Temperature vegetation dryness index (TVDI) for drought monitoring in the Guangdong Province from 2000 to 2019

Study area

Guangdong Province, located in southern China (longitude 109°39′∼117°19′East, latitude 20°08′∼25°32′North)(Fig. 2), encompasses temperate, subtropical, and tropical climatic zones. It registers an annual mean temperature of 19–23 °C. Guangdong benefits from extensive solar irradiance, manifesting an with an average daily sunshine duration of 1745.8 h and an annual total solar radiation between 4,200–5,400 MJ/m2, ensuring ample photothermal resources. The province’s topography exhibits higher elevations in the south compared to the north, influencing the spatiotemporal distribution of precipitation. Notably, between April and September, over 80% of the yearly rainfall is recorded. This period witnesses substantial inter-annual variability, where precipitation during pluvial seasons can surpass twice that of arid years. Guangdong is recurrently subjected to climatic adversities such as floods, droughts, and typhoons. Additionally, meteorological events like springtime chill and rainfall, autumnal cold dew winds, and late autumn to early spring frost events and cold surges are prevalent. The province’s drought-afflicted zones vary annually, spanning from a few ten thousand to several hundred thousand hectares, with particularly arid years impacting over a million hectares. By the decade’s close in the 1990s, recurrent droughts affected approximately 24% of the province’s cultivated area, approximating 2.3 million hectares.

The boundaries of the study area were sourced from China National Catalog Service For Geographic lnformation (https://www.webmap.cn/commres.do?method=result100W) 1:1 million public version of basic geographic information data (2021) the boundry of the study area; the Digital elevation Model Map of the study area is derived from the Shuttle Radar Topography Mission (SRTM), version V4.1 jointly measured by the United States Space Agency (NASA) and the National Mapping Agency (NIMA) of the Department of Defense. Download from the Chinese Academy of Sciences Geographic Spatial Data Cloud: https://www.gscloud.cn.

Data source and preprocessing

The traditional TVDI model

The Temperature Vegetation Dryness Index (TVDI) serves as a pivotal tool in remote sensing drought monitoring. It appraises the conditions of soil surface moisture conditions based upon the spatial features of a two-dimensional scatter plot similar to a triangle, created from the surface temperature (LST) and vegetation index (VI) data. This LST-VI feature space is a scatter plot of LST and VI values for a given pixel in a two-dimensional coordinate system (Fig. 3).

Figure 3 illustrates the triangle’s vertices represent three extreme situations in the LST-VI feature space: the upper left corresponds to to arid, barren soil (low VI values and high LST values), the lower left represents wet soil (minimum VI and LST values in that region), and the bottom right vertex symbolizes the most saturated soil with highest vegetation cover (high VI values and low LST values). For bare soil areas, changes in surface temperature are highly correlated with soil moisture changes. From the top left corner to the bottom left corner of the feature triangle, the amount of surface soil water evaporation gradually increases from zero to the maximum value. The trajectory from the upper left to the lower left of this triangle indicates an incremental ascent in surface soil water evaporation. The bottom right vertex represents the surface soil being completely covered by vegetation and having sufficient water content, with a water stress index of 0 in this region (Carlson, Gillies & Perry, 1994). The triangle’s oblique side signifies a compromised soil moisture efficiency, and minuscule surface water evaporation, thereby being characterized as a “dry edge” state during drought periods. Conversely, the base suggests ample soil moisture content, unrestrictive to plant growth, with the surface evapotranspiration being equated to potential evapotranspiration, thus falling within a “wet edge” state (Price, 1990).

By simplifying the feature space, Sandholt constructed the Temperature Vegetation Dryness Index (TVDI), the mathematical expression is shown below: (1)${\displaystyle TVDI=\frac{T_s-T_{s_{min}}}{T_{s_{max}}-T_{s_{min}}}}$

Eq. (1), T_s represents the surface temperature value (°C) of any pixel on the remote sensing image, T_smin represents the minimum surface temperature value of all grid values under the same vegetation cover condition (with a constant VI value), and T_smax represents the corresponding maximum surface temperature value under the same VI value.

From the remote sensing data amalgamating LST and VI data, the maximum and minimum surface temperature values for the same vegetation index are extracted, and the vegetation indices are linearly fitted to obtain the dry edge (Eq. 2) and wet edge (Eq. 3) equations in the feature space: (2)${\displaystyle T_{s_{min}}=a_1+b_1\times VI}$ (3)${\displaystyle T_{s_{max}}=a_2+b_2\times VI}$

where, a1 and b1 are the coefficients of the wet edge equation, while a2 and b2 are the coefficients of the dry edge equation.

According to the principle of TVDI, the dry edge corresponds to a TVDI value of 1, while the wet edge corresponds to a TVDI value of 0, and the TVDI value of any point is calculated to be between 0 and 1. The higher the TVDI value, the more severe the corresponding soil drought; the lower the TVDI value, the lighter the corresponding soil drought. The TVDI drought level is classified as wet (0∼0.2), normal (0.2∼0.4), mild drought (0.4∼0.6), moderate drought (0.6∼0.8), and severe drought (0.8∼1.0).

Data reconstruction to the TVDI model

The NDVI, EVI, and LST datasets, once processed for quality control, exhibit missing values in raster cells deemed to have unreliable quality. To build a consistent time series, it is imperative to optimize and smooth these continuous raster values over temporal dimensions. In this study, the Inverse Distance Weight (IDW) interpolation technique is employed to estimate values for the missing raster cells, facilitating spatial data reconstruction. Subsequently, the Savitzky–Golay (S–G) time series filtering approach is applied to refine the temporal data series, ensuring comprehensive time reconstruction of the datasets.

DEM correction to the TVDI model

The accuracy of TVDI in drought assessment is intrinsically related to LSTand V). However, it is also influenced by topographical elevation and geographical latitude, two paramount factors that dictate atmospheric background variance and solar radiation when acquiring LST data via remote sensing instrumentation. Such influences can precipitate direct discrepancies in TVDI computation. Consequently, rectifying LST data is essential to enhance the computational accuracy of TVDI. The formula for correcting LST data is given in (Eq. 9): (9)${\displaystyle T_c=T_s+a\times H+b\times L+c}$

where, T_s and T_c represent the temperature values before and after correction, respectively. H is the elevation, L is the latitude, a is the elevation correction coefficient, and b and c are the latitude correction coefficients. Figure 4 shows the DEM corrected LST data comparison.

The LST data is derived from the NASA EOSDIS Land Processes Distributed Active Archive Center (https://doi.org/10.5067/MODIS/MOD11A2.006).

Assessment method

This study validates the accuracy of the remote sensing-based TVDI by comparing it with the widely used meteorological drought indices, namely Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI). McKee, Doesken & Kleist (2012) leveraged historical records to delineate the probabilistic distribution of cumulative precipitation across temporal intervals, introducing the SPI. This innovation addressed the challenge of juxtaposing standalone rainfall data across disparate spatial–temporal scales and subsequently garnered global acclaim for its efficacy in drought surveillance. Augmenting the SPI framework, Vicente-Serrano, Beguería & López-Moreno (2010) integrated both precipitation and evapotranspiration metrics, heralding the SPEI, a tool proficient in appraising drought scenarios across diverse intervals. Its universal application in global drought research and praxis is attested by subsequent works (Noorisameleh et al., 2020; Shi, Wu & Ding, 2020). Given the multifaceted temporal scales of SPI and SPEI (encompassing 1-month, 3-month, 6-month, 12-month, and beyond), our study has elected January as the representative temporal metric, harmonizing it with the monthly granularity of TVDI.

For accuracy assessment, this study employed the Pearson Correlation Coefficient and F-test. The Pearson Correlation Coefficient, renowned for gauging the linearity between paired variables (X and Y), with values ranging between -1 and 1. Complementing this, the F-test scrutinizes the hypothesis suggesting that a given statistical figure adheres to an F distribution under the null premise (H0). Historically, the Pearson Correlation Coefficient traces its roots to Karl Pearson’s adaptation of a concept posited by Francis Galton in the late 19th century. Concurrently, the F-test, occasionally referenced as the variance ratio test, assesses variance uniformity, producing results consistent with the F-distribution under the null premise (H0).

The calculation formula for the Pearson correlation and F-test are Eqs. (10) and (11): (10)${\displaystyle r=\frac{\sum _{i=1}^n\left(X_i-\overline{X}\right)\left(Y_i-\overline{Y}\right)}{\sqrt{\sum _{i=1}^n{\left(X_i-\overline{X}\right)}^2}\sqrt{\sum _{i=1}^n{\left(Y_i-\overline{Y}\right)}^2}}}$ (11)${\displaystyle F=\frac{1}{n-1}\sum _{i=1}^n{\left(X_i-\overline{X}\right)}^2/\frac{1}{n-1}\sum _{i=1}^n{\left(Y_i-\overline{Y}\right)}^2}$

TVDI_{NDV I} and TVDI_{EV I} construction results

LST-VI feature space construction

In this study, IDW and Savitzky–Golay (S–G) Filtering were applied to reconstruct NDVI, EVI and LST data with the parameters set as shown in the table below, while the parameters for topography correction are shown in Table 2 below:

Table 2:

Data reconstruction parameter setting.

	Parameter	Setting value
IDW	Distance parameters	2
	Variable retrieval radius value	12
S-G	Half-width of filter window m	7
	Polynomial fitting order d	2
Terrain correction	Elevation correction coefficient a	0.006
	Latitude correction factor b	0.4
	Latitude correction factor c	−16

DOI: 10.7717/peerj.16337/table-2

In previous studies, the range of 0.2 < NDVI < 0.8 was directly selected to fit the dry and wet edges of the characteristic space when calculating TVDI values, because NDVI is more sensitive to changes in soil background, and when NDVI < 0, the soil surface layer is mainly water or snow, and surface moisture can be considered 100%. When vegetation cover is less than 20%, NDVI values are difficult to indicate the degree of vegetation cover in the area. When vegetation cover is greater than 80%, the increase in NDVI value will be delayed and reach saturation, resulting in reduced sensitivity to vegetation cover recognition (Sha et al., 2011). However, for Guangdong Province, which has a large population density and wide urban coverage, some extreme cells (such as urban green belts, etc.) will appear in the range of 0.2∼0.8, and these very few cells need to be eliminated in order to make the dry and wet edge fitting equation more reasonable. Moreover, the principle of enhanced vegetation index EVI and NDVI indicating vegetation coverage is different, and the interval fitting of 0.2∼0.8 cannot be directly selected, and the fitting range needs to be selected according to the actual situation of the spatial distribution of the features. In addition, the results show that the pixel histogram (the frequency distribution of pixels with different VI values) is closely related to the shape of the feature space, and plays an important auxiliary role in the determination of dry and wet edges in the feature space (Zhang & Feng, 2012).

Based on this, by extracting the maximum and minimum land surface temperature values corresponding to different NDVI/EVI pixel grid values, the distribution of LST-VI feature space in Guangdong Province from September 2000 to December 2019 is obtained on a monthly basis. Twelve months of feature space and corresponding VI pixel histograms are selected for display, representing the distribution of feature space in different years for each of the 12 months (Figs. 5 and 6).

Figure 5 delineates the LST-NDVI feature space, where the x-axis represents NDVI values, and the y-axis represents LST (°C). The shape of the feature space, demarcated by the dry and wet edges, can be characterized as a tapered spindle with acuminated termini. Transitions along the dry and wet edges can be segmented into three distinct phases. Initially, the LST value of the points on the dry edge gradually increases from NDVI = 0, conversely, those on the wet edge descend. Subsequently, with escalating NDVI values, the data points on both edges expand laterally. In the third stage, the points on the dry edge decrease with the increase of NDVI values, while those on the wet edge decrease correspondingly, and finally converge at the maximum NDVI value. Considering the corresponding histogram, the points in the first stage have extremely few pixels and can be ignored when fitting the dry and wet edges. The points in the second and third stages should be reasonably removed according to the distribution of the histogram. Between January and March, the NDVI histogram exhibits a broad distribution, predominantly spanning 0.35 to 0.75, indicative of sparse vegetative cover. Progressing through the months, NDVI value ranges migrate rightward, peaking beyond 0.9, thus extending the LST-NDVI feature space into higher NDVI sectors. By August, the NDVI histogram primarily encompasses a 0.65–0.85 range, manifesting a more pronounced and integral triangle in the feature space. From September, the histogram range shifts to the left, demonstrating significant seasonal characteristics.

Figure 6 illustrates the LST-NDVI feature space, the LST-EVI feature space also exhibits a spindle shape, but the shape changes significantly with the month. From May to October, it is a more elongated spindle shape, while from November to April, it is a shorter and rounder spindle shape. The range of the feature space fluctuates more than that of LST-NDVI, with the maximum value in January and February not exceeding 0.5, reaching almost 0.8 in July and August, and then shifting to around 0.6 in November and December. Combining with the histogram, in the warmer climate from May to October, the EVI values are concentrated between 0.45 and 0.7, while from November to April, they are distributed between 0.35 and 0.6.

Pixels with exceptionally high or low NDVI/EVI values are typically sparse, representing unique soil moisture conditions, making it difficult to ensure that there are different corresponding values from dry to wet at that NDVI. Therefore, the values of the pixels at both ends are extremely unreliable and cannot be used for fitting the dry edge and wet edge (Sha et al., 2011; Zhang & Feng, 2012). When fitting the dry and wet edges, the VI interval with more pixels and after the critical point in the feature space should be selected for fitting. The analysis of the feature spaces constructed by the two vegetation indices shows that there are obvious critical points, such as the NDVI = 0.55 in Fig. 5C, where the values on the dry edge begin to decrease gradually and the values on the wet edge increase gradually. The pixel volume corresponding to NDVI = 0.55 in the histogram also stands robust, making edge fitting from this value onward align seamlessly with TVDI principles.

LST-VI dry and wet edge fitting

In Guangdong Province, the LST-VI feature space exhibits pronounced monthly variations and distinct segmentation. Therefore, it is necessary to adjust the fitting of the dry and wet edges based on the results analyzed in the previous section. During the interval spanning September 2000 to December 2019, the dry and wet edges were reformulated by excluding anomalous NDVI and LST values. For this refinement, the range 0.55 < NDVI < 0.9 was employed within the LST-NDVI feature space. The least squares regression technique was utilized to fit the dry and wet edges, with select outcomes presented in Figs. 7 and 8.

Figurs 7 and 8 illustrate that by setting the fitting range according to the specific characteristics of the feature space in different months in Guangdong Province and removing the points with fewer pixels at both ends of the histogram, the fitting of the dry-wet edge equation was optimized to ensure that there are continuous pixel values from dry to wet for each VI value. The R² coefficient for the wet edge fitting across various intervals ranges from 0.5 to 0.9, with the mean R² coefficient for the dry edge being 0.6. This results in a characteristic triangular configuration marked by an inverse relationship in the dry edge and a direct relationship in the wet edge for the dry-wet edge fitting curve. This restructured dry-wet edge feature space offers a notable improvement over the initial space delineated in the earlier segment.

TVDI_{NDV I} and TVDI_{EV I} selection

The SPI and SPEI values were computed using monthly precipitation data from 22 meteorological stations in Guangdong Province from 2000 to 2019. The 1-month SPI values were calculated, and the average of the 22 station values was used to represent the province-level 1-month SPI situation. The TVDI_{NDV I} and TVDI_{EV I} values of the corresponding pixels for the 22 meteorological stations from 2000 to 2019 were extracted. Correlation analysis and F-tests were then conducted by comparing these values with the SPI and SPEI values for each meteorological station.

Taking the Dongyuan meteorological station as an example, the R-value in Fig. 9 represents the correlation coefficient, and P represents the P-value of the F-test. According to the calculation principle of TVDI, the higher the value, the more significant the drought. In contrast, the meteorological drought index SPI/SPEI is the opposite, with lower values indicating more severe drought. Therefore, the two should show a negative correlation. The Pearson correlation coefficients between the monthly N-TVDI and SPI-1 from September 2000 to December 2019 at the Dongyuan meteorological station were −0.6025, and −0.231 with SPEI-1, both passing the significant test of p < 0.001. To compare the strength of the correlation between N-TVDI/E-TVDI and meteorological drought indices, the correlation coefficients and P-values of TVDI and SPI-1/SPEI-1 at 22 meteorological stations were plotted in a table (Table 3).

Table 3 indicates that among the 22 meteorological stations, TVDI_{NDV I} was negatively correlated with SPI in 20 stations with exceptions being Huiyang and Xuwen.The Pearson correlation coefficient for the association between TVDI_{NDV I} and SPI in Lianxian did not pass the significance test of P < 0.05. Furthermore,16 stations passed the significance test of P < 0.001. The Pearson correlation coefficient ranged from −0.23 to −0.66, with an average of −0.44. TVDI_{NDV I} was negatively correlated with SPEI in 21 stations (excluding Luoding). Four stations (Huiyang, Nanxiong, Shaoguan, and Xuwen) did not pass the significance test of P < 0.05, while coefficients for the remaining stations varied between −0.13 and −0.45.

In the correlation analysis between SPI-1/SPEI-1 and TVDI_{EV I}, 19 stations were negatively correlated with SPI (except for Huiyang, Xuwen, and Nanxiong), and the Shaoguan station did not pass the significance test of P < 0.05. A total of fifteen stations passed the significance test of P < 0.001, with the correlation coefficient ranging from −0.23 to −0.60 and an average of −0.43. TVDI_{EV I} was negatively correlated with SPEI in 21 stations (excluding Shaoguan), and six stations (Gaoyao, Huiyang, Luoding, Nanxiong, Wuhua, and Xuwen) did not pass the P < 0.05 significance test. The correlation coefficient for the remaining stations ranged from −0.11 to −0.43.

Upon evaluating the correlation between the two monthly TVDI drought indices and the corresponding SPI-1/SPEI-1 at each station, it was found that TVDI_{NDV I} surpassed TVDI_{EV I} in both the number of stations achieving statistical significance at the P < 0.001 and P < 0.05 levels, and in the average Pearson correlation coefficient. This suggests that TVDI_{NDV I} offers a more sensitive response to drought in Guangdong province.

Spatial and temporal distribution characteristics of drought in Guangdong Province from 2000 to 2019

Figures 10 and 11 shows the average spatial distribution pattern of TVDI in Guangdong Province from 2000 to 2019. Predominantly, the province manifests a humid climate. However, a marked contrast in drought severity is evident between its northern and southern regions. The most pronounced drought conditions pervade the southernmost and mid-eastern coastal zones, exhibiting a multifaceted range from mild to severe degrees of drought. The northeastern coast sporadically experiences moderate to acute drought conditions, whereas the northwest to central-west areas remain largely devoid of drought. The Leizhou Peninsula, is the most severely affected area, with Suixi County and Xuwen County experiencing the most severe drought, and moderate to severe drought areas widely distributed. The next most affected area is the southern coastal area of Maoming City, where most of the areas are affected by moderate drought and some areas by severe drought. The drought in the central and eastern areas is mainly concentrated in Dongguan, Shenzhen, Huizhou, and Guangzhou, with the most extreme drought levels occurring in the urban areas of Dongguan, the entire county of Huiyang, the urban areas of Huizhou, and the eastern part of Bao’an County. In addition to the areas with large-scale drought distribution mentioned above, small-scale moderate to severe drought occurs in Huizhou County, Haifeng County, Lufeng County, Chaoyang County, Shantou City, Chenghai County in the northeastern coastal areas, and Fengshun County, Dapu County, and Chaozhou City in the northeastern central area. Apart from these focal areas of extensive drought, localized moderate to severe drought episodes also transpire within the northeastern coastal territories and northeastern central districts.

Guangdong Province frequently experiences winter-spring droughts, a phenomenon attributed to diminished rainfall, augmented evaporation, reduced soil moisture, and elevated temperatures during these seasons. To analyze the specific areas affected by these droughts, mean TVDI To analyze the specific areas affected by these droughts, the mean TVDI values from December of the previous year to April of the following year were calculated. The resulting maps depict the distribution of droughts during the winter and spring seasons for each year between 2001 and 2019 (Fig. 12).

As shown in Fig. 12, the drought-prone areas during winter-spring transition predominantly align with the annual TVDI distribution, especially within Suixi County of the Leizhou Peninsula, which consistently endures such drought conditions. The spatial distribution and extent of drought vary over time, with the largest affected area occurring in 2003 when the entire southern Leizhou Peninsula was in a severe drought state. The drought eased after 2004 but intensified again in 2005–2006, before subsiding in 2007. In 2009, the winter and spring drought was the lightest, but in 2010, it was the second most severe after 2003, with Xuchang County in the southernmost part of the Leizhou Peninsula shifting from moderate drought in the previous years to severe drought. The area of moderate drought in the northeastern coastal region also significantly expanded. The drought gradually eased until 2013, but intensified again in 2014, with severe impacts continuing until 2019, with winter and spring droughts affecting large areas of Guangdong Province almost every year. For a quantitative delineation of the extent and intensity of winter-spring droughts in Guangdong Province, an annual assessment of the impacted area’s proportion, stratified by drought severity, was executed (Fig. 13). The proportion of areas unaffected by winter and spring droughts in Guangdong Province ranged from 15.78% (2003) to 50.98% (2009), with a significant fluctuation range. The proportion of areas affected by mild drought ranged from 42% to 64.6%, with moderate drought ranging from 6.96% to 27.92%, and severe drought ranging from 0.002% to 1.84%. In 2003, the drought affected the largest area, accounting for 84.2% of the province’s total area, followed by 2010, with the drought affecting 74.98% of the province’s area. The lowest coverage occurred in 2009, with only 49.02% of the province’s area affected by drought. Although the proportion of severe drought areas in Guangdong Province is not high, the Leizhou Peninsula region experiences severe droughts every winter and spring, causing significant damage to local agriculture.