Agricultural drought dynamics in China during 1982–2020: a depiction with satellite remotely sensed soil moisture

ABSTRACT

Agricultural drought (AD) is a serious threat to food security for many regions worldwide. Understanding the dynamics of AD contributes to preventing or mitigating its adverse impacts. Soil moisture (SM) anomaly is a relatively straightforward indicator of AD. However, most of the previous studies on AD dynamics of China were conducted with non-remotely sensed SM indicators due to the lack of long-term and spatial-continuous SM datasets. Here, such an SM dataset was created by enhancing a satellite remote sensing SM dataset with a machine learning method XGBoost, various remote sensing datasets, and some surface or meteorological parameters from reanalysis data. The new SM dataset has a period of 1982–2020, a spatial resolution of 0.25°, and a temporal resolution of 1 month. Furthermore, Standardized SM Index at one-month scale (SSMI1) was calculated, and AD events were identified using the SSMI1 and a 3-dimensional clustering method. Results demonstrated that 1) the new SM presented comparable or even better performances with the original SM as evaluated with spatial distributions, in-situ SM observations, and manufactured data gaps. 2) The AD was most frequent in North China, followed by the western parts of East China, Northeast, and Southwest China. The centroids of identified AD events were found chiefly in the Northeast, North, Southwest, and western parts of East China. 3) The severity of AD events presented a decreasing trend from 1982 to 2020, while significant drying trends were found mostly in the southern parts of North China, western parts of East China, and Southwest China. 4) The AD dynamics revealed in this study are basically consistent with other studies but also have unique features such as more space details and less drought frequency and count than that of meteorological drought. Further studies are expected to create a long-term satellite SM with faster timeliness, higher resolution, and greater depth.

KEYWORDS:

• Remote sensing of drought
• CCI soil moisture
• gap-filling method
• agricultural drought events

1. Introduction

Drought refers to a natural occurrence characterized by an extended period when a region experiences significantly reduced water availability compared to its usual conditions. This phenomenon can persist for varying durations, ranging from weeks and months to years or even decades (Xu et al. 2015). What’s worse, it can occur anywhere in the world with an imperceptible beginning and end. As a natural hazard, drought occurs with high frequency, prolonged duration, and widespread impact worldwide. It also significantly affects agriculture, water resources, environment, ecosystems, and human livelihoods. According to “Atlas of Mortality and Economic Losses from Weather, Climate and Water Extremes (1970–2019),” there were 4 drought events among the top 10 disasters that led to the most enormous human losses during the period, which caused almost 650 000 deaths, due to the severe African droughts of 1975, 1983, and 1984 (WMO 2021). Moreover, the frequency of concurrent heatwaves and droughts are projected to increase (high confidence) in the context of global climate change (IPCC 2023). Agricultural drought is a specific type of drought in addition to meteorological, hydrological, socioeconomic, and ecological droughts, etc. It chiefly corresponds to the phenomenon that a deficit in soil moisture affects the plant growth or production. It is a significant threat to food security in most countries including China. For example, according to the “Bulletin of Flood and Drought Disaster in China” from 2006 to 2019, the average disaster area of crops caused by drought was 7,526,000 hectares per year, the loss of economic crops was about 24.9 billion yuan per year, and the food loss was about 22 billion kilograms per year. In this context, it is necessary to understand agricultural drought’s spatial and temporal variations to prevent or mitigate its adverse impacts.

Historically, agricultural drought can be monitored by in-situ observed soil moisture (SM) through the oven drying method, which is destructive, time-consuming, and labor-intensive. These shortcomings can be circumvented using advanced in-situ SM measurement techniques, such as neutron scattering, time-domain-reflectometer, and resistive sensor (Babaeian et al. 2019). Additionally, the in-situ observed precipitation and evapotranspiration (ET) can also be employed as an indirect measure of SM. The drought indexes based on precipitation and ET include SPEI (Standardized Precipitation Evapotranspiration Index), RDI (Reconnaissance Drought Index), PDSI (Palmer Drought Severity Index), SPI (Standardized Precipitation Index), etc. For these precipitation and ET-based indexes, the 3-month to 6-month time scales are generally used for agricultural drought (Sun et al. 2017). Regardless, the above in-situ direct or indirect measurements of soil moisture are site-based methods, which have inherent disadvantages such as unsatisfactory spatial representation and discontinuous spatial distribution.

Currently, remote sensing technology provides spatially complete and temporally continuous information for monitoring agricultural drought. In a review study, the remote sensing-based indexes of agricultural drought were divided into bare surface-based indexes and vegetation-based indexes (Liu et al. 2016). The bare surface-based indexes were considered as including remote sensing thermal inertia methods and microwave SM inversion methods. The vegetation-based indexes were further classified into crop morphological indices, crop physiological indices, and crop comprehensive indices. The existing researches indicated that soil moisture is a significant diagnostic criterion of agricultural drought (AghaKouchak et al. 2015; Liu et al. 2016). Even, the agricultural drought was referred to as soil moisture drought in some studies (Liu et al. 2016; West, Quinn, and Horswell 2019). Thanks to the development of remote sensing of soil moisture, there are currently various remotely sensed SM datasets that can be used to directly monitor the agricultural drought. The representative methods of directly using SM to depict drought include Standardized Soil Moisture Index (SSMI) (Hao and AghaKouchak 2013), Soil Moisture Percentile (Sheffield et al. 2004), Soil Moisture Condition Index (SMCI) (Souza, Neto, and Souza 2021), Soil Water Deficit Index (SWDI) (Bai et al. 2018), and Normalized total depth Soil Moisture (NSM) (Dutra, Viterbo, and Miranda 2008), etc. The available remote sensing SM datasets include Soil Moisture Active Passive (SMAP), Soil Moisture and Ocean Salinity (SMOS), Advanced Microwave Scanning Radiometer 2 (AMSR2), and some fusion SM datasets such as a Daily all weather surface soil moisture data set with 1 km resolution in China (2003–2022) (Song et al. 2021) and ESA (European Space Agency) Climate Change Initiative (CCI) Soil Moisture dataset (Liu et al. 2011; Sun et al. 2019), etc. Among them, the ESA CCI SM product merges data from multiple active and passive microwave sensors, and it has the longest temporal span from 1978 to the present. Numerous studies have assessed the ESA CCI SM data, consistently demonstrating that the merged products exhibit comparable or superior performance when compared to the single-sensor input products (Dorigo et al. 2015; Dorigo et al. 2017). Some researchers have demonstrated the validity of the remote sensing SM data in monitoring agricultural drought (Bai et al. 2018; Eswar et al. 2018; Martínez-Fernández et al. 2016; Mishra et al. 2017; Wang et al. 2016).

China is located on the east of the Eurasian continental plate and on the west coast of the Pacific Ocean. It also has the Qinghai-Tibet Plateau, known as the “Roof of the World.” Such special geographical and climatic conditions determine frequent droughts occurring in China and bring very serious consequences to agriculture. Moreover, it is facing an increasing drought risk in the twenty-first century under the changing climate (Xu et al. 2015). In this context, understanding the spatiotemporal variation of China’s agricultural drought becomes very necessary for the prevention and mitigation of the adverse effects of drought disasters. As we know, such an investigation has been conducted using some indirect indicators of soil moisture such as Vegetation Health Index (VHI) (Zeng et al. 2020), Integrated Surface Drought Index (ISDI) (Zhou et al. 2017), SPI3/SPEI3/RDI3 (Xu et al. 2015; Yu et al. 2014), Normalized Difference Vegetation Index (NDVI) anomaly (Wang et al. 2015), and Process-based Accumulated Drought Index (PADI) (Zhang et al. 2017). However, there were very few studies using remotely sensed SM datasets to reveal the spatiotemporal variation of agricultural drought episodes in China due to a lack of long-term and spatial-continuous SM datasets. As reviewed above, the ESA CCI SM dataset has a long-term span. Nevertheless, the CCI SM dataset encounters significant challenges in addressing “data gap” issues caused by inherent limitations in microwave-based soil moisture retrieval methods. A recent study has highlighted the frequent occurrence of data gaps in the ESA CCI SM dataset, particularly in China, where the maximum ratio of data gap can exceed 80%. On average, the ratio of data gap in this region is approximately 40% (Sun and Qian 2021). These data gaps are bound to limit the study from revealing the spatiotemporal variation of agricultural drought.

Motivated by the above research background, we first developed a novel approach in this study to address the data gaps that existed in the ESA CCI SM dataset specifically for China. Subsequently, we calculated a standardized soil moisture index (SSMI) based on the gap-filled SM like the calculation of SPI and SPEI. Then, a clustering method suggested by Lloyd-Hughes (2012), a temporal connecting procedure, and the Mann-Kendall trend analysis was implemented to identify and characterize the agricultural drought episodes based on the SSMI drought index. More details about the method, materials, and results can be found in the following sections.

2. Data sources

2.1 ESA CCI SM

In this study, we utilized the ESA CCI SM dataset version 7.1, which can be accessed from the website (https://www.esa-soilmoisture-cci.org/), to compute the drought indices. The SM dataset employed in this study comprises three types of products: active microwave, passive microwave, and a combination of both. These products provide daily temporal resolution and a spatial resolution of 0.25 degrees. The combined and passive product spanning from November 1978 to December 2020, and the active product cover only from August 1991 to December 2020. Only the combined SM was used here for the gap-filling procedure, considering its longest temporal span and the best performance (Gruber et al. 2019). The original daily CCI SM data was first merged into a monthly product using averaging method. Then, the gap-filling technology shown in the following Methods section was conducted on the monthly CCI SM to obtain a gap-filled monthly ESA CCI SM dataset which is the basis to calculate drought indices and identify the drought episodes.

2.2 GIMMS & MODIS NDVI

Two NDVI datasets were utilized in this study to cover the period spanning from 1982 to 2020. The first NDVI dataset utilized in this study is the Global Inventory Monitoring and Modelling System (GIMMS) NDVI dataset (NASA ECOCAST GIMMS NDVI3g.v1), available at https://ecocast.arc.nasa.gov/data/pub/gimms/. It is derived from NOAA AVHRR and spans the period from 1982 to 2000. This NDVI dataset has a spatial resolution of 1/12 degrees (about 0.083 degrees) and a temporal resolution of 1/24 a year (about 15 days). In addition to the GIMMS NDVI dataset, we also included MODIS NDVI data covering the period from 2001 to 2020. The MODIS NDVI data were obtained from the MOD13C2 Version 6, accessible at https://www.earthdata.nasa.gov/. Both GIMMS NDVI and MODIS NDVI products were aggregated into a spatial resolution equaling that of ESA CCI SM through averaging method. Note that the GIMMS NDVI and MODIS NDVI have been compared in many studies and the overall consistency between them in indicating vegetation changes has been demonstrated (Gallo et al. 2005; Du et al. 2014; Fensholt and Proud 2012). Therefore, the GIMMS and MODIS NDVI were combined to provide continuity of NDVI data.

2.3 CMA-RA/Land reanalysis data

CMA-RA is a reanalysis dataset provided by the China Meteorological Administration (CMA). This dataset contains atmospheric reanalysis data and matched land surface reanalysis data (CMA-RA/Land) with a temporal resolution of 6 h and a spatial resolution of 34 km. In this study, monthly CMA-RA/Land products spanned from January 1979 to December 2020 were collected. Specifically, the land surface variables, including soil moisture, surface albedo, surface temperature, air temperature, precipitation rate, and potential evapotranspiration were used from the CMA-RA/Land dataset.

2.4 Other materials

Besides the data sources mentioned above, GTOPO30 digital elevation model (DEM), soil texture from Harmonized World Soil Database (HWSD), and in-situ measurements from the International Soil Moisture Network (ISMN) (Dorigo et al. 2011) were also used in the training of machine learning and evaluation of gap-filled ESA CCI soil moisture. GTOPO30 DEM was downloaded using Google Earth Engine and HWSD was downloaded from the website of the Food and Agriculture Organization of the United Nations. Note that the soil texture was converted from the Global Mapping Unit identifier (field named after “MU_GLOBAL”) which provides the link between the GIS layer and the attribute database. Then, the soil texture was divided into three bands, i.e. the proportion of sand, silt, and clay. In-situ measurements of soil moisture from six networks, i.e. CHINA (40 stations), CTP-SMTMN (57 stations), MAQU (27 stations), NAQU (11 stations), NGARI (23 stations) and SMN-SDR (34 stations) (Su et al. 2011; Yang et al. 2013; Zhao et al. 2020; Zheng et al. 2022) were averaged by month and compared to the ESA CCI original/gap-filled soil moisture. All the data sources are summed up in Table 1.

Table 1. Data sources used in this study.

Variables	Category	Unit	Spatial Resolution	Temporal Resolution
Soil Moisture	ESA CCI v07.1	cm^3cm^−3	0.25°	daily
NDVI	MOD13C2 & GIMMS	-	0.05° & 0.083°	monthly & 15 days
Background Soil Moisture	CMA-RA/Land	cm^3cm^−3	34 km	monthly
Albedo	CMA-RA/Land	-	34 km	monthly
Surface Temperature	CMA-RA/Land	K	34 km	monthly
Air Temperature	CMA-RA/Land	K	34 km	monthly
Precipitation	CMA-RA/Land	mm	34 km
Potential Evapotranspiration	CMA-RA/Land	mm	34 km	monthly
Soil Texture	HWSD	(Sand, Silt, Clay)%	0.0083°	-
DEM	GTOPO30	m	0.0083°	-
In-situ observations	ISMN	cm^3cm^−3	-	-

2.5 Study area

The Chinese Mainland was selected as the study area as shown in Figure 1. Generally, the Chinese Mainland is divided into three parts: Northern China, Southern China, and the Tibetan Plateau where Northern China consists of Northeast, North, and Northwest China, and Southern China consists of Southwest, East, and South China. To facilitate analysis, the study area was partitioned into three primary regions according to the classification of the aridity index. These regions include the arid region, the transition region encompassing sub-humid and semi-arid areas, and the humid region. The aridity indices for these regions were defined as follows: 0–0.2 for the arid region, 0.2–0.75 for the transition region, and >0.75 for the humid region (Xu et al. 2015). As shown in Figure 1, Northwest China and most parts of the Tibetan Plateau are in the arid region. North and Northeast China are in the transitional region. South China belongs to the humid region. What is more special is the East and Southwest China regions because they consist of transitional and humid climate regions. In this study, the arid region (including Northwest China and most parts of the Tibetan Plateau) was eliminated from the identification procedure due to the following reasons: (1) the bread baskets of China are mainly located in transition and humid regions; (2) water resources replenishment in the arid region primarily relies on the thawing of glaciers or perennial frozen soil, rather than precipitation; (3) to ensure a comprehensive and reliable analysis, it is crucial to enhance the extent of meteorological observations in this region and gather more extensive data, etc.

Figure 1. Climatic sub-regions in Chinese Mainland, locations of the in-situ SM sites, and six validation regions of interest (ROIs) under Asia North Albers equal area conic projection.

3. Methods

3.1 Generation of gap-filled SM dataset

The gap-filling method designed in this study is based on the machine learning (ML) method as shown in Figure 2, which can be expressed as follows:

(1) $\begin{array}{rl}& {\theta }_a=f_a\left({\alpha }_1^a,{\alpha }_2^a,{\alpha }_3^a\cdots {\alpha }_n^a\right)\\ & \downarrow \\ & {\theta }_g=f_a\left({\alpha }_1^g,{\alpha }_2^g,{\alpha }_3^g\cdots {\alpha }_n^g\right)\end{array}$(1)

Figure 2. The flowchart to generate gap-filled SM and identify droughts utilizing the spatial-temporal analysis frame.

where ${\theta }_a$ and ${\theta }_g$ are available SM and the gap of SM, respectively. $f_a$ is the machine learning algorithm established with available SM data. ${\alpha }_1^a$, ${\alpha }_2^a$, ${\alpha }_3^a$ … ${\alpha }_n^a$ are SM covariates at the pixel of available SM. ${\alpha }_1^g$, ${\alpha }_2^g$, ${\alpha }_3^g$ … ${\alpha }_n^g$ are SM covariates at the gap pixel of SM. The main idea of this method is employing the ${\theta }_a$ and various SM covariates to establish an estimation model by a machine learning algorithm. Subsequently, the specified model was applied for the gap pixels with available SM covariates to estimate SM. The estimated SM was finally used to fill the gaps. There are two key parts involved in this gap-filling method: the machine learning algorithm and the SM covariates.

In this study, the XGBoost model was selected since it is an improved algorithm based on the gradient-enhanced decision tree utilizing a practical construction of enhanced trees while running parallel computing. As to the SM covariates, geographical information and meteorological or climate factors were adopted according to previous researches (Sun and Cui 2021; Sun and Qian 2021), which include Normalized Differential Vegetation Index (NDVI), Albedo (A), Land Surface Temperature (LST), Air Temperature (AT), Precipitation (P), Antecedent Precipitation (AP), Potential Evapotranspiration (PET), Soil Texture (ST), Elevation (DEM), background SM from CMA-RA/Land, and time information, i.e. “year,” since machine learning model was established by month.

To evaluate the performance of the gap-filled SM dataset, the comparisons with the in-situ observed SM and with the original CCI SM over man-made gaps were conducted. Four metrics were utilized in the comparison, including unbiased root-mean-square error (ubRMSE), root-mean-square error (RMSE), correlation coefficient (R), and mean BIAS (Sun and Cui 2021). The six manufactured gaps, i.e. regions of interest (ROI) were distributed in transition and humid regions as shown in Figure 1.

3.2 Calculation of drought indices

Three drought indices were calculated in this study, including SPI, SPEI, and SSMI, where the SSMI was used as the primary agricultural drought indicator. This was implemented with an open-source Python package “climate_indices” developed by James Adams (GitHub – monocongo/climate_indices: Climate indices for drought monitoring, community reference implementations in Python). This library provides reference implementations of commonly used climate indices such as SPI and SPEI. Here, precipitation and potential evapotranspiration from the CMA-RA/Land dataset were used to calculate SPI and SPEI, while gap-filled SM was used to calculate the SSMI. The input datasets were first converted to a 3-dimensional array. The drought indices were calculated pixel by pixel based on the sequences along the temporal dimension. Here are some detailed parameters: (1) the timescale was set to 3 months for SPI and SPEI, while 1 month for SSMI; (2) the distribution type to be used for the internal fitting and computation was set as “gamma”; (3) the periodicity of the time series was set as “monthly”; (4) the initial year of the input datasets was set to 1982. SPI, SPEI, and SSMI are computed using a similar methodology, and their values share the same statistical significance. Consequently, these indicators are comparable to each other. Therefore, we classify the drought indices into four levels in Table 2 to uniformly express the drought conditions (Yu et al. 2014).

Table 2. Drought classifications according to the drought indices.

Drought Class	Index Value	Drought Class	Index Value
No drought	≥−1.0	Severe drought	−2.0～−1.5
Moderate drought	−1.5～−1.0	Extreme drought	≤−2.0

3.3 Identification of drought events

Upon computing the drought indices, −1 is applied as the threshold to determine whether the pixel is under drought conditions. The aforementioned drought indices were utilized to identify drought events over the past few decades utilizing a space-time continuous technique initially proposed by Andreadis et al. (2005) and further enhanced by Lloyd-Hughes (2012). This approach can be outlined in the subsequent steps, as illustrated in Figure 3.

Figure 3. Schematic diagram of spatial-temporal analysis frame to identify drought events.

The first step is clustering the drought patches in single monthly drought indices using the clustering algorithm suggested by Andreadis et al. (2005). A drought patch is defined as a collection of grid cells with a drought index value lower than −1, and these grid cells are interconnected. This process is repeated for all the monthly drought indices throughout the study period. Then the drought patches are all identified and marked with different labels. Finally, the patches are stored as an object which belongs to a class named “drought patch.”

Secondly, the algorithm should establish the connectivity between drought patches in two consecutive months. When considering two patches, denoted as a and b, extracted from two adjacent months in chronological order, if their overlapping area exceeds a predefined threshold, they will be linked together as part of a single drought event.

The third part is the identification of drought events over the past 40 years. For each patch, search all the other patches from the following month until the temporal connection is interrupted. In other words, the searching procedure ends when there is a gap of patches between the adjacent months. These connected patches are packed as a new object belonging to a class named “drought event.” Subsequently, the connected patches are excluded from the searching list consisting of drought patches. Repeat the steps until the searching list is empty. Finally, all drought events can be regarded as a set of temporal-continuous drought patches and marked with period stamps. The identification procedure can be expressed by the following figure.

For the convenience of characterizing the drought events, some parameters are calculated. The definitions are:

(1) Duration (D) is the persistent period of a drought event. The duration of a drought event is computed as the difference between the start and end times of the event. (2) Severity (S) is an index to express the total amount of water shortage during a drought event both spatially and temporally. It can also be more intuitively interpreted as a summation of drought index values during drought events. The severity of drought event is defined as:

(2) $S_n=\sum _{i=1}^{n_{lon}}\sum _{j=1}^{n_{lat}}\sum _{k=1}^{n_t}s\left(i,j,k\right)$(2)

(3) $s\left(i,j,k\right)=\mathrm{D}\mathrm{I}\left(i,j,k\right)\times \mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\left(i,j,k\right)\times \mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\left(i,j,k\right)$(3)

where $S_n$ is the severity of the $n$-th drought event, s is the severity of a single drought patch. $\mathrm{D}\mathrm{I}\left(i,j,k\right)$ is the drought indices value, $\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\left(i,j,k\right)$ is the area of drought patch which is treated as one grid (0.25° × 0.25°), and $\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\left(i,j,k\right)$ is treated as 1 month.

• (3) Affected Area (A) is the area affected by a drought event. It is here treated as the total number of grids of drought events in both space and time fields.

• (4) Intensity (I) is defined to express the average level of drought events. Using intensity, the development of drought events can be more easily seen. It aids in distinguishing drought events characterized by extended durations across large areas from those occurring over shorter periods and limited areas. It can be expressed as the following equation.

  (4) $I=S/\left(D\times A\right)$(4)

• (5) Centroid (C) is the center of the drought event over the three-dimensional space-time domain (longitude, latitude, and time). The space center is calculated by the positions of each grid during the drought event.

3.4 Depiction of drought dynamics

The drought dynamics were depicted in two parts: spatial distribution of drought frequency and drought centroid, temporal variation of drought severity and evolutionary trend of drought index. The drought frequency was obtained at each pixel by calculating the ratio of pixel numbers that were identified as drought to the amounts of all pixels based on the drought index during 1982–2020, where the drought identification criterion is listed in Table 2. The drought frequency varied between 0% and 100%. The drought event centroid can be obtained by the method in the above section.

The temporal changes in drought were assessed using the Mann-Kendall trend test for the following reasons: (1) it is a nonparametric test, making it suitable for analyzing data with various distributions; (2) being a rank-based method, it is robust to the influence of outliers in the data; and (3) it has been extensively employed in studying the spatiotemporal patterns of land surface variables. The normalized test statistic Z is the main output of Mann-Kendall analysis, and it can be expressed as follows:

(5) $Z=\left\{\begin{array}{c}\left(S-1\right)/\sqrt{\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)},\mathrm{i}\mathrm{f}S>0\\ 0,\mathrm{i}\mathrm{f}S=0\\ \left(S+1\right)/\sqrt{\mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)},\mathrm{i}\mathrm{f}S<0\end{array}\right.$(5)

where the statistics S and its variance Var(S) are calculated as:

(6) $\left\{\begin{array}{c}S=\sum _{k=1}^{N-1}\sum _{j=k+1}^N\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\left(R_j-R_k\right)\\ \mathrm{V}\mathrm{a}\mathrm{r}\left(S\right)=\left[N\left(N-1\right)\left(2N+5\right)-\sum _{p=1}^g{t}_p\left(t_p-1\right)\left(2t_p+5\right)\right]/18\end{array}\right.$(6)

where $N$ the total number of data points in the time series; $R_j$ and $R_k$ are the data at time $j$ and $k$, respectively; $\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\left(R_j-R_k\right)$ is equal to 1 when $R_j>R_k$, to 0 when $R_j=R_k$, and to −1 when $R_j<R_k$; $g$ representing the number of tied groups in the dataset, where a tied group refers to a set of data points with the same value; $t_p$is the number of data points in the $p$-th group. The Z score values can be divided into five levels where the Z score below −2.57 is recognized as a very significant decreasing trend, the Z score between [−2.57, −1.96] is recognized as a significant decreasing trend, the Z score between [−1.96, 1.96] is recognized as no significant change trend, the Z score between [1.96, 2.57] is recognized as a significant increasing trend, and the Z score above 2.57 is recognized as a very significant increasing trend. A specific description of the Mann–Kendall method can refer to Sun et al. (2021).

4. Results

4.1 Evaluation of the gap-filled CCI SM dataset

The original CCI SM dataset from 1982 to 2020 was firstly gap-filled in this study, because it was found that the ratio of gap pixel numbers varied around 10% to 80%, and its average is around 40% in the entire Chinese Mainland (Sun and Qian 2021). Figure 4 compares the original and gap-filled CCI SM in several randomly selected months. The discontinuous data gaps can be easily found in the original CCI SM data, especially in the subfigures (a2) and (a4). After the gap-filling process, these data gaps were filled with estimated values by the proposed gap-filling method. As shown in Figure 4, the gap-filled SM data present satisfactory performance in describing the spatial variations of soil moisture as compared with the original CCI SM data.

Figure 4. Comparisons of original and gap-filled CCI SM where the subfigures(a1), (a2), … , and (a6) are original CCI SM in May 1982, June 1990, July 1998, August 2006, September 2014, and October 2020, respectively; the subfigures (b1), (b2), … , and (b6) are gap-filled CCI SM corresponding to the gap-filled SM.

Further, the original SM and gap-filled SM were evaluated with in-situ SM measurements from six networks of ISMN. Figure 5 shows the comparisons with box plots. CTP_SMTMN showed the best consistency among the six networks according to the statistical indicators. The median R of original SM and gap-filled SM at CTP_SMTMN were 0.778 and 0.784, the median RMSE was 0.064 and 0.094 cm³cm⁻³, the median ubRMSE were 0.04 and 0.061 cm³cm⁻³, and the median mean biases were 0.003 and −0.031 cm³cm⁻³, respectively. The Network CHINA showed the lowest consistency with median R of 0.249 and 0.229, median RMSE of 0.077 and 0.076 cm³cm⁻³, median ubRMSE of 0.054 and 0.054 cm³cm⁻³ and median bias of −0.032 and −0.032 cm³cm⁻³. Despite the differences among various ISMN networks, the gap-filled SM presented comparable or even better performances as compared with the original CCI SM. For example, the gap-filled SM outperformed the original SM in the metric of R for some networks, such as the MAQU, NAQU, and SMN-SDR.

Figure 5. The comparisons between gap-filled SM and in-situ observed SM at six ISMN networks (CHINA, CTP-SMTMN, MAQU, NAQU, NGARI, SMN-SDR) with the metrics of (a) R, (b) RMSE, (c) ubRMSE, and (d) bias.

Besides, scatter plots are used to examine the overall performance as shown in Figure 6. The original CCI SM with 8055 paired samples has R of 0.45, RMSE of 0.0907 cm³cm⁻³, ubRMSE of 0.0884 cm³cm⁻³ and bias of −0.0203 cm³cm⁻³. The gap-filled SM has R of 0.554, RMSE of 0.0977 cm³cm⁻³, ubRMSE of 0.0894 cm³cm⁻³ and bias of −0.0395 cm³cm⁻³. After gap-filling, the R has been improved, and the other indicators have similar performance.

Figure 6. Scatterplots of the comparisons (a) between original SM and in-situ SM and (b) between gap-filled SM and in-situ SM.

As for the six ROIs, data distributed in these areas were first deleted in order to create manufactured gaps. The proposed gap-filling method was then operated over these manufactured gaps to generate predicted values. These predicted values were then compared with their original values with scatter plots as shown in Figure 7. Comparison results show that the R values of the six ROIs are 0.738, 0.695, 0.73, 0.733, 0.854, and 0.822; RMSEs are 0.0292, 0.0324, 0.0334, 0.03, and 0.181 cm³cm⁻³; ubRMSEs are 0.0291, 0.0316, 0.0321, 0.0182, 0.0297, and 0.018 cm³cm⁻³; and the biases are 0.0026, 0.0071, −0.009, 0.058, −0.0044 and 0.0017 cm³cm⁻³. Such smaller RMSE, ubRMSE, and bias and larger R indicated again that the proposed gap-filling method had a fine ability to fill the data gaps in the ESA CCI SM dataset.

Figure 7. Scatterplots between the original and predicted CCI SM over the six man-made data gaps corresponding to the ROIs in fig. 1.

In summary, the spatial distributions, the comparisons with in-situ SM, and the comparisons over manufactured gaps all demonstrate the reasonability of the gap-filled SM dataset.

4.2 Evaluation of the identified agricultural drought events

Based on the gap-filled CCI SM dataset, we calculated the SSMI drought index at 1-month scale which was labeled as SSMI1-CCI. Moreover, we also calculated the SPI3 and SPEI3 with reanalysis precipitation and air temperature dataset, i.e. CMA-RA/Land reanalysis data (labeled as SPI3-CRA and SPEI3-CRA). The SSMI indicates the severity of agricultural drought, while the SPI3 and SPEI3 indicate the severity of meteorological drought. With the help of the long-term drought indexes and the drought event identification method, we can depict the specific drought events. Here, a specific drought event, i.e. the 2011 drought event in South China was characterized by the above drought indexes. Figures 8, 9 and Figure 10 illustrate the spatial-temporal developing process of the 2011 drought event in South China identified by SSMI1-CCI, SPI3-CRA, and SPEI3-CRA, respectively. The pixels under drought condition were clustered as patches based on the drought indexes. The black dots are the centroids of each patch. The variation of affected area and intensity of this drought event are shown in Figure 11.

Figure 8. Spatiotemporal description of the 2011 drought event in South China based on the SSMI1-CCI index.

Figure 9. Spatiotemporal description of the 2011 drought event in South China based on the SPI3-CRA index.

Figure 10. Spatiotemporal description of the 2011 drought event in South China based on the SPEI3-CRA index.

Figure 11. Intensity and affected area of the 2011 drought event in South China depicted by the SSMI, SPI, and SPEI indexes.

From the position of the drought event centroids in Figure 8, it can be found that this agricultural drought event was near the Yangtze River basin during the first months. Subsequently, this agricultural drought event moved to the southwest of China during the last months and was finally located in the junction of Yunnan, Guizhou, and Sichuan. From the areas of the drought event in Figure 8, it can be found that this agricultural drought event was more severe in March, April, and May than in other months. Such movement trajectory and severity differences in different months were consistently found by SPI3 and SPEI3. Indeed, there are some differences between the drought events revealed by different drought indexes. First, the centroid of drought events did not coincide exactly as shown in Figures 8, 9 and Figure 10. Secondly, the area and intensity of agricultural drought events were generally lower than those of meteorological drought events as shown in Figure 11. Thirdly, the agricultural drought events depicted by the SSMI1-CCI were relatively discrete as compared with the meteorological drought events depicted by the SPI3-CRA and SPEI3-CRA. The above consistency demonstrates the reasonability of characterizing the agricultural events with the gap-filled CCI SM data. Correspondingly, the above differences demonstrate the necessity of characterizing agricultural drought with the gap-filled CCI SM data.

To further evaluate the identified drought events, the remotely sensed NDVI was utilized. Since NDVI is directly related to vegetation growth which is driven by soil moisture to a great extent, the agriculture drought event with lower soil moisture should have lower NDVI. Therefore, the NDVI was first fitted to the gamma distribution according to the time series of NDVI at each pixel. The aggregation scale of NDVI was set to 1 month. Finally, the sum of gamma-distributed NDVI over each drought event patch (called NDVI drought severity) was compared to the severity of each drought event identified by the SSMI1-CCI, SPI3-CRA, and SPEI3-CRA. The comparison scatter plots are shown in Figure 12. The drought severity identified by the SSMI1-CCI, SPI3-CRA, and SPEI3-CRA presented a good consistency with the NDVI drought severity. The R² values are 0.85, 0.88, and 0.79 for the SSMI1-CCI, SPI3-CRA, and SPEI3-CRA, respectively. Certainly, there are some exceptional drought events where NDVI is above normal conditions during the drought period. The exceptional events that appeared may be because of artificial irrigation, groundwater replenishment, energy-limited vegetation growth period, etc. Those exceptional events were relatively rare compared to reasonable events. Overall, the comparisons with NDVI further demonstrate the reasonability of characterizing the agricultural events with the gap-filled CCI SM data.

Figure 12. Correlation between the drought severity identified by SSMI1-CCI, SPI3-CRA, and SPEI3-CRA against the NDVI drought severity.

To sum up, the gap-filled CCI SM data was further employed to calculate a soil moisture drought index SSMI and identify agricultural drought events. Evaluations with SPEI3, SPI3, and NDVI anomaly demonstrated the validity of the SSMI for indicating agricultural drought.

4.3 Spatial variations of agricultural drought

Based on the SSMI1-CCI during 1982–2020, the agricultural drought frequency was obtained at each pixel. Figure 13 illustrates the spatial distribution of drought frequencies for moderate drought, severe drought, extreme drought, and all droughts, respectively. To facilitate statistical analysis, box plots of the drought frequency over different regions of China are presented in Figure 14.

Figure 13. Spatially distributed frequencies of moderate, severe, extreme, and all droughts by the drought indexes of SSMI1-CCI, SPEI3-CRA, and SPI3-CRA.

Figure 14. Statistics of the drought frequency over different regions of China, where (a) is the drought indicated by SSMI1-CCI, (b) is that by SPI3-CRA, and (c) is that by the SPEI3-CRA.

Firstly, the frequency of moderate drought is greater than that of severe drought and further greater than that of extreme drought. Extreme drought has the lowest frequency for both agricultural drought and meteorological drought. For moderate drought, the frequency of agricultural drought is higher in North and Northeast China, while it is relatively lower in East and South China. Comparatively, the frequency of meteorological drought is not only higher in North and Northeast but also higher in the whole of Southern China. For severe drought, the frequency of agricultural drought is roughly equivalent in North, Northeast, and the entire Southern China, while it is relatively lower in the Tibetan Plateau, which is consistent with that of meteorological drought. For extreme drought, the frequency of agricultural drought is higher in the Northeast and the whole of Southern China. The Northeast and Southwest regions are also frequent areas for extreme meteorological droughts. Overall, the agricultural drought was most frequent in North China, while the meteorological drought identified by SPEI3 was most frequent in Southwest China. The second most frequent regions of agricultural drought were the western parts of East China, Northeast, and Southwest China. The least frequent region of agricultural drought was the South China.

Furthermore, the identified drought events were analyzed. Figure 15 (a) presents the spatially distributed centroids of agricultural drought events during 1982–2020. Additionally, the centroids of meteorological drought events by SPI3 and SPEI3 during 1982–2020 are also presented in Figure 15 (b) and (c). The circles in the figures represent the centroid of drought events, while the colors represent drought durations and the sizes represent drought severities. As shown in Figure 15, the number of identified agricultural drought events is less than that of meteorological drought events. The identified agricultural drought events are mostly located in the Northeast, North, Southwest, and western parts of East China. In South China, there are some identified meteorological drought events, whereas there are rare agricultural drought events. The centroids of relatively more severe agricultural drought events are usually distributed between the Yangtze River basin and the Yellow River basin, which is roughly consistent with the meteorological drought events.

Figure 15. Spatial distribution of drought events during 1982–2020 identified by (a) SSMI1-CCI, (b) SPI3-CRA, and (c) SPEI3-CRA.

4.4 Temporal variations of agricultural drought

The severity of each identified drought event was used to illustrate the temporal variation of drought events in the study area. Figure 16 shows that there were mainly three drought periods during the past decades, the first roughly started in 1983 and reached its peak in 1986, the second appeared from 1994 to 1998 and reached its peak in 1998, and the third one was in 2010. Moreover, the agricultural drought events identified by the SSMI1-CCI droughts are sparser than the meteorological drought events by SPI3 and SPEI3, which is consistent with the results in Figure 15. Additionally, Mann–Kendall and Theil–Sen trends were performed on the temporal series. The Z Score values of Mann-Kendall for SSMI1-CCI, SPI3, and SPEI3 are − 1.36, −4.35, and − 2.93, respectively. The Theil-Sen slope values for SSMI1-CCI, SPI3, and SPEI3 are − 583.3, −1865.2, and − 864.8, respectively. The results demonstrate that the severity of identified agricultural drought events as well as the meteorological drought events has a decreasing trend during the past 40 years.

Figure 16. Temporal variation of drought severity during 1982–2020 based on (a) SSMI1-CCI, (b) SPI3-CRA, and (c) SPEI3-CRA.

To reveal the evolutionary trends of droughts, Mann-Kendall trend test was conducted on the SSMI1-CCI, SPI3-CRA, and SPEI3-CRA drought indexes. The results of Mann–Kendall trend analysis are shown in Figure 17. For the agricultural drought, as shown in Figure 17 (a), there are significant wetting trends over eastern parts of Tibet and Northeast China, western parts of North China, and South China. In contrast, there are significant drying trends over southern parts of North China, western parts of East China, and Southwest China. The identified regions with drying trends in agricultural drought were basically consistent with the result by SPEI3, as shown in Figure 17 (c). However, there are some differences between Figure 17 (a) and (c). The first significant difference is the area of drying trends, where that by SSMI1-CCI is greater than that by SPEI3-CRA. The second significant difference is in Northeast China, where the SPEI3-CRA indicated a significant drying trend, while the SSMI1-CCI did not indicate significant change. The third significant difference is in North China and East China, where the SPEI3-CRA indicated some significant wetting trend areas, while the SSMI1-CCI did not indicate significant change. Such differences may be because meteorological droughts do not necessarily cause agricultural droughts. Additionally, there are differences between Figure 17 (b) and (c) which can be interpreted by the fact that the SPI does not take air temperature or potential evapotranspiration into account for drought monitoring, whereas they are considered in the SPEI.

Figure 17. Mann-Kendall analysis for varied drought indexes: (a) SSMI1-CCI, (b) SPI3-CRA, and (c) SPEI3-CRA.

5. Discussion

5.1 Further explanation of the results

Understanding the spatial and temporal variations of drought is very important for preventing or mitigating the adverse impacts of drought. Previous research have paid more attention to meteorological drought but less attention to agricultural drought due to the lack of long-term and spatiotemporal continuous SM data. Recently, the ESA CCI SM dataset has presented great potential in investing agricultural drought dynamics. However, many data gaps exist in CCI SM dataset over China, which limits the spatial and temporal investigations. To address this issue, we first constructed a method for gap-filling based on a machine learning method and a set of SM covariates. Based on this gap-filling method, a long-term (1982–2020) and spatially continuous distributed SM dataset over China was produced. We evaluated this SM dataset with in-situ SM observations and manufactured SM data gaps. All evaluations demonstrated the reasonability of the gap-filled SM dataset. To the best of our knowledge, such SM dataset that prepared for drought monitoring and evaluation over China was rarely seen before this study.

Subsequently, we investigated the agricultural drought dynamics in China using this SM dataset by calculating a drought index SSMI and identifying the drought event with a space-time continuous technique. Identifying agricultural drought events based on the remotely sensed SM data is another unique feature of this study. Results of spatial variation analysis demonstrate the following: 1) the agricultural drought was most frequent in North China. The second most frequent region included the western parts of East China, Northeast, and Southwest China. The least frequent region was South China. 2) The identified agricultural drought events are mostly located in Northeast, North, Southwest, and western parts of East China. The centroids of relatively more severe agricultural drought events were usually distributed between the Yangtze River basin and the Yellow River basin. Results of temporal variation analysis demonstrated the following: 1) The severity of agricultural drought events varied in a decreasing trend from 1982 to 2020. 2) Significant drying trends were found for southern parts of North China, western parts of East China, and Southwest China. The comparisons with other studies are presented in the following section.

5.2 Comparison with other studies

The spatial and temporal variations of agricultural drought identified in this study are basically consistent with some previous studies but also have unique characteristics. For spatial variation, Yu et al. (2014) investigated the drought dynamics in China during 1951–2010 using the SPEI index at 609 meteorological stations, where 3-month SPEI was used for drought duration and frequency analysis. They found that the North, Northeast, and western Northwest of China had a higher frequency of persistent multi-year severe droughts. Instead of using drought indexes at meteorological stations, Xu et al. (2015) investigated the drought dynamics in China during 1961–2012 based on gridded SPI, SPEI, and RDI. Moreover, a three-dimensional clustering method was developed to identify drought events rather than identifying drought in space and time separately. They found that large-magnitude droughts were usually centered in the region from North China Plain to the downstream of the Yangtze River. Zhou et al. (2017) investigated the drought dynamics in China during 2001–2013 using an Integrated Surface Drought Index (ISDI) which is a composite of SPI, PDSI, remote sensing vegetation conditions, and some biophysical datasets, such as DEM and land cover type. They found that the North of China has higher drought probabilities. Comparatively, we found that the agricultural drought indicated by remotely sensed SM anomaly was the most frequent in North China and the least frequent in South China. We also found that the identified agricultural drought events were mostly located in Northeast, North, Southwest, and western parts of East China.

From the perspective of temporal variation, Wang et al. (2015) found that the drought-impacted areas in China as a whole decreased slightly from 1982 to 2011. Moreover, dry trends were identified in northeastern and southwestern China. Yu et al. (2014) found significant drying trends in North China, the southwestern parts of Northeast China, the central and southwestern parts of Southwest China. Xu et al. (2015) demonstrated that the western part of the North China Plain, Loess Plateau, Sichuan Basin, and Yunnan-Guizhou Plateau showed drying trends significantly. Zhou et al. (2017) indicated that the drying trend was in the northeast of China and south of the Yangtze River. Recently, Zeng et al. (2020) investigated the agricultural drought dynamics in China during 1981–2019 using a remote sensing index: VHI. They found that the overall drought-affected area of China decreased from 1981 to 2019. Comparatively, we demonstrated that the severity of agricultural drought events identified by remotely sensed SM anomaly varied in a decreasing trend from 1982 to 2020. We also found significant drying trends in southern parts of North China, western parts of East China, and Southwest China. However, no significant dry trends were found by the remotely sensed SM anomaly in most areas of Northeast China. Moreover, no significant wet trends were found by the remotely sensed SM anomaly in most areas of North and East China. The area of drying trends identified by the remotely sensed SM anomaly is not the same as the meteorological drought by SPEI. Besides, we demonstrated that the number of identified agricultural drought events was significantly less than that of meteorological drought events during 1982–2020. These differences are unique characteristics found in this study.

To demonstrate the reasonability of the above findings, we collected the economic crop losses, drought-affected cropland area, drought-damaged cropland area, and the area of crop failure from 2006 to 2021 according to the “China Flood and Drought Disaster Prevention Bulletin” by the Ministry of Water Resources of the People’s Republic of China. Figure 18 presents the temporal variation of these statistical indicators during 2006–2021. According to the findings of this study, there are more areas with no significant dry/wet trends identified by SM anomaly than by meteorological drought indexes, the frequency of agricultural drought was less than the meteorological drought, and the severity of agricultural drought events varied in a decreasing trend during the past decades. As a result, drought disaster losses have decreased during the past decades. The statistical indicators shown in Figure 18 do support the findings of this study. A recent study by Deng et al. (2022) made a longer time-series analysis of the drought-affected area and drought-suffering area than that in Figure 18. They found that the multi-year average drought-affected area rate and drought-suffering area rate are 17.12% and 8.44% in 1981–2000 and 10.35% and 5.11% in 2001–2020, which also presented significant decreasing trend during the past decades.

Figure 18. Statistic of drought disaster losses during 2006–2021 from “China Flood and drought disaster prevention Bulletin”.

5.3 Contributions and limitations of this study

This study has three main contributions. First, a gapless and long-term satellite remote sensing SM dataset in China was contributed, which can be directly obtained from https://doi.org/10.57760/sciencedb.07849. This SM dataset includes monthly average soil moisture from 1982 to 2020 with a spatial resolution of 0.25° in the whole Chinese Mainland. Such soil moisture dataset is very useful for agriculture, hydrology, meteorology, global climate change, and other fields. Moreover, based on this dataset, a soil moisture dataset with higher spatial resolution can be produced with our previously developed downscaling techniques (Sun and Gao 2023; Sun, Zhang, and Zhao 2022).

The second contribution is the gap-filling method suggested in this study including the combination of input variables and the selection of machine learning methods. As for the machine learning methods, we have tried Gated Recurrent Unit (GRU) which is one of the Recurrent Neural Networks (RNNs). The GRU is recognized to be good at dealing with time-series data. However, our results showed that the gap-filling with GRU was slightly inferior to that with XGBoost when we compared the gap-filled SM with the in-situ measurements. The tested parameters of GRU contained the number of GRU layers and hidden size, the learning rate, maximum epochs, and batch size, which were set to 0.001, 1000, and 1024, respectively. The number of layers was changed from 1 to 3, and the hidden size had two options which were 256 and 512. The results are listed in Table 3 which showed that the performance of XGBoost was better than that of GRU. The results demonstrate the validity of the suggested gap-filling method. Likewise, recent research indicated that tree-based models (e.g. Random Forest, XGBoost, and Gradient Boosting Tree) still outperform deep learning methods on tabular data (Grinsztajn, Oyallon, and Varoquaux 2022). To some extent, the CCI SM is such a form of tabular data. Certainly, more attempts at deep learning methods are encouraged to modify the suggested gap-filling method.

Table 3. Comparison between in-situ SM and SM predicted by GRU or XGBoost.

Model	R	RMSE	ubRMSE	bias	Layers	Hidden Size
GRU-1	0.2946	0.1025	0.1023	0.0060	1	256
GRU-2	0.3755	0.0976	0.0975	0.0040	2	256
GRU-3	0.2736	0.1033	0.1032	−0.0025	3	256
GRU-4	0.3033	0.1015	0.1014	0.0037	1	512
GRU-5	0.2805	0.1029	0.1029	−0.0011	2	512
GRU-6	0.2617	0.1037	0.1037	−0.0009	3	512
XGBoost	0.5540	0.0977	0.0894	−0.0395	-	-

Revealing the importance of input variables in the gap-filling method contributes to understanding the validity of the suggested gap-filling method in producing spatially continuous distributed SM. In this study, the input variables include three soil component proportions (sand proportion, silt proportion, and clay proportion), air temperature, surface temperature, precipitation, antecedent precipitation, NDVI, albedo, elevation, time, and reanalysis SM from CRA-Land. The importance was measured by the increased percentage of RMSE. First, all the input variables were involved in the gap-filling method to predict SM and calculate the basic RMSE between the predicted SM and the original SM. Second, a combination of input variables that excluded a target variable was involved in the gap-filling method to calculate the target RMSE. The importance of this target variable can then be obtained by importance = (target RMSE − basic RMSE)/basic RMSE. Figure 19 presents the mean importance of each input variable during the study period. The results demonstrated that elevation was found to be the most important variable with a mean importance of 12.75%, which is reasonable because of the complex terrain over the Chinese Mainland. Reanalysis SM also contributes greatly to the gap-filling method with a mean importance of 10.1%. Temporal information denoted as year showed the third importance (7.47%) due to the annual variation of soil moisture. Precipitation, NDVI, and albedo also made important contributions (1.24%, 1.66%, and 2.38%) to the gap-filling method, because of their ability to reflect vegetation growth status and surface energy fluxes which are highly related to soil moisture. In addition to these, air temperature and three soil component proportions were found to have relatively low importance. The importance analysis of input variables implies that more efforts can be put into developing more effective SM covariates for the gap-filling method.

Figure 19. Importance of input variables in the gap-filling method depicted by the increased percentage of RMSE.

Third, we presented a depiction of agricultural drought dynamics in China from 1982 to 2020 with the gapless and long-term satellite SM dataset. As it has been found, the original CCI SM data have many data gaps. Based on the original CCI SM data, the spatially continuous drought index cannot be obtained. It is difficult to reveal the complete characteristics of agricultural droughts with the original CCI SM because the spatial clusters and temporal connections are interrupted by the gaps (see Figure 20). Trend analysis is also hard to perform on incomplete time series. Therefore, the accurate temporal change trend of drought cannot be estimated with the original CCI SM. In contrast, the suggested gap-filling technique promoted the integrality of CCI SM data greatly. The difference between the original CCI SM and gap-filled CCI SM can be easily found in Figure 20, taking the data in December 1982, for example. As we know, the agricultural drought dynamics have been investigated using some indirect indicators of soil moisture such as Vegetation Health Index (VHI) (Zeng et al. 2020), Integrated Surface Drought Index (ISDI) (Zhou et al. 2017), SPI3/SPEI3/RDI3 (Xu et al. 2015; Yu et al. 2014), and Normalized Difference Vegetation Index (NDVI) anomaly (Wang et al. 2015). Here, we provided an investigation using remotely sensed SM dataset, which is an important complement to existing researches.

Figure 20. Visualization of the (a) original CCI SM and (b) gap-filled SM in December, 1982.

Certainly, there is some room for further improvement in this study. First, the gapless and long-term satellite remote sensing SM dataset provided in this study has a temporal resolution of 1 month and a spatial resolution of 0.25 degree. The original temporal resolution of the ESA CCI SM dataset is daily. If the set of SM covariates constructed in this study can be obtained in a daily resolution, the gapless and long-term SM dataset can also be produced in a daily resolution even though the process is very computationally intensive. If so, the agricultural drought can be depicted on a finer timescale. As to the spatial resolution, the original size was retained. Such a spatial resolution of around 20–30 km would be too coarse for investigating drought dynamics over a small watershed or a local area. However, it is probably suitable for investigating the drought dynamics over the whole area of China. For example, Xu et al. (2015) investigated the spatiotemporal variation of drought in China during 1961–2012 using parameters such as precipitation, air temperature, and wind speed at the same spatial resolution of 0.25 degree. What’s more, Deng et al. (2022) investigated the spatiotemporal pattern of drought impacts on agriculture in China at a provincial spatial scale. Certainly, higher spatial resolution is essential for depicting the agricultural drought in more detail, which can be further improved with downscaling techniques (Sun, Zhang, and Zhao 2022). At last, obtaining timeliness and detection depth of the gapless and long-term satellite remote sensing SM dataset deserves more improvement research. The original ESA CCI SM dataset was only produced every year, i.e. there is a one-year delay. Such delay limits near-real-time drought monitoring, which can be further improved by integrating the historical CCI SM data with near-real-time SM data such as SMAP. As for the detection depth issue, it has always been a “pain point” of the remote sensing SM techniques. Fortunately, methods based on land surface models, machine learning methods, or a combination of them have been explored to estimate root-zone SM (A et al. 2022). The ESA CCI is also working on developing long-term satellite-based root-zone soil moisture products (Dorigo et al. 2017), which would play a significant role in understanding agricultural drought. To sum up, gapless and long-term satellite remote sensing SM with faster obtaining-timeliness, higher spatiotemporal resolution, and greater detection depth are required for understanding agricultural drought more comprehensively and accurately.

6. Conclusion

This study can be mainly concluded as follows:

1. To investigate the agricultural drought dynamics of China, a gapless satellite remote sensing SM dataset during 1982–2020 was created with a spatial resolution of 0.25° and a temporal resolution of 1 month. The original SM and gap-filled SM were evaluated with in-situ SM measurements from six networks of ISMN. Overall, the original SM has R of 0.45, RMSE of 0.0907 cm³/cm³, ubRMSE of 0.0884 cm³/cm³, and a bias of −0.0203 cm³/cm³. The gap-filled SM has R of 0.55, RMSE of 0.0977 cm³/cm³, ubRMSE of 0.0894 cm³/cm³, and a bias of −0.0395 cm³/cm³. The gap-filled SM presented comparable or even better performances as compared with the original SM.

2. An agricultural drought index, i.e. SSMI was calculated using the gapless and long-term satellite SM dataset for investigating the spatial and temporal variation of agricultural drought. The agricultural drought was found to be the most frequent in North China. The second most frequent regions are the western parts of East China, Northeast, and Southwest China. The least frequent region is South China. The regions with a significant drying trend were found mostly in the southern parts of North China, western parts of East China, and Southwest China.

3. Agricultural drought events were identified from long-time series SSMI with a 3-dimensional clustering method. The centroids of identified agricultural drought events were found mostly in Northeast, North, Southwest, and western parts of East China. The centroids of relatively more severe agricultural drought events were usually distributed between the Yangtze River basin and the Yellow River basin. The severity of agricultural drought events presented a decreasing trend from 1982 to 2020.

4. The agricultural drought dynamics revealed in this study are basically consistent with the other studies. Nevertheless, the agricultural drought dynamics investigated with SM have their own unique features, such as more space details and less drought frequency or count than that depicted by the 3-month SPI and SPEI.

Monitoring, evaluating, and understanding agricultural drought requires long-time series SM data with good quality and spatiotemporal continuity. Here, such an SM dataset was presented by reprocessing the CCI SM dataset. The SM dataset provided in this study can be further used in agriculture, hydrology, meteorology, climate change, and other fields. The methods suggested in this study can also be referred for studies in other areas. The agricultural drought dynamics investigated in this study contribute to national or regional drought risk management. Future studies for such agricultural or ecological drought are encouraged in developing long-term satellite remote sensing SM with faster obtaining timeliness, higher spatiotemporal resolution, and greater detection depth.

Acknowledgments

The author would like to thank any agency or organization that provided the remote sensing data resources, reanalysis data resources, and in-situ meteorological data resources used in this study.

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

The data supporting the findings of this research are publicly available at https://www.esa-soilmoisture-cci.org/data, https://doi.org/10.5067/MODIS/MOD13C2.006, https://developers.google.com/earth-engine/datasets/catalog/NASA_GIMMS_3GV0, http://data.cma.cn/data/cdcdetail/dataCode/NAFP_CRA40_FTM_MON_NC.html, https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/, https://earthexplorer.usgs.gov/, https://ismn.earth/en/, https://www.scidb.cn/en/detail?dataSetId=434a6fad0daa4c32a7396dc7db575fc8.

Additional information

Funding

This work was supported by the Beijing Natural Science Foundation under Grant 6222045.