A Modified Temperature Vegetation Dryness Index (mTVDI) for Agricultural Drought Assessment Based on MODIS Data: A Case Study in Northeast China

Abstract

Satellite-based drought indices have been shown to be effective and convenient in detecting drought conditions. The temperature vegetation dryness index (TVDI) is one of the most frequently used drought indices; however, it is not suitable for areas with high fractional vegetation cover (FVC). In this study, a modified temperature vegetation dryness index (mTVDI) was constructed by using the multispectral vegetation dryness index (MVDI) proposed by a PROSAIL simulation and water stress experiments which was based on the theory of the TVDI and utilized MODIS data. Compared with the TVDI, the mTVDI presents a more triangular feature space, as well as obviously increased R² values for dry and wet edges (from 0.37–0.90 to 0.53–0.91 for dry edges and from 0.00–0.77 to 0.24–0.80 for wet edges). The mTVDI was evaluated using standardized precipitation evapotranspiration indices (SPEIs), precipitation, potential evapotranspiration (PET), and the crop water deficit index (CWDI), and the results confirmed that the mTVDI can better reflect the actual spatial changes, compared to the TVDI, under high FVC, as well as presenting an increased Pearson correlation coefficient (by 0.06–0.10) when compared with SPEIs. Moreover, the good performance of the mTVDI in major drought events indicates its reliability and accuracy for drought monitoring. Overall, the mTVDI is a reliable and accurate satellite-based dryness index suitable for high FVC conditions.

1. Introduction

Drought is one of the most serious natural disasters, causing serious water and food security problems all over the world [1,2,3,4]. The mechanisms of drought are complex and poorly understood. From different perspectives, droughts can be divided into meteorological, agricultural, hydrological, and socioeconomic droughts [1,2,5]. Given the severe impacts of drought on agriculture, it is urgent and important to develop effective methods for agricultural drought assessment.

In recent years, remote sensing has been shown to be capable of providing fast, economical, non-destructive, and spatio-temporal measurements of agricultural drought. A number of drought indices have been proposed to monitor agricultural drought based on visible (VIS), near-infrared (NIR), shortwave infrared (SWIR), and thermal infrared (TIR) remote sensing data. The drought indices can be divided into vegetation, temperature, and temperature–vegetation indices [6,7].

A vegetation index can detect drought conditions by combining spectral information from various regions that are sensitive to the physiology and biophysical characteristics of vegetation. The normalized difference vegetation index (NDVI) is the earliest and most widely used vegetation index, which uses non-linear stretching to enhance the reflectance contrast in the NIR and red spectral regions [8]. When the fractional vegetation cover (FVC) is higher than 80%, the increase in the NDVI is delayed, showing a saturation state, which leads to decreased sensitivity of vegetation detection under high FVC [9]. To overcome the saturation phenomenon of the NDVI, further vegetation indices have been established, such as the enhanced vegetation index (EVI) [9,10,11]. These indices are often referred to as greenness indices because they reflect the greenness of the vegetation [12,13].

In addition to the change in vegetation index, stomatal closure on leaves under water stress can lead to an increase in canopy temperature, which occurs earlier and provides a more sensitive indicator than a change in the vegetation index. Based on this physiological characteristic of crops, many indices based on land surface temperature (LST) have been proposed for drought monitoring [14]. These indices mainly include the temperature condition index (TCI) [15,16] and normalized difference temperature index (NDTI) [17]. Temperature indices have better performance than vegetation indices in accurately monitoring soil moisture due to evapotranspiration and leaf temperature [18]; however, a temperature index cannot distinguish the differences in drought resistance among different vegetation types, leading to certain inaccuracies in drought assessment [19].

In recent years, various approaches to combine vegetation and temperature indices for the detection of drought have been explored [16,18,20,21]. The most widely used temperature–vegetation indices include the vegetation supply water index (VSWI), vegetation health index (VHI), and temperature vegetation dryness index (TVDI). The TVDI is constructed according to the NDVI–LST triangle space, relying on the relationship between LST and the NDVI (which are usually negatively correlated) [22,23,24]. Compared with the individual vegetation and temperature indices, the TVDI is more reliable for drought monitoring and impact assessment [25,26,27]. However, the TVDI has limitations regarding drought monitoring in high FVC areas, due to the saturation of the NDVI [28,29]. Therefore, it is necessary to modify the TVDI in order to improve the accuracy and reliability of agricultural drought monitoring.

The aim of this study is to improve the accuracy of drought monitoring in high FVC areas. We present a modified drought index, termed the modified temperature vegetation dryness index (mTVDI), based on the multispectral vegetation dryness index (MVDI) and LST.

2. Materials and Methods

2.1. Study Area

The study area covers an area of 1,245,202 km² in Northeast China (38°40′—53°20′N, 111°37′—135°5′E). The terrain of the study area is mainly mountainous and plain, among which the Northeast plain is rich in black soil resources and fertile soil, making it an important grain base in China. The main climate type of the region is temperate monsoon climate, with an annual average temperature of 5.4 °C. The annual average precipitation is about 600 mm, which gradually decreases from east to west. The eastern part is humid, while the western part is semi-humid. The main local crop types in drylands are maize and soybean, and the growing period is from May to September. In this paper, the areas of corn and soybean were masked. In the extracted areas, the FVC was low in May due to seeding, and the FVC gradually increased with the growth of crops from June to September.

The most frequent natural disaster in Northeast China throughout the 20th century has been drought, which causes severe water shortages, environmental degradation, and negative socioeconomic effects. Extreme drought occurrences have become increasingly frequent, leading to significant agricultural losses [30,31].

2.2. Data

The Moderate Resolution Imaging Spectroradiometer (MODIS) data from May to September, 2007–2017, was chosen from the Google Earth Engine platform (https://code.earthengine.google.com, accessed on 1 June 2022). The MVDI and NDVI data were obtained from the MOD09A1 product corresponding to the Terra platform, with a spatial resolution of 500 m and a temporal resolution of 8 days. The LST data were generated, with a spatial resolution of 1 km and a temporal resolution of 8 days, using the MYD11A2 product corresponding to the Aqua platform. To match the spatial resolution of LST, the resolution of the MOD09A1 product was resampled to 1 km by cubic convolution interpolation. In general, the temporal resolution of the mTVDI was 8 days. When compared with other indices, such as the SPEI, the mTVDI was averaged monthly, in order to match the temporal scale.

The crop classification data were obtained from the 10 m crop type maps of the major crops (i.e., maize, soybean, and rice) in Northeast China in 2017 [32]. The dataset used the Google Earth Engine platform, Sentinel-2 remote sensing image, and random forest algorithm to classify major crops (i.e., maize, soybean, and rice) in Northeast China from 2017 to 2019. The resultant maps have overall accuracies spanning from 0.81 to 0.86.

The meteorological data were selected from the National Meteorological Centre of China Meteorological Administration (NMCCMA; https://data.cma.cn, accessed on 1 July 2022). In particular, daily precipitation, temperature, and sunshine duration were downloaded from 115 meteorological stations (Figure 1c) from May to September, 1978–2017.

The documented records of major drought events in Northeast China were obtained from the China Drought Dataset (CDD), which is released by China Engineering Science and Technology Knowledge Center (https://www.ckcest.cn, accessed on 15 July 2022).

2.3. Methodology

2.3.1. The Technical Route

The flowchart of this study is shown in Figure 2. We used the PROSAIL model to simulate the canopy spectral reflectance response of drought crops. Through water stress experiments of maize and soybean, it was shown that the spectral characteristic bands can detect the change of crops in response to drought. Using characteristic bands based on simulation and experiments, the MVDI was established. Then, we constructed the mTVDI by constructing the triangular feature space of the MVDI and LST based on MODIS data. We used the standardized precipitation evapotranspiration index (SPEI), the TVDI, precipitation, potential evapotranspiration (PET), the crop water eeficit index (CWDI), and the major drought events to evaluate the mTVDI. Finally, we used the mTVDI for monitoring drought in Northeast China.

2.3.2. Canopy Reflectance Simulation by PROSAIL

Model simulation can help to quantify the contributions of the canopy’s biophysical and biochemical properties to canopy reflectance. Since about thirty years ago, researchers have studied plant canopy spectral and directional reflectance in the solar domain using the PROSAIL radiative transfer model, which takes into account all biochemical parameters of leaves, vegetation canopy structure, and bidirectional scattering characteristics. PROSAIL-5B [33], which combines the 4SAIL and PROSPECT-5B [34,35] models, was employed in this study. The former defines how light is transmitted on the leaf’s surface and within, and is appropriate for both thick and sparse leaf vegetation. The latter model, known as the scattering by arbitrarily inclined leaves (SAIL) model, has been frequently used to investigate the vegetation canopy. The plant canopy in the model is thought to be a mixed medium, made up of isotropic leaves that are uniformly horizontal, endlessly long, and arbitrarily oriented. The transmission of light through the vegetation canopy is explained by radiative transfer theory, and canopy reflectance can be calculated.

For this study, 168 canopy spectral reflectance curves were simulated by successively changing the PROSAIL input parameters—equivalent water thickness (EWT), chlorophyll a+b content (Cab), and the leaf area index (LAI)—which are the most sensitive to water stress. The values and steps of the parameters are given in

Table 1. Parameter settings used in the PROSAIL model.

Model	Parameter	Quantity	Value	Step	Unit
PROSPECT	N	Leaf structure parameter	2.5	-	-
	Cab	Chlorophyll a+b content	10~100	1	μg/cm2
	Car	Carotenoid content	8	-	μg/cm2
	CW	Equivalent water thickness	0.005~0.03	0.001	cm
	Cm	Dry matter content	0.009	-	g/cm2
	Cbp	Brown pigments content	0	-	-
SAIL	LAI	Leaf area index	1~6	0.1	-
	LIDF	Leaf inclination distribution function	Beta distribution	-	-
	SL.	Hot spot parameter	0.01	-	-
	sza of θS	Solar zenith angle	30	-	deg
	vza of θV	Viewing zenith angle	10	-	deg
	raa of θSV	Relative azimuth angle	90	-	deg

.

A sensitivity analysis was conducted to identify the characteristic bands of the spectral reflectance, which are strongly responsive to parameters. The sensitivity calculation formula is as follows:

$$S={\displaystyle \sum _i^{n-1}(\left|R_i-R_{i+1}\right|/R_i})$$

(1)

where $R_i$ is the spectral reflectance of the corresponding parameter.

Figure 3 shows the canopy spectral reflectance and sensitivity analysis, as simulated by the PROSAIL model.

The EWT can be used for the characterization of crop water content and, further, to characterize crop drought [36]. It can be seen, in Figure 3a,b, that the range sensitive to crop water is in the NIR and SWIR, with the sensitive bands being 970, 1200, 1450, 1895, and 1980 nm. The function of the Cab is to convert light energy into chemical energy that plants can use. As can be seen in Figure 3c,d, the sensitive bands in this regard were 610 and 670 nm. It can be seen, in Figure 3e,f, that the spectral reflectance increased in the VIS and SWIR, decreased in the NIR, and changed changes in the LAI. The sensitive bands were 670, 780, 1450, and 1930 nm. This can be explained by interior scattering within the leaves at the air, cell, and water interfaces [37]. The PROSAIL simulation indicated that the spectral responses were generated by the different characteristic changes of crops subjected to water stress.

2.3.3. Water Stress Experiments

Experiments were conducted in the potted proving ground of Jilin University, Jilin, China (43°57′N, 125°15′E) during May–September of 2020–2021, using maize (Xianyu 335) and soybean (Jipin 609). To ensure stress environments, the maize and soybean were subjected to four water stress levels by adjusting the field capacity (FC) in the following ways: no water stress (N: 70–80% FC), light water stress (L: 50–60% FC), moderate water stress (M: 40–50% FC), and severe water stress (S: 30–40% FC). Nine replicates were set for each treatment, arranged in parallel (3 × 3). The samples were planted in flower pots, with one maize plant in each pot and four soybean plants in each pot. The quantity and quality of soil in the flowerpots were basically the same. The pots were 30 cm in diameter, 40 cm in height, and 35 cm in soil thickness. The soil was black in texture and the soil moisture retention was 23.8%. The soil moisture content of maize and soybean was controlled by an awning and irrigation.

The hyperspectral data were collected from the crop canopy by using an ASD Field Spec 3 portable spectrometer. This spectrometer had the following characteristics: a spectral region of 350–2500 nm, a spectral resolution of 3 nm in the VIS, 8.5 nm in the NIR, and 6.5 nm in the SWIR, a spectra sampling interval of 1.4 nm (350–1000 nm) and 2 nm (1000–2500 nm), and a 1 nm spectra resampling interval [38]. Reflectance measurements were made on sunny and windless days at 10:00–14:00 Beijing time. The field of view (FOV) was 25° and the distance between the optical head of the spectrometer and the top of the crop was kept at 1 m for all observations. During the measurements, standard whiteboard calibration was performed every 30 min. The spectral characteristics of the crops were measured during the critical period of crop growth (see

Table 2. Dates of spectral reflectance measurements.

Stage	Maize		Soybean
	Growth Period	Acquisition Time (2021)	Growth Period	Acquisition Time (2020)
Early	Jointing	24 June	Seedling	16 June, 28 June
Middle	Tasseling	14 July	Branch	9 July, 15 July
Late	Milking	9 August	Pod bearing	21 July, 25 July, 6 August
Mature	Maturing	30 August	Pod filling	12 August, 20 August, 26 August

).

The spectral reflectance of maize and soybean under water stress have similar characteristics in different growth stages. Figure 4 shows examples of the spectral reflectance of maize and soybean under different water stress levels. With an increase in water stress, the spectral reflectance of the maize and soybean canopy increased in the VIS and SWIR, while decreasing in the NIR. These spectral changes were related to the physiological and biochemical characteristics of crops under water stress. For example, with an increase in water stress, the decrease in chlorophyll content led to the weakening of spectral absorption which, in turn, led to the reflectance increase in VIS, while the decrease in the LAI led to the reflectance decrease in NIR, and the decrease in water content was reflected by the increased reflectance in the SWIR [6,39].

Figure 5 shows the sensitivity analysis of the canopy spectra of soybean and maize under different water stress levels. It can be seen that the sensitive bands of water stress for soybean were 480, 670, 780, 1200, 1450, and 2020 nm, while the sensitive bands of water stress for maize were 430, 670, 780, 1200, and 1450 nm. This is the result of a combination of crop water, structure, and chlorophyll. As can be seen in Figure 4 and Figure 5, the results of the water stress experiments for soybean and maize verified the results of the PROSAIL simulation, confirming that spectral information can be used to accurately detect the state of water stress in crops.

2.3.4. Multispectral Vegetation Dryness Index (MVDI)

Based on the results of the PROSAIL simulation and water stress experiments, we summarized the spectral response characteristic bands caused by different physiological and biochemical changes. The 430, 480, 610, and 670 nm wavelengths were chosen as indicators for chlorophyll content; the 970, 1200, 1450, 1895, 1980, and 2020 nm wavelengths were chosen as indicators for water content; and the 670, 780, 1450, and 1930 nm wavelengths were chosen as indicators for variability in crop structure.

Crops respond physiologically and biochemically to water stress in a variety of ways, such as by stunting growth, coiling or necrosis of photosynthetically active sections, wilting or, in extreme circumstances, a reduction in leaf area due to severe defoliation. In order to monitor crop drought, some bands may be helpful in spotting changes in crop water, structure, and chlorophyll. Determining an index based on a mixture of bands from spectral regions responsive to various crop attributes altered by water stress would, therefore, be valuable. To formulate a new spectral index, we combined the characteristic bands from different spectral responses. By combining different bands of chlorophyll content and crop structure, we found that the reflectance difference between 780 nm and 670 nm was most sensitive to water stress. On this basis, the difference between 780 nm and 670 nm and the water content bands were combined. Finally, the hyperspectral vegetation dryness index (HVDI), which is most relevant to water stress, was selected, as shown in Equation (2):

$$\mathrm{HVDI}=(R_{780}-R_{670})/R_{1200}$$

(2)

where $R_x$ is the reflectance at the corresponding wavelength.

Considering the application of the HVDI, we transform it into a multispectral form that can be applied to the MODIS. The MODIS has been widely used for drought monitoring, due to its advantages of large imaging width, wide spectral range, and high updating frequency [1,9,40,41]. Therefore, MODIS imagery was selected in this study, such that the HVDI in Equation (2), by changing the narrow band channel to the wide band channel, as shown in Equation (3):

$$\mathrm{MVDI}=(R_{{\mathrm{NIR}}_1}-R_{\mathrm{Re}\mathrm{d}})/R_{{\mathrm{NIR}}_2}$$

(3)

where MVDI is the multispectral HVDI based on MODIS, $R_{\mathrm{Red}}$ denotes the MODIS red band (620–670 nm), and $R_{{\mathrm{NIR}}_1}$ and $R_{{\mathrm{NIR}}_2}$ are the MODIS near-infrared bands (${\mathrm{NIR}}_1$: 841–876 nm, ${\mathrm{NIR}}_2$: 1230–1250 nm). Although the center of the MODIS channel is off-HVDI, the features in both channels are comparable, as we discuss in Section 4.1.

In general, the value of the MVDI is greater than 0; the smaller the MVDI, the stronger the water stress.

2.3.5. Modified Temperature Vegetation Dryness Index (mTVDI)

On the basis of the TVDI theory [23,42,43], the mTVDI was developed by combining the MVDI and LST to construct a triangle feature space (Figure 6). In the MVDI–LST feature space, point P represents dry bare soil and point Q represents wet bare soil. With an increase in FVC, the LST decreases. Point R represents where the FVC is highest and soil moisture is sufficient, such that evaporation and transpiration are maximal. Therefore, the dry edge indicates low soil water availability and small surface evapotranspiration, while the wet edge denotes adequate soil moisture and normal vegetation growth.

The MVDI–LST feature space has a number of contours related to soil moisture (Figure 6). Based on this correlation, the mTVDI is defined as follows:

$${\mathrm{mTVDI}=(\mathrm{LST}-\mathrm{LST}}_{\min })/({\mathrm{LST}}_{\max }{-\mathrm{LST}}_{\min })$$

(4)

$${\mathrm{LST}}_{\min }{=\mathrm{a}}_1{+\mathrm{b}}_1\times \mathrm{MVDI}$$

(5)

$${\mathrm{LST}}_{\max }{=\mathrm{a}}_2{+\mathrm{b}}_2\times \mathrm{MVDI}$$

(6)

where LST is the land surface temperature; ${\mathrm{LST}}_{\min }$ and ${\mathrm{LST}}_{\max }$ are the lowest and highest LST corresponding to a certain MVDI, respectively, which can be determined by extracting the wet edge using Equation (5) and dry edge using Equation (6) through linear regression analysis; and ${\mathrm{a}}_1$, ${\mathrm{b}}_1$, ${\mathrm{a}}_2$, and ${\mathrm{b}}_2$ are undetermined coefficients. The mTVDI value ranges from 0 to 1, with values closer to 1 indicating more severe drought.

In this study, LST was extracted by using 0.01 as the interval of the MVDI, and a 95% confidence interval was selected to leave out possible abnormal LST. Then, a series of maximum and minimum LSTs corresponding to the MVDI were calculated. After that, the dry and wet edges of the mTVDI were extracted using the linear regression method. The performance evaluation of the mTVDI is discussed in the Results and Discussion sections.

2.3.6. Other Drought Indices

In addition, nine vegetation indices that have been evaluated by other researchers using various vegetation data and have shown stable performance were calculated in order to comparatively assess the effectiveness of the MVDI. According to the number of bands and the selected algorithm, the spectral indices were classified into three types: structural, chlorophyll, and water indices (

Table 3. Vegetation indices used in this study.

Abbreviation	Index Name	Equation	References
Structural Indices
1. NDVI	Normalized difference vegetation index	$(R_{800}-R_{680})/(R_{800}+R_{680})$	Tucker [44]
2. RDVI	Renormalized difference vegetation index	$(R_{800}-R_{670})/\sqrt{(R_{800}+R_{670})}$	Roujean [45]
3. EVI	Enhanced vegetation index	$\begin{array}{l}2.5(R_{800}-R_{670})/(R_{800}+6R_{670}\\ -7R_{450}+1)\end{array}$	Huete [9]
Chlorophyll Indices
4. PRI	Photochemical reflectance index	$(R_{570}-R_{531})/(R_{570}+R_{531})$	Gamon [46]
5. SIPI	Structure-intensive pigment index	$(R_{800}-R_{445})/(R_{800}+R_{680})$	Penelas [47]
6. MCARI	Modified chlorophyll absorption ratio index	$\begin{array}{l}\left[\right(R_{700}-R_{670})-0.2(R_{700}-R_{550}\left)\right]\\ /(R_{700}/R_{670})\end{array}$	Daughtry [48]
Water Indices
7. NDII	Normalized difference infrared index	$(R_{819}-R_{1649})/(R_{819}+R_{1649})$	Jackson [12]
8. NDWI	Normalized difference water index	$(R_{860}-R_{1240})/(R_{860}+R_{1240})$	Gao [49]
9. WI	Water index	$R_{900}/R_{970}$	Penuelas [50]

). We compared the MVDI with the nine conventional vegetation indices in Section 3.1.

In addition, the TVDI, SPEI, and crop water deficit index (CWDI) were calculated to assess the effectiveness of the mTVDI. The TVDI is a method for measuring soil moisture based on the triangle feature space of NDVI–LST, which is common and reliable for use in drought monitoring [22].

The SPEI is a drought index calculated according to precipitation and temperature, which can be calculated at various time scales and has been widely used in drought research [51,52]. In this study, the time scales of one month (SPEI1), three months (SPEI3), and six months (SPEI6) were calculated for comparison with the mTVDI.

The CWDI is an indicator used to characterize agricultural drought through the relationship between precipitation and PET [53,54]. The CWDI was used to verify the drought monitoring results of the mTVDI.

2.3.7. Statistical Metrics

In order to quantitatively evaluate the ability of the MVDI and the nine other vegetation indices to identify the degree of water stress, the normalized mean distance was introduced. According to the Fisher criterion, the relative distance between categories can be used as a measure of category separability; the greater the distance, the greater the separability [55]. The normalized mean distance principle formula is as follows [56]:

$$d_{\mathrm{norm}}=\frac{\left|{\mu }_m-{\mu }_n\right|}{{\sigma }_m+{\sigma }_n}$$

(7)

where ${\mu }_m$ and ${\mu }_n$ are the respective mean values of the vegetation indices under different water stress levels and ${\sigma }_m$ and ${\sigma }_n$ are the respective variances of the vegetation indices under different water stress levels.

To quantitatively evaluate the performance of the mTVDI with that of the TVDI and SPEI, the Pearson correlation coefficient was adopted as a statistical metric. Its calculation formula is as follows:

$$P_{\mathrm{X},\mathrm{Y}}=\frac{Cov(\mathrm{X},\mathrm{Y})}{\sqrt{Var(\mathrm{X})}\sqrt{Var(\mathrm{Y})}}$$

(8)

where X and Y are the variables for comparison, $Cov$ is the covariance of X and Y, and $Var$ is the variance.

To verify the match between the HVDI bands and MODIS channels, a residual analysis was adopted as an evaluation metric for deviation. Its calculation formula is as follows:

$$\epsilon =y-\widehat{y}$$

(9)

where $\epsilon$ is the residual between the values of the HVDI and MVDI, and y and $\widehat{y}$ are the values of the HVDI and MVDI.

3. Results

3.1. Comparison of the MVDI with Other Vegetation Indices

The MVDI and nine other vegetation indices were selected to identify crop water stress levels using the hyperspectral data measured in the water stress experiments (Figure 7). As maize and soybean have similar spectral characteristics (Figure 4 and Figure 5), they are discussed together. As can be seen in Figure 7, the NDVI, EVI, PRI, SIPI, WI, NDII, and NDWI can identify water-stressed crops in the early, middle, and late stages, but could not accurately identify the water stress degree in the mature stage. The MCARI values of control and water-stressed crops presented no constant rule, except in the early stage, which cannot allow for accurate identification of water-stressed crops. The MVDI and RDVI could identify water-stressed crops throughout the whole growth period.

Moreover, to quantitatively evaluate the ability of the MVDI and the other nine vegetation indices to identify the degree of water stress, the normalized mean distance was calculated to judge the index recognition ability, as shown in

Table 4. The normalized mean distance between different water stress levels during the main growth stages of the crop (empty cell indicates that the index could not distinguish different water stress levels).

Type	Index	Early Stage				Middle Stage
		N–L	L–M	M–S	Average	N–L	L–M	M–S	Average
Structural Indices	NDVI	2.77	0.93	2.49	2.06	1.27	0.86	0.97	1.03
	RDVI	2.52	0.49	1.44	1.48	1.24	0.66	0.25	0.72
	EVI	2.56	0.58	1.67	1.60	1.26	0.75	0.43	0.81
Chlorophyll Indices	PRI	3.40	1.22	2.37	2.33	1.22	1.11	2.25	1.53
	SIPI	2.14	0.73	1.70	1.52	1.29	0.83	0.76	0.96
	MCARI	0.83	0.96	1.82	1.20	-	-	-	-
Water Indices	NDII	3.35	4.18	4.11	3.88	0.87	0.95	1.19	1.00
	NDWI	2.15	2.93	2.83	2.64	0.90	0.62	1.51	1.01
	WI	2.73	2.28	1.60	2.20	0.79	0.43	1.16	0.79
This Study Index	MVDI	4.32	4.47	5.03	4.61	1.59	0.89	1.45	1.31
Type	Index	Late stage				Mature stage
		N–L	L–M	M–S	Average	N–L	L–M	M–S	Average
Structural Indices	NDVI	1.55	2.64	2.44	2.21	-	-	-	-
	RDVI	2.57	1.22	1.55	1.78	1.22	1.09	1.43	1.25
	EVI	2.30	1.76	1.93	2.00	-	-	-	-
Chlorophyll Indices	PRI	1.00	0.44	1.96	1.13	-	-	-	-
	SIPI	1.58	3.27	2.38	2.41	-	-	-	-
	MCARI	-	-	-	-	-	-	-	-
Water Indices	NDII	0.41	0.05	0.27	0.24	-	-	-	-
	NDWI	0.84	0.78	0.52	0.71	-	-	-	-
	WI	0.82	0.86	0.38	0.69	-	-	-	-
This Study Index	MVDI	2.55	2.65	2.83	2.68	1.60	1.39	1.40	1.46

. In the early, late, and mature stages, the mean distance values for the MVDI were 4.61, 2.68, and 1.46, respectively, which were all larger than those of the other indices. In the middle stage, the mean distance of the MVDI was 1.31, which was second only to that of the PRI (1.53). However, as the PRI could not distinguish the water stress levels in the mature stage; it could be considered that the MVDI presented the best comprehensive performance.

In conclusion, the MVDI had a stronger ability to identify the physiological and biochemical changes of crops than the NDVI, RDVI, EVI, PRI, SIPI, MCARI, WI, NDII, and NDWI, and it had strong stability throughout the whole growth period of the crops.

3.2. Evaluating the Feature Space of the MVDI–LST

Figure 8 shows the MVDI–LST space of the mTVDI and the NDVI–LST space of the TVDI. The results show that the shape of the MVDI–LST space was more triangular than the NDVI–LST space, especially from June to September when the FVC is high. Moreover, the linear fitting of dry and wet edges in the MVDI–LST space was better than that in the NDVI–LST. The dry edge R² values for the NDVI–LST space were 0.90, 0.78, 0.72, 0.45, and 0.37, while those for the MVDI–LST space were 0.91, 0.88, 0.90, 0.82, and 0.53. The wet edge R² values for the NDVI–LST space were 0.77, 0.01, 0.00, 0.00, and 0.49, while those for the MVDI–LST space were 0.80, 0.26, 0.52, 0.24, and 0.35.

3.3. Applying the mTVDI in Drought Monitoring in Northeast China

Given the excellent results of the MVDI–LST feature space for the proposed mTVDI, MODIS data were used to calculate the mTVDI in order to assess the spatio-temporal distribution of droughts in Northeast China in 2017 (Figure 9). Figure 9 shows the evolution process of drought over the crop growth period. The western part of Northeast China experienced widespread drought in May, which may have been caused by insufficient precipitation in the early period. Then, the drought spread in June, due to an increase in water demand as crops began to grow rapidly. As temperatures increased, the drought covered almost the whole of Northeast China in July, which may have been due to the increasing evapotranspiration and water demand. From August to September, the drought finally eased in Northeast China, except for the persistent drought in the southwest part.

Moreover, we used precipitation, PET, and the CWDI to explain the spatio-temporal distribution of droughts in Northeast China monitored by the mTVDI (Figure 10). Due to the spatial difference in drought occurrence, we divided Northeast China into four regions (Heilongjiang, Jilin, Liaoning, and Neimenggu) for comparison and analysis. The mTVDI showed an aggravating trend of drought from May to July, which was caused by the continuous low precipitation and the increase in crop PET due to an increase in temperature. From August to September, the mTVDI indicated that the drought eased significantly, which was caused by the significant increase in precipitation in August and the decrease in crop PET due to the lower temperatures, consistent with the results of the CWDI. Among the four regions in Northeast China, the mTVDI showed that Neimenggu suffered the worst drought, which was caused by the low precipitation and high PET during the whole crop-growing period. The next worst droughts were in Liaoning and Jilin, while the drought was lightest in Heilongjiang.

In addition, the distribution of drought in the four regions may be related to crop types. Studies have shown that soybean is more drought-resistant than maize; this is because, in the growth process of soybean, its small leaf area leads to lower plant transpiration and, thus, less water is needed [57,58,59]. The north and east parts of Heilongjiang are soybean-growing regions, as shown in Figure 1b, and the drought tolerance of soybean plays a regulating role in drought. Jilin and Liaoning are mostly maize-growing regions and, thus, are less resistant to drought. In Neimenggu, although the soybean-growing region is wide, the drought is more severe, which may be mainly caused by the intolerance of soybean with respect to high temperatures [59,60].

3.4. Comparison of the mTVDI with Other Drought Indices

To evaluate the performance of the mTVDI and its consistency with other drought indices, the mTVDI was compared with the TVDI and various SPEIs, and spatio-temporal distribution maps of drought were created, as shown in Figure 9.

In general, these indices behaved similarly, in terms of the spatial trends of drought. However, there were some differences in the duration, distribution, and degree of drought. For example, both the mTVDI and the SPEIs indicated that the drought in Northeast China gradually intensified from May to July and gradually eased from August to September; however, the TVDI results showed that the drought was most severe in May, relieved from June to July, and intensified from August to September. This may be caused by NDVI saturation in areas with high FVC (usually an LAI greater than 3 [61]). In addition, the results of the mTVDI, SPEI3, and SPEI6 indicated that the most severe drought occurred in the southwest of Northeast China in July, while SPEI1 and the TVDI presented no significant drought in the south. This may be due to the bias caused by SPEI1, taking into account only one month’s precipitation and PET.

In order to compare the performance of the mTVDI and TVDI quantitatively, Pearson correlation coefficients were calculated between the mTVDI, TVDI, and SPEIs, as shown in

Table 5. Pearson correlation coefficients between the mTVDI, TVDI, and SPEIs.

Mon	Method	Heilongjiang			Jilin			Liaoning			Neimenggu			Mean
		SPEI1	SPEI3	SPEI6	SPEI1	SPEI3	SPEI6	SPEI1	SPEI3	SPEI6	SPEI1	SPEI3	SPEI6
May	mTVDI	−0.83	−0.86	−0.84	−0.39	−0.75	−0.76	−0.62	−0.78	−0.88	−0.61	−0.80	−0.82	−0.75
	TVDI	−0.80	−0.81	−0.80	−0.38	−0.65	−0.66	−0.62	−0.68	−0.77	−0.60	−0.72	−0.74	−0.69
Jun	mTVDI	−0.36	−0.39	−0.48	−0.52	−0.69	−0.68	−0.64	−0.72	−0.81	−0.69	−0.78	−0.78	−0.63
	TVDI	−0.37	−0.32	−0.40	−0.48	−0.57	−0.58	−0.66	−0.66	−0.46	−0.50	−0.52	−0.48	−0.55
Jul	mTVDI	−0.68	−0.65	−0.68	−0.86	−0.76	−0.81	−0.72	−0.69	−0.51	−0.75	−0.70	−0.66	−0.71
	TVDI	−0.66	−0.62	−0.65	−0.68	−0.68	−0.72	−0.64	−0.50	−0.38	−0.66	−0.60	−0.58	−0.61
Aug	mTVDI	−0.54	−0.79	−0.80	−0.45	−0.59	−0.52	−0.73	−0.67	−0.66	−0.57	−0.68	−0.66	−0.64
	TVDI	−0.47	−0.79	−0.77	0.29	−0.35	−0.61	−0.68	−0.66	−0.69	−0.29	−0.60	−0.69	−0.56
Sep	mTVDI	−0.77	−0.74	−0.81	−0.62	−0.69	−0.89	−0.65	−0.76	−0.83	−0.68	−0.73	−0.84	−0.75
	TVDI	−0.75	−0.63	−0.73	−0.65	−0.62	−0.76	−0.66	−0.71	−0.69	−0.69	−0.65	−0.72	−0.69

. It can be seen that the mTVDI had a stronger correlation with SPEIs than the TVDI. For example, from May to September, the average Pearson correlation coefficients of the mTVDI and SPEIs were −0.75, −0.63, −0.70, −0.64, and −0.75 (p < 0.01), which were all greater than the correlation coefficients with the TVDI. The results for the mTVDI, TVDI, and SPEIs showed that the mTVDI has a more accurate spatial distribution than the TVDI under high FVC.

3.5. Performance of the mTVDI in the Major Drought Events

To further verify the validity and applicability of the mTVDI, we looked for records of drought events that had occurred in Northeast China (

Table 6. The documented records of the major drought events from the China Drought Dataset (CDD).

Region	Drought Month
Heilongjiang	September 2008, September 2010
Jilin	July 2009, August 2009, September 2009, September 2010
Liaoning	September 2008, July 2009, August 2009, September 2009
Neimenggu	June 2007, July 2007, September 2009

). Due to the small number of drought records, we only obtained drought event data for the four considered regions (Heilongjiang, Jilin, Liaoning, and Neimenggu) during 2007–2010. The mTVDI values of the month when the drought events occurred were calculated, with mTVDI values closer to 1 indicating more drought. In addition, SPEI3 was calculated to assess the accuracy of the mTVDI due to its reliability in agricultural drought monitoring [3,42,62,63,64]. For ease of comparison, the SPEI results were normalized into the range from 0 to 1, with values closer to 0 indicating more drought. The results are shown in Figure 11, which demonstrates the drought records using red marks.

The results showed that the drought events identified by the mTVDI and SPEI3 were basically consistent with those recorded in the CDD. For example, in the four drought events in Liaoning, the mTVDI showed high values and SPEI3 showed low values, both of which represent drought. The performance of the mTVDI regarding the major drought events indicated its reliability and accuracy in drought monitoring.

4. Discussion

4.1. Matching between HVDI Bands and MODIS Channels

In this study, the MVDI was proposed for application to MODIS on the basis of the HVDI. It is based on the fact that the HVDI bands and the MVDI channels have the same meanings: $R_{670}$ of the HVDI and $R_{\mathrm{Red}}$ of the MVDI represent the strong reflectance absorption in the red spectral region, $R_{780}$ of the HVDI and $R_{{\mathrm{NIR}}_1}$ of the MVDI represent the high reflectance platform in the NIR, and $R_{1200}$ of the HVDI and $R_{{\mathrm{NIR}}_2}$ of the MVDI represent the water absorption peaks in the NIR [10,36]. However, the bands of the HVDI do not exactly correspond to the central wavelength of the channels of MODIS; thus, it is necessary to verify the match between the HVDI bands and MODIS channels. The HVDI and MVDI were calculated using the hyperspectral data measured in the water stress experiments, and the MVDI channels were obtained by convolution of the spectral response function of MODIS. The results of the HVDI and MVDI, and the corresponding residual analysis, are shown in Figure 12.

As can be seen in Figure 12a–d, there is a good consistency between the HVDI and MVDI at different stages of crop growth. The residual plots provided a reasonable pattern in which residuals are distributed randomly and unpredictably (Figure 12e–h) [65]. The results show that the HVDI bands match well with MODIS channels; thus, the MVDI can express the spectral response characteristic represented by the HVDI.

4.2. Anti-Saturation of the mTVDI in High FVC

In the MVDI–LST feature space of the mTVDI, the MVDI reflects not only the red edge but also the water absorption characteristics of crops [36,61]. According to previous research, there are four significant absorption peaks—at about 970, 1200, 1450, and 1950 nm—which can be used for the diagnosis of the water status in plants [66,67]. However, only the water absorption bands at 1200 and 1450 nm were detected in the water stress experiments of maize and soybean (Figure 5), which may be caused by the difference in canopy geometry and the influence of the soil background [68,69]. In addition, although absorption at 1200 nm is not as strong as that at 1450 nm, it is less susceptible to atmospheric influence in satellite applications [49]. Therefore, the addition of 1200 nm to the MVDI can better reflect the physiological and biochemical state of vegetation than the NDVI. The MVDI incorporates water information based on vegetation information, which makes the mTVDI perform good for drought monitoring in high FVC.

On the other hand, the NDVI uses non-linear stretching to enhance the reflectance contrast in the NIR and red spectral regions, which causes saturation in high FVC [9,70]. Under high FVC conditions, the increase in the NDVI slows down with the increase in FVC. The form of the MVDI formula also avoids the saturation effect for the mTVDI in high FVC conditions.

4.3. Limitation and Future Research

In this study, we verified the ability of the mTVDI in drought monitoring of major crops in Northeast China and achieved good results. However, the mTVDI still has some uncertainties, such as those relating to the adaptability of other vegetation types and other areas. Therefore, the mTVDI should be further validated by utilizing multiple vegetation types and wider areas for drought monitoring.

More recently, hyperspectral remote sensing technology has played an important role in quantitative remote sensing because of its high spectral resolution. In terms of monitoring and application of drought, it also has advantages that traditional multispectral remote sensing does not have, which can capture and discriminate subtle drought response signals and monitor vegetation drought [36,55,71]. In future research, a hyperspectral-modified TVDI should be constructed by the triangular feature space of the HVDI and LST, and further verified and applied to hyperspectral images.

5. Conclusions

In this study, in order to improve the applicability of drought monitoring based on the TVDI under high FVC, we constructed the mTVDI from the MVDI based on simulation and experimental data. By comparing the feature spaces of the mTVDI and TVDI, the results indicated that the mTVDI presented a more triangular feature space than the TVDI, and had obviously increased the R² values for the dry and wet edges (from 0.37–0.90 to 0.53–0.91 for dry edges and from 0.00–0.77 to 0.24–0.80 for wet edges). The Pearson correlation coefficient of the mTVDI was 0.06–0.10 higher than that of the TVDI; thus, it can better reflect the actual spatial changes, indicating its practical value in high FVC areas. As a result, the more trustworthy results generated using the mTVDI can aid in better field management, irrigation planning, and agricultural drought risk assessment.