variables typically result in increased evapotranspiration [12,17,18,19]. A particularly strong correlation was observed between ET₀ and R_s (ρ = 0.97), while the Spearman correlation coefficients for T and VPD with ET₀ were 0.62 and 0.72, respectively. A weak positive correlation (ρ = 0.21) was observed between ET₀ and u₂.

The variable importance analysis, measured as the mean decrease in node impurity in the RF model, ranked R_s (28,021,493) as the most influential factor, followed by T (11,445,772), u₂ (9,601,995), and VPD (9,175,731), with relative importance values of 48.1%, 19.7%, 16.5%, and 15.8%, respectively. These results indicate that R_s is the primary driver of ET₀ in Taiwan, followed by T, u₂, and VPD. Temperature and solar radiation are generally regarded as essential inputs for estimating ET₀ [50,51]. In addition, Fang et al. [52] identified T and R_s as the dominant factors influencing ET₀ in Taiwan.

3.2. Trends of Important Climate Variables in Taiwan

This study presents the results of trend analysis using z-values and Sen’s slopes. The z-values indicate the statistical significance of each trend. In other words, larger absolute z-values suggest the existence of the trend. Moreover, the Sen’s slopes quantify the magnitude of trends and enable comparisons of effect sizes across different locations. Among the 18 studied stations, ET₀ exhibits a predominantly upward trend, except at Station 82A750, which shows downward trends across all temporal units (

Table 2. Trend analyses of annual climate variables from 18 meteorological stations. The common trend of a region is tested using the MK-type multivariate test. The trend of each station is tested using the modified MK test. The bold numbers indicate that the trend is significant at $\alpha =0.05$.

Region	Station	Statistic	T (°C /yr)	VPD (kPa/yr)	u2 (m/s/yr)	Rs (MJ/m2/yr)	PP (mm/yr)	ET0 (mm/yr)	Arid Index
Northern	All	z	1.76	0.86	−4.71	3.42	−0.06	3.53	−0.72
	72C440	Slope	0.0274	0.0017	−0.0379	64.4018	−2.8776	14.8468	−0.0121
		z	1.85	1.66	−3.08	5.13	−0.26	4.48	−2.80
	72D080	Slope	0.0287	0.0006	−0.0069	−7.3054	−10.8861	−1.6226	−0.0040
		z	1.27	2.02	−1.02	−0.15	−0.57	−0.24	−0.14
	82A750	Slope	−0.0269	−0.0060	−0.0170	−51.2209	11.1150	−18.5568	0.0477
		z	−1.31	−1.62	−0.97	−0.68	0.69	−1.15	1.24
	82C160	Slope	0.0285	−0.0005	−0.0636	66.2084	−3.6080	14.7723	−0.0222
		z	1.75	−0.21	−3.88	4.68	−0.34	3.09	−1.56
	K2E360	Slope	0.0300	0.0049	−0.0153	38.2372	2.4963	13.5855	−0.0111
		z	1.98	2.10	−1.88	3.48	0.10	6.13	−1.13
Central	All	z	3.40	1.43	−4.40	2.83	0.11	3.50	−1.40
	72G600	Slope	0.0337	0.0026	0.0029	54.7603	−0.2111	16.7439	−0.0162
		z	2.54	2.20	1.03	1.02	0.00	1.15	−1.80
	72K220	Slope	0.0451	0.0000	−0.0283	2.8293	−1.2749	5.4492	−0.0104
		z	4.33	0.14	−2.93	0.07	−0.10	0.70	−1.05
	72M360	Slope	0.0336	0.0081	−0.0335	28.2188	0.8462	12.9502	−0.0066
		z	3.52	2.59	−3.42	1.47	0.06	2.36	−0.93
	82H840	Slope	0.0092	−0.0010	−0.0041	29.9815	6.2557	8.2366	−0.0095
		z	0.36	−0.49	−2.08	1.90	0.41	2.22	−0.89
	G2F820	Slope	0.0359	0.0053	−0.0103	82.7472	−1.1488	26.2093	−0.0216
		z	3.42	2.38	−1.19	2.38	−0.14	3.53	−2.00
	U2H480	Slope	0.0233	−0.0025	−0.0231	22.7595	6.0904	4.9367	−0.0073
		z	3.32	−3.89	−3.43	1.89	0.39	2.39	−0.41
Southern	All	z	2.32	1.53	−5.74	1.75	0.03	2.00	−1.02
	B2Q810	Slope	0.0080	0.0007	−0.0487	3.2620	−1.1748	0.1380	−0.0350
		z	1.34	0.54	−3.24	0.36	−0.22	0.03	0.06
	B2N890	Slope	0.0224	0.0039	−0.0316	63.2969	9.4717	19.6473	−0.0102
		z	2.77	2.00	−4.35	0.97	0.69	1.44	−0.81
	72Q010	Slope	0.0178	−0.0019	−0.0446	44.1143	−8.2721	14.592	0.0005
		z	2.29	−0.42	−4.34	1.22	−0.38	1.15	−1.88
Eastern	All	z	1.22	−3.40	−5.02	2.92	−1.94	2.02	−2.33
	72S200	Slope	0.0337	−0.0038	−0.0156	−3.6189	−18.5293	0.4108	−0.0185
		z	3.30	−2.11	−2.82	−0.13	−1.64	0.10	−1.48
	72S590	Slope	−0.0394	−0.0129	−0.0110	59.6778	−34.4626	11.3245	−0.0639
		z	−2.56	−2.21	−3.68	1.43	−3.06	0.99	−1.72
	72T250	Slope	0.0194	−0.0045	−0.0015	34.5151	−19.3100	10.3225	−0.0404
		z	1.15	−1.65	−0.12	1.87	−1.52	1.73	−2.98
	72U480	Slope	0.0215	−0.0043	−0.0400	17.3881	−1.0984	2.1031	−0.0032
		z	1.54	−2.40	−3.38	3.68	−0.02	1.29	−0.18

,

Table 3. Trend analyses of spring climate variables of 18 meteorological stations. The common trend of a region is tested using the MK-type multivariate test. The trend of each station is tested using the modified MK test. The bold numbers indicate that the trend is significant at $\alpha =0.05$.

Region	Station	Statistic	T (°C /yr)	VPD (kPa/yr)	u2 (m/s/yr)	Rs (MJ/m2/yr)	PP (mm/yr)	ET0 (mm/yr)	Arid Index
Northern	All	z	0.69	−0.03	−4.27	3.08	0.33	2.77	0.18
	72C440	Slope	0.0138	0.0013	−0.0415	15.9971	−0.8813	3.8408	−0.0138
		z	0.90	0.75	−3.52	4.07	−0.09	4.45	−0.85
	72D080	Slope	0.0095	0.0000	−0.0050	0.5928	1.6524	0.4268	0.0330
		z	0.61	−0.08	−0.82	0.04	0.30	0.11	0.97
	82A750	Slope	−0.0272	−0.0050	−0.0249	−13.2194	6.6875	−4.6200	0.0598
		z	−0.86	−0.96	−2.00	−0.70	0.89	−1.14	1.96
	82C160	Slope	0.0223	−0.0025	−0.0568	17.7454	−0.5704	4.2049	−0.0248
		z	0.91	−0.75	−3.58	4.05	−0.14	3.22	−1.80
	K2E360	Slope	0.0188	0.0038	−0.0154	12.5493	2.007	3.6716	−0.0084
		z	1.49	1.25	−1.52	3.34	0.55	3.18	−0.26
Central	All	z	2.26	1.44	−4.14	2.55	−0.42	3.36	−1.14
	72G600	Slope	0.0312	0.0023	0.0050	15.8833	−0.7666	4.4678	−0.0153
		z	2.57	1.61	0.99	0.99	−0.26	1.57	−1.48
	72K220	Slope	0.0276	0.0006	−0.0223	−0.9159	−3.7076	0.8555	−0.0102
		z	3.03	0.20	−3.62	−0.31	−0.81	0.76	−1.01
	72M360	Slope	0.0271	0.0080	−0.0315	10.5334	1.7123	3.2939	−0.0018
		z	2.57	2.39	−3.22	1.55	0.49	1.75	−0.18
	82H840	Slope	0.0117	0.0011	−0.0039	6.0227	−5.4466	2.1120	−0.0214
		z	0.59	0.34	−1.15	1.70	−1.21	2.23	−1.40
	G2F820	Slope	0.0304	0.0070	−0.0121	18.0075	−1.4439	5.3146	−0.0172
		z	3.17	2.59	−1.18	4.40	−0.26	4.64	−1.36
	U2H480	Slope	0.0181	−0.0018	−0.0218	4.3164	−0.6457	1.0435	−0.0133
		z	1.42	−2.48	−3.37	2.19	−0.16	2.11	−0.77
Southern	All	z	1.30	1.35	−5.41	1.26	−0.25	2.20	−0.75
	B2Q810	Slope	0.0000	0.0000	−0.0503	2.1671	0.9743	0.5306	−0.0171
		z	0.00	0.00	−2.80	1.27	0.26	0.65	0.49
	B2N890	Slope	0.0184	0.0054	−0.0300	16.3192	−0.1426	4.9767	−0.0060
		z	1.67	2.22	−5.44	0.95	0.00	1.92	−0.65
	72Q010	Slope	0.0243	0.0030	−0.0460	8.5095	−2.0792	4.2178	0.0029
		z	2.08	0.83	−4.57	0.28	−0.89	0.90	−1.68
Eastern	All	z	0.04	−3.34	−4.91	1.46	−0.04	0.79	−0.45
	72S200	Slope	0.0253	−0.0050	−0.0147	−0.1023	−0.4531	−0.2667	−0.0044
		z	2.29	−2.22	−3.58	−0.01	–0.38	−0.04	−0.42
	72S590	Slope	−0.0455	−0.0118	−0.0128	9.5870	−2.1091	1.2536	−0.0155
		z	−2.61	−2.57	−3.15	0.84	−0.69	0.44	−0.85
	72T250	Slope	0.0113	−0.0057	−0.0084	12.9683	0.1131	3.1416	−0.0119
		z	1.05	−2.42	−0.66	2.59	0.08	2.07	−1.42
	72U480	Slope	0.0078	−0.0067	−0.0339	0.4366	3.0147	−0.4624	0.0147
		z	0.63	−3.01	−3.26	0.18	0.47	−0.49	0.69

,

Table 4. Trend analyses of summer climate variables of 18 meteorological stations. The common trend of a region is tested using the MK-type multivariate test. The trend of each station is tested using the modified MK test. The bold numbers indicate that the trend is significant at $\alpha =0.05$.

Region	Station	Statistic	T (°C /yr)	VPD (kPa/yr)	u2 (m/s/yr)	Rs (MJ/m2/yr)	PP (mm/yr)	ET0 (mm/yr)	Arid Index
Northern	All	z	2.33	1.82	−4.65	3.05	−0.34	3.23	−0.33
	72C440	Slope	0.0273	0.0038	−0.0313	22.5775	0.0663	6.1090	−0.0056
		z	2.65	1.88	−3.08	7.11	0.00	4.88	−0.61
	72D080	Slope	0.0198	0.0019	−0.0063	−8.8424	−5.1326	−2.7048	0.0131
		z	1.51	0.89	−1.02	−0.68	−0.40	−0.82	0.30
	82A750	Slope	−0.0034	−0.0058	−0.0289	−13.2536	−4.8894	−4.8098	0.0193
		z	−0.16	−1.02	−1.54	−0.68	−0.59	−0.94	0.61
	82C160	Slope	0.0367	0.0023	−0.0571	21.7534	−1.6133	6.2891	−0.0160
		z	1.84	0.20	−5.89	3.23	−0.34	3.58	−1.17
	K2E360	Slope	0.0345	0.0123	−0.0300	8.4488	0.7233	4.0278	−0.0103
		z	3.16	2.67	−4.35	3.32	0.02	4.33	−0.38
Central	All	z	3.15	0.01	−4.67	2.96	0.49	3.18	−0.72
	72G600	Slope	0.0276	0.0039	−0.0080	17.4792	4.7500	5.1360	−0.0140
		z	2.55	1.37	−1.94	1.08	0.36	1.00	−0.69
	72K220	Slope	0.0361	−0.0015	−0.0416	6.6182	4.685	2.4933	−0.0020
		z	3.46	−0.37	−3.38	0.90	0.30	0.99	−0.02
	72M360	Slope	0.0237	0.0100	−0.0310	3.3041	5.5972	2.1189	0.0007
		z	3.75	1.74	−2.99	1.10	0.43	1.64	0.06
	82H840	Slope	0.0004	−0.0034	−0.0084	19.4879	8.1727	5.8423	−0.0579
		z	0.00	−1.15	−2.29	7.31	0.73	4.26	−1.48
	G2F820	Slope	0.0310	0.0038	−0.0169	26.4735	5.7336	7.8225	−0.0292
		z	3.03	1.51	−1.81	2.35	0.53	3.28	−1.32
	U2H480	Slope	0.0220	−0.0038	−0.0228	3.9075	2.2013	0.9119	−0.0031
		z	2.59	−3.10	−3.12	1.40	0.15	1.13	−0.02
Southern	All	z	1.53	0.73	−5.93	1.79	−0.15	2.26	−0.87
	B2Q810	Slope	0.0120	0.0011	−0.0356	0.9970	−3.6320	0.8960	−0.0597
		z	1.09	0.61	−4.74	0.18	−0.49	0.58	−0.34
	B2N890	Slope	0.0132	0.0028	−0.0366	16.4809	5.3810	5.5894	−0.0223
		z	1.35	1.31	−4.09	1.43	0.45	1.75	−0.89
	72Q010	Slope	0.0207	−0.0031	−0.0547	18.6575	−9.0125	5.8650	−0.0113
		z	1.56	−0.36	−5.52	1.12	−0.34	1.27	−1.01
Eastern	All	z	1.55	−2.55	−5.03	2.95	−2.89	2.69	−3.02
	72S200	Slope	0.0439	−0.0024	−0.0139	5.0445	−12.4762	1.5901	−0.0416
		z	3.04	−0.90	−3.20	0.68	−1.88	0.66	−2.00
	72S590	Slope	−0.0265	−0.0123	−0.0163	27.8268	−17.7186	6.2989	−0.0914
		z	−3.35	−1.92	−3.94	1.34	−5.17	1.41	−1.95
	72T250	Slope	0.0286	−0.0041	−0.0026	17.7681	−15.8333	5.3625	−0.0533
		z	2.02	−1.09	−0.53	2.51	−2.00	2.67	−3.14
	72U480	Slope	0.0327	−0.0045	−0.0419	8.7681	−7.5373	2.1215	−0.0180
		z	1.38	−1.46	−5.53	2.94	−1.50	3.76	−2.19

,

Table 5. Trend analyses of autumn climate variables of 18 meteorological stations. The common trend of a region is tested using the MK-type multivariate test. The trend of each station is tested using the modified MK test. The bold numbers indicate that the trend is significant at $\alpha =0.05$.

Region	Station	Statistic	T (°C /yr)	VPD (kPa/yr)	u2 (m/s/yr)	Rs (MJ/m2/yr)	PP (mm/yr)	ET0 (mm/yr)	Arid Index
Northern	All	z	1.85	0.99	−3.66	2.52	0.22	2.76	−0.03
	72C440	Slope	0.0348	0.0046	−0.0515	14.3124	0.8896	3.7999	−0.0033
		z	4.00	2.45	−3.26	3.14	0.34	3.50	−0.44
	72D080	Slope	0.0421	0.0011	−0.0054	1.2075	−3.1292	0.5960	0.0019
		z	2.39	1.01	−1.06	0.20	−0.32	0.32	0.15
	82A750	Slope	−0.0168	−0.0062	−0.0141	−8.4723	14.2560	−2.7750	0.0760
		z	−1.01	−1.42	−0.99	−0.38	1.17	−0.87	1.09
	82C160	Slope	0.0268	−0.0021	−0.0732	14.4537	2.9914	3.1722	0.0012
		z	1.84	−0.71	−4.64	2.51	0.61	2.79	0.06
	K2E360	Slope	0.0263	0.0046	−0.0063	10.0347	−2.8631	3.2296	−0.0111
		z	2.02	1.10	−0.73	2.85	−0.88	7.92	−1.15
Central	All	z	2.92	2.27	−3.60	2.51	−0.49	3.30	−0.72
	72G600	Slope	0.0385	0.0038	0.0079	15.5101	−0.8418	5.3769	−0.0052
		z	3.08	2.96	1.51	1.12	−0.49	1.69	−0.65
	72K220	Slope	0.0563	0.0005	−0.0222	3.9542	−3.2175	2.3691	−0.0100
		z	3.62	0.40	−2.68	1.02	−0.78	1.46	−0.82
	72M360	Slope	0.0382	0.0125	−0.0350	7.1597	−2.6696	3.9382	−0.0066
		z	2.57	3.52	−3.30	1.95	−1.32	5.50	−1.28
	82H840	Slope	0.0206	0.0000	−0.0055	1.8700	−1.8821	0.5400	−0.0056
		z	1.80	0.00	−1.64	0.38	−0.40	0.46	−0.55
	G2F820	Slope	0.0384	0.0079	−0.0106	21.3231	−0.0025	6.7380	−0.0055
		z	2.73	2.71	−0.65	2.52	0.00	3.06	−0.53
	U2H480	Slope	0.0267	−0.0023	−0.0235	3.3565	2.0857	0.6812	0.0077
		z	1.52	−2.83	−3.52	1.91	0.71	1.19	0.58
Southern	All	z	2.35	1.47	−5.42	2.42	−0.84	2.35	−1.51
	B2Q810	Slope	0.0080	0.0019	−0.0500	15.8513	−7.0033	5.0068	−0.0241
		z	1.80	1.13	−3.24	0.00	−0.97	−0.03	−0.97
	B2N890	Slope	0.0274	0.0050	−0.0316	16.6725	−3.0893	5.0218	−0.0159
		z	2.98	2.36	−5.57	1.29	−0.73	1.53	−1.36
	72Q010	Slope	0.0259	−0.0029	−0.0435	15.8513	−1.2771	5.0068	−0.0152
		z	2.35	−1.33	−4.31	1.78	−0.20	1.55	−1.00
Eastern	All	z	0.82	−2.90	−4.14	2.57	−0.47	1.81	−0.78
	72S200	Slope	0.0373	−0.0033	−0.0159	2.1030	−0.1292	0.5913	−0.0263
		z	2.55	−1.70	−3.74	0.43	0.00	0.44	−0.41
	72S590	Slope	−0.0420	−0.0150	−0.0055	14.5092	−9.7230	2.1243	−0.0634
		z	−3.85	−1.93	−1.37	1.52	−1.17	0.61	−1.32
	72T250	Slope	0.0224	−0.0050	−0.0022	10.4072	0.9208	2.8632	−0.0247
		z	1.68	−1.51	−0.15	2.37	0.14	2.40	−0.45
	72U480	Slope	0.0180	−0.0032	−0.0442	4.3741	−8.2595	0.3188	−0.0211
		z	1.03	−1.25	−3.38	1.66	−0.55	0.30	−0.38

and

Table 6. Trend analyses of winter climate variables of 18 meteorological stations. The common trend of a region is tested using the MK-type multivariate test. The trend of each station is tested using the modified MK test. The bold numbers indicate that the trend is significant at $\alpha =0.05$.

Region	Station	Statistic	T (°C /yr)	VPD (kPa/yr)	u2 (m/s/yr)	Rs (MJ/m2/yr)	PP (mm/yr)	ET0 (mm/yr)	Arid Index
Northern	All	z	0.46	−1.31	−3.74	2.16	0.32	0.52	0.42
	72C440	Slope	0.0117	−0.0028	$-$0.0373	10.1099	1.4158	0.1685	0.0026
		z	0.67	−1.57	−2.75	3.18	0.46	0.12	0.17
	72D080	Slope	0.0211	0.0000	−0.0050	1.6807	−2.9744	0.4835	−0.0132
		z	1.38	0.18	−0.74	0.47	−0.66	0.43	−0.41
	82A750	Slope	−0.0526	−0.0025	−0.0174	−9.0904	6.0176	−3.4102	0.1324
		z	$-$1.31	−1.10	−1.02	−0.98	1.13	−1.61	2.27
	82C160	Slope	0.0207	−0.0018	−0.0720	10.6727	0.2100	1.4448	−0.0006
		z	1.25	−1.40	−3.26	3.18	0.02	1.64	0.00
	K2E360	Slope	0.0074	−0.0009	−0.0048	8.3438	0.6833	1.5369	0.0007
		z	0.38	−0.72	−0.81	2.63	0.34	2.26	0.02
Central	All	z	1.37	0.38	−3.60	2.69	−0.16	3.14	−0.66
	72G600	Slope	0.0219	−0.0002	0.0118	13.7243	−0.5242	2.8750	−0.0051
		z	1.05	−0.26	2.08	1.69	−0.36	2.20	−1.17
	72K220	Slope	0.0393	0.0011	−0.0211	3.7026	0.6141	1.6703	$-$0.0008
		z	2.17	0.81	−3.22	0.54	0.54	1.54	−0.14
	72M360	Slope	0.0187	0.0036	−0.0374	8.9057	0.4833	2.4898	0.0011
		z	0.97	1.56	−3.84	1.64	0.67	2.81	0.50
	82H840	Slope	0.0088	−0.0008	−0.0044	−3.2119	−0.6905	−0.9288	−0.0024
		z	0.30	−0.68	−1.09	−0.76	−1.03	−0.87	−0.59
	G2F820	Slope	0.0308	0.0033	−0.0133	13.4413	−0.4618	3.9408	−0.0063
		z	1.68	1.44	−0.68	4.50	−0.38	6.25	−1.28
	U2H480	Slope	0.0212	−0.0013	−0.0234	7.0263	−0.5671	1.6263	−0.0105
		z	1.05	−1.99	−3.72	2.87	−0.26	4.60	−0.81
Southern	All	z	0.54	0.67	−4.95	2.36	0.44	2.20	−0.44
	B2Q810	Slope	0.0035	0.0010	−0.0443	0.0951	1.5446	0.0244	−0.0034
		z	0.36	0.49	−4.11	0.00	1.03	0.02	1.05
	B2N890	Slope	0.0180	0.0020	−0.0264	16.3217	−0.1795	3.8575	−0.0023
		z	0.79	1.71	−4.07	1.49	−0.08	2.24	−1.39
	72Q010	Slope	0.0075	−0.0024	−0.0329	10.7046	0.1898	2.4606	0.0036
		z	0.43	−1.13	−3.50	2.19	0.14	1.39	−1.24
Eastern	All	z	0.11	−3.52	−4.77	2.26	1.09	1.28	0.16
	72S200	Slope	0.0280	−0.0015	−0.0195	0.9660	1.2325	0.8138	0.0041
		z	2.14	−5.18	−11.71	0.30	1.93	0.38	0.71
	72S590	Slope	−0.0548	−0.0133	−0.0092	9.4480	1.8243	0.4558	0.0027
		z	−3.95	−2.07	$-$2.41	0.81	1.54	0.16	0.22
	72T250	Slope	0.0139	−0.0040	−0.0026	9.3053	−2.0009	1.8083	−0.0229
		z	0.87	−2.32	−0.38	1.80	−1.30	1.88	−1.68
	72U480	Slope	0.0100	−0.0033	−0.0417	3.6258	9.4475	0.0444	0.0539
		z	0.71	−2.26	−3.64	1.96	1.48	0.06	2.41

; Figure S1). This anomaly was attributed to the station’s geographical location on the windward slope in the northeastern corner of Taiwan (Figure 1), where unique meteorological conditions resulted in significantly higher rainfall compared to other stations (Table 1). In terms of the magnitude of the trend, the Sen’s slopes for annual ET₀ ranged from $-$18.56 to 26.21 mm/yr, with the minimum and maximum values recorded at Stations 82A750 and G2F820, respectively (Table 2). Annual and seasonal ET₀ values indicate that upward trends are dominant, with more than 80% of positive slopes for the annual period and each season (Figure 3), and the statistically significant increasing trends ranged from 33.3% to 38.9% across different temporal units.

Changes in PP and four other climate variables related to ET₀ (i.e., T, R_s, u₂, and VPD) were also observed, as shown in Table 2, Table 3, Table 4, Table 5 and Table 6 and Figures S2–S6. The trend in PP mainly decreased but was not significant in the spring, autumn, and annual temporal units (Table 2, Table 3, Table 4, Table 5 and Table 6; Figure S6). A general upward trend of T and R_s was observed in more than 80% of the series, with over 25% of the series showing statistically significant trends (p < 0.05), except for T in winter (Figure 3 and Figures S2 and S3). Regarding u₂, decreasing trends were predominant (about 95%), and over 50% of the series showed a significant decreasing trend (Figure 3 and Figure S4). The trend in VPD did not show a clear directionality; i.e., the proportions of increasing and decreasing trends were similar in each temporal unit, except for winter (Figure 3 and Figure S5). Considering the relative impact and trend of each climate variable, the recent increase in ET₀ in Taiwan may be attributed to the increases in T and R_s.

3.3. Findings from Multi-Station Trend Analysis

Although the trend analysis for each station can show the detailed trend at each location, it is difficult to interpret the massive amount of results [47]. Therefore, considering the station-wise and multi-station analyses cohesively may capture a complete and accurate picture of the trend characteristics [24]. Multi-station trend analyses using the CE method indicate that annual trends of T and R_s in the four geographical regions of Taiwan have increased (Table 2). Conversely, the annual trend of u₂ decreased significantly (p < 0.05), and the PP showed a non-significant decreasing trend. Significant downward trends in annual VPD were found for the eastern region. For the seasonal series, u₂ in the four geographical regions showed significant changes (Table 3, Table 4, Table 5 and Table 6).

The annual ET₀ series increased significantly in each geographic region of Taiwan (Table 2). The seasonal ET₀ series showed an increasing trend for all seasons and geographic regions (Table 3, Table 4, Table 5 and Table 6). The non-significantly increasing trends of the seasonal ET₀ series mainly occur in the eastern region, possibly because the corresponding T trends are not significant. On the contrary, annual and seasonal PP in Taiwan mainly decreased (Table 2, Table 3, Table 4, Table 5 and Table 6). Moreover, PP increased in central Taiwan in the summer (Table 4) but decreased in other seasons (Table 3, Table 5 and Table 6), indicating that PP in this region may become more uneven. With increasing evapotranspiration and insufficiently increasing PP, it is expected that Taiwan’s water resources may become scarce.

Generally, the results obtained by the CE method are consistent with the results of the station-wise analysis (Table 2, Table 3, Table 4, Table 5 and Table 6; Figure 3 and Figures S1–S6). However, a deeper examination of the multi-station analysis reveals that the particularities of different stations within the same region cannot be ignored. For example, the winter u₂ in central Taiwan shows a significant downward trend, but the trend at Station 72G600 is significantly upward (Table 6). This study also compares the results of multi-station analyses using the UM and CE methods. It was found that the results of the UM and CE methods were inconsistent (i.e., the significance or direction of the trend was different) in 36 out of 140 cases (7 variables × 4 regions × 5 temporal units) (Tables S1–S5). Obviously, there are many possibilities for the two multi-station analysis methods to yield different conclusions.

3.4. Trends of the Arid Index for the Geographic Regions in Taiwan

In addition to analyzing the trends of ET₀ and PP individually, this study also examined the effect of climate change on the arid index to provide a more comprehensive view of the possible impact on crop production. The net irrigation requirements (NIR) for crop production can be roughly estimated by subtracting PP from the crop water demand (Equation (10)) [53]. If PP is less than the crop water demand, artificial irrigation is required to provide the necessary water for crop growth. After simple derivation (Equations (11)–(13)), it can be inferred that the arid index represents the upper bound of the K_c of crops suitable for planting in the region under rainfed cultivation. When the K_c of a crop is higher than the arid index, additional water resources must be provided to support its cultivation. It was found that the values of the seasonal and annual arid index mainly showed a downward trend, except for winter (Table 2, Table 3, Table 4, Table 5 and Table 6; Figure 3). These results imply that Taiwan would have fewer crop options suitable for cultivation from a water resource perspective.

$$\mathrm{N}\mathrm{I}\mathrm{R}={\mathrm{E}\mathrm{T}}_{\mathrm{c}}-\mathrm{P}\mathrm{P}={\mathrm{K}}_{\mathrm{c}}{\mathrm{E}\mathrm{T}}_0-\mathrm{P}\mathrm{P}$$

(10)

$$0\ge \mathrm{N}\mathrm{I}\mathrm{R}={\mathrm{K}}_{\mathrm{c}}{\mathrm{E}\mathrm{T}}_0-\mathrm{P}\mathrm{P}$$

(11)

$$0\ge {\displaystyle \frac{{\mathrm{K}}_{\mathrm{c}}{\mathrm{E}\mathrm{T}}_0}{{\mathrm{E}\mathrm{T}}_0}}-{\displaystyle \frac{\mathrm{P}\mathrm{P}}{{\mathrm{E}\mathrm{T}}_0}}={\mathrm{K}}_{\mathrm{c}}-\mathrm{A}\mathrm{r}\mathrm{i}\mathrm{d} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}$$

(12)

$${\mathrm{K}}_{\mathrm{c}}\le \mathrm{A}\mathrm{r}\mathrm{i}\mathrm{d} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}$$

(13)

4. Discussion

4.1. Comparison of the Effects of Climate Change in Taiwan and Other Regions

In this study, a general upward trend in ET₀ was observed (Figure 3). Similar increasing trends in annual ET₀ have been reported in studies analyzing the effects of climate change in Romania (0.69 mm/year) [7], Slovenia (3.27 mm/year) [9], and Spain (2.45 mm/year) [15]. In China, annual ET₀ has shown an increasing trend of 2.24 mm/year over the past two decades [17]. Comparing the annual ET₀ trends in Taiwan over the past 20 years (Table 2) with those studies, it is obvious that the trend magnitude in Taiwan is larger. Wang et al. [12] and Fan et al. [19] pointed out that the trend of ET₀ under climate change varies in direction and magnitude depending on regional climate differences. A study in China found that provinces with a warm temperate monsoon climate experienced a greater increase in drought hazards compared with those having a continental monsoon climate [54]. Therefore, the larger ET₀ trend observed in Taiwan is likely due to its warm climate. Regarding u₂, a predominant decreasing trend was observed in Taiwan, with approximately 95% of the series showing a declining tendency, and over 50% exhibiting a statistically significant decreasing trend (Figure 3 and Figure S4). Significant reductions in wind speed due to climate change have also been reported in numerous studies [9,17,19,55].

In South Africa, the frequency of dry years in rainfed agricultural systems has increased dramatically since 2010, with larger areas of irrigated cropland affected by drought since 2012 [56]. In Iran, agricultural production has declined due to droughts, with total production expected to decrease by 6% to 18% [57]. However, de la Casa and Ovando [18] found that in central Argentina, annual ET₀ did not change significantly between 1941 and 2010 in 91.4% of the regions, while increased rainfall was observed in 54.4% of the regions. Consequently, climate change has led to a greater water availability for agricultural production in central Argentina. Comparing the ET₀ and PP trends in Taiwan (Table 2, Table 3, Table 4, Table 5 and Table 6; Figure 3) with those in other regions highlights the site-specific nature of climate change impacts. Given the significant variations in climate, topography, and vegetation worldwide, localized research and tailored policy frameworks are essential for effectively solving climate-related challenges at the regional level [58].

In addition to statistical trend analysis, climate models are a major approach in climate change research. Early climate models focused solely on the fundamental physics of the atmosphere and ocean, but they have been expanded in complexity to the Earth System Models. These models now simulate not only the physical aspects of the climate system but also the biogeophysical and chemical interactions within the fully coupled atmosphere-ocean-land-sea ice-land ice system [59]. Future research could integrate trend analysis with climate models to more accurately assess the progression and impacts of climate change.

4.2. Changes in Important Climate Variables and Their Possible Impacts on Crop Production

In this study, a general upward trend of T was observed (Figure 3), which may affect rice production. According to research [60], global warming may result in a 10% to 15% yield loss in major rice-producing regions of Asia. Specifically, a 1 °C increase in nighttime temperature can reduce yield by 10%, and for daytime temperatures between 28 °C and 34 °C, each 1 °C increase can result in a yield decrease of up to 7% to 8%. The primary reason for this reduction is the shortened vegetative growth period due to increased temperatures, leading to insufficient biomass accumulation and poorly filled grains. Additionally, changes in the nutritional content of grains should be a concern. Research indicates that with increased atmospheric temperatures and higher carbon dioxide concentrations, the nutrients of rice will notably decline, including protein, micronutrients, and vitamin B [61]. Therefore, increasing nutritional value through breeding, which is known as biofortification, is one of the future directions for crop breeding to contend with climate change [62]. Furthermore, the shortening of low-temperature durations and insufficient chilling in winter adversely affects the flowering and fruiting of deciduous fruit trees. This also impacts honey production, as bees lack flowering sources. To cope with future warming and unpredictable climate conditions, it is recommended to accelerate the breeding of low-chilling fruit tree varieties [63] or to replace temperate fruit trees with tropical evergreen fruit trees for cultivation [64].

This study identified that the VPD in eastern Taiwan decreased significantly (Table 2, Table 3, Table 4, Table 5 and Table 6). With the increase in T in eastern Taiwan, the decrease in VPD implies that RH increased in this region. In addition, a predominant decreasing trend of u₂ was observed in this region (Table 2, Table 3, Table 4, Table 5 and Table 6). Warm, moist, and poorly ventilated environments are generally recognized as increasing the risk of pests and plant diseases. Therefore, pest and disease control in crop production in eastern Taiwan needs attention. Monoculture systems lose resilience to climate-related disasters and may lead to large-scale outbreaks of pests and diseases. Small-scale and diversified cropping systems are recommended to reduce production risks while maintaining biodiversity [65].

Annual and seasonal ET₀ values indicate that upward trends are dominant, with more than 80% of positive slopes for the annual period and each season (Figure 3). The increasing evaporative demand will cause net volume loss in natural lakes and storage losses in reservoirs [66]. In addition, due to the increase in evapotranspiration, the water demand for crop production will also rise [31]. Besides, it was found that the values of the seasonal and annual arid index mainly showed a downward trend (Figure 3), suggesting a shift toward drier conditions. A trend of more severe and frequent droughts globally could be predicted [67,68,69,70,71]. These findings imply that fewer crop options may be suitable for cultivation due to water resource constraints. Relocating cultivation areas to mitigate exposure to adverse climate conditions could serve as a theoretical adaptive response. For example, Sloat et al. [72] demonstrated that global rainfed crops have shifted toward regions with more favorable temperatures. Additionally, the negative impacts of warming on these crops have been significantly mitigated through migration and the expansion of irrigation.

Su et al. [73] revealed that high temperatures and dry climate have been the primary climate patterns in Taiwan since 2012. In addition, Su and Kuo [74] pointed out that central and southern Taiwan have a higher risk of heavy rain damage from mid-May to June compared to other regions, and the damage could be mitigated by reducing the cropping areas during high-risk periods. If farmers want to cultivate crops during high-risk periods, it is recommended to choose lodging-resistant varieties [74] or cultivate the crops in greenhouses [75]. Conversely, to produce crops in the dry season, it is necessary to reserve sufficient water during the wet season and establish a complete irrigation system to supply water during the dry season.

In Taiwan, the utilization of water resources primarily relies on surface water. Due to the temporal and spatial variabilities of PP, hydraulic facilities are required to regulate water distribution. However, the regulation capacity of these hydraulic facilities is limited, and sediment accumulation further reduces their capacity. Moreover, during drought periods, the competition among different water use sectors—domestic, agricultural, and industrial—makes water allocation even more challenging. Under these situations, decision-makers should allocate water resources appropriately, and farmers should cultivate drought-tolerant crops and cultivars, as well as adopt water-saving management practices, such as the alternate wetting and drying technique of rice cultivation [76] and regulated deficit irrigation [77,78] in the dry season, to avoid economic losses.

4.3. Comparison of Different Multi-Station Analysis Methods

In this study, the results obtained by the CE method are generally consistent with the results of the station-wise analysis (Table 2, Table 3, Table 4, Table 5 and Table 6; Figure 3 and Figures S1–S6). However, the particularities of different stations within the same region cannot be ignored. In order to address the discrepancies between the results of station-wise and multi-station analyses, it is recommended to group stations based on both geographical and meteorological characteristics [79,80], thereby increasing the homogeneity of stations in each group for multi-station analysis.

When reviewing previous studies that conducted trend analyses involving multiple stations, we found that several studies still used the UM approach instead of the multivariate one [6,17,19]. This study compares the results of multi-station analyses using the UM and CE methods. It was found that the results of the UM and CE methods were inconsistent (Tables S1–S5). The benefit of the UM method is its simplicity and ease of implementation; however, it has two possible drawbacks, in our opinion. First, the arithmetic mean is not an appropriate measure of central tendency when outliers exist. Second, the UM method ignores the spatial dependence between the stations. In the domain of geography, there is a core principle that “everything is related to everything else, but near things are more related than distant things” [81]. This principle implies that many geographic phenomena may be highly sensitive to the relationships among nearby things. Because the distances between meteorological stations are not the same, calculating the arithmetic mean with equal weight for each station is unsuitable to represent the reference characteristics of the region. Conversely, the multivariate method considers the unique variation of each station and the covariance between stations [47]. In the CE method, the multi-station data are transferred to a multivariate setting where each station represents one variable, and then the tests account for the dependence between variables. We think that the multivariate approach is more appropriate when investigating the common trend across multiple stations. However, the superiority of the multivariate method needs to be validated with more rigorous simulation studies based on statistical power.

5. Conclusions

Temporal changes in ET₀ and PP may have important hydrological, ecological, and agricultural consequences. Thus, this study analyzed the trends in ET₀, PP, and several important climate variables in Taiwan during 1995–2022. It was found that T, VPD, u₂, and R_s were all positively correlated with ET₀. According to the importance evaluation using RF, ET₀ was mainly impacted by R_s, followed by T, u₂, and VPD in Taiwan. The station-wise analysis revealed increasing annual and seasonal T, R_s, and ET₀ values, whereas u₂ decreased. Considering the relative impact and the trend of each climate variable, the increase in ET₀ in Taiwan in recent years may be attributed to the increase in T and R_s. The multi-station analysis indicated an increasing trend in seasonal ET₀ across all seasons and geographic regions, suggesting a rising water demand for crop production. Eastern Taiwan showed significantly decreasing trends in annual VPD and u₂, highlighting the need for enhanced pest and disease control in crop production in this region. Additionally, the annual and seasonal PP values slightly decreased in Taiwan, with the distribution of PP in central Taiwan becoming more uneven. However, in multi-station analysis, the unique characteristics of individual stations within the same region cannot be overlooked. Additionally, a comparison of different multi-station analysis methods revealed that the CE and UM methods often produced inconsistent results. Therefore, further research is needed to determine the superiority of these multi-station analysis methods. Finally, this study investigated the effect of climate change on the arid index and its possible impact on crop production. The seasonal and annual arid index values exhibited predominantly a downward trend in all seasons, except in winter. Consequently, drier conditions are expected in the future. Therefore, it is recommended that policies in Taiwan should focus on improving cultivation methods for water conservation and enhancing hydraulic infrastructure and irrigation systems in agricultural production.