Detection of a Real Time Remote Sensing Indices and Soil Moisture for Drought Monitoring and Assessment in Jordan

Drought monitoring represents a challenge for water and agricultural sector as this natural hazard accelerates water deficiency and leads to adverse environmental and socioeconomic impacts. The use of remote sensing data and geospatial techniques to monitor and map drought severity expanded in the last decades with progressive developments in data sources and processing. This study investigates the correlations among drought indices derived with soil moisture stress (K) obtained from ground data collected from fields cultivated with barley. The study, carried out in Yarmouk basin in the north of Jordan, includes NDVI, PDI, MPDI and PVI derived from Landsat 8-OLI and Sentinel 2-MSI. Results showed different behavior among the indices and throughout the 2016/2017 growing season, with maximum correlation between PDI and MPDI followed by NDVI with PVI. Correlations among the remote sensing indices and K for different soil depths during March-April were significant for most indices with a maximum (R 2 ) of 0.82 for K 30-50 and MPDI, followed by K 30-50 with NDVI. Drought severity maps for the month of March showed different trends for the different indices, with similarities between MPDI and PDI. The map of drought severity combined from the remote sensing indices and K showed that PDI and soil moisture could significantly explain 56% of variations in spatial patterns of drought, while the combination of MPDI, PDI and NDVI could significantly explain up to 59% of variations in drought severity map. Therefore, the study recommends the adoption of these remotely sensed indices for monitoring and mapping of agricultural droughts.


Introduction
Drought is a natural disaster that has many interrelated environmental and socioeconomic impacts related to the lack or shortage of water. With time, definitions of drought included four types that are: meteorological, agricultural, hydrological and socioeconomic [1]. The scientific consensus on drought defines this phenomenon as the condition of insufficient moisture caused by deficit precipitation over a period of time. Therefore, agricultural drought remains the most important type as the level of soil moisture content largely controls plant growth through the growing season. Since agricultural drought is highly correlated with soil moisture, monitoring of drought using soil moisture observations or indices related to soil moisture can provide important information for drought early warning systems.
Among many challenges that stand as obstacles for monitoring soil moisture conditions and drought severity during the different seasons is the good spatial distribution of instruments for measuring soil moisture and high cost of in-situ measurements for large geographical areas. The alternatives for the in-situ soil moisture measurements are the land surface models that integrate climatic and remote sensing data [2]. These coupled and uncoupled models for soil moisture estimation are limited by the level of accuracy which is highly limited by the input data for initial conditions and the coarse spatial resolution [3]. Alternatively, several techniques have revealed the ability of remote sensing data to extract indices that could indicate soil moisture and reflect drought conditions. The advantages of remote sensing techniques as potential tool for detecting for drought monitoring are their abilities to reveal the main three characteristics of drought which are intensity, duration and spatial coverage [4]. The techniques of remote sensing data are based on the assimilation of digital numbers (DNs), representing surface spectral reflectance at certain wavelength, to derive indices that are related to drought and reflecting soil moisture conditions. Most of these indices are derived from the red and near-infrared bands and assumed to reflect vegetation fractions and conditions. During the last two decades, progressive developments in the geospatial techniques of remote sensing and geographic information systems (GIS) encouraged the use of remote sensing indices for monitoring drought and modeling its spatial patterns at different spatial and temporal scales [5].
The use of remote sensing data also increased due to its availability at reason- tion Index (NDVI). The index shows good response to cumulative rainfall, except in arid areas or in environments with dense vegetation [6].
For drought monitoring, the concept of perpendicular drought index (PDI) was proposed to reflect moisture distribution in the red (R)-near-infrared (NIR) space [7]. The limitations of PDI are related to the assumptions of homogenous land cover and soil types. Therefore, a modified perpendicular drought index (MPDI) that considers vegetation fraction and soil moisture content was proposed for drought monitoring [8] The MPDI derived from the Landsat Enhanced Thematic Mapper Plus (ETM+) and the Moderate Resolution Imaging Spectroradiometer (MODIS) showed better performance than PDI in terms of drought monitoring [8].
In Jordan, adverse trends of climate changes and increased frequency and severity of droughts are expected to add more stresses to the countries scarce water resources [9] [10]. The country's water sector policy for drought management emphasized the need for a national drought early warning system to form the basis for decision-making and effective drought management planning [11]. This would require the adoption and use of credible indices that can detect drought and reflect soil moisture conditions, for the case of rainfed areas. At present, a unit for monitoring agricultural droughts is operating at the National Agricultural Research Center (NARC). The maps produced by this unit are based on the NDVI data of MODIS with spatial resolution of 1000. In these maps, severity of drought is based on the degree of NDVI deviation from the long-term means [12].
As a possible improvement in drought mapping, the drought monitoring unit at NARC started to use the 250 m resolution data of MODIS to produce maps of drought [13]. The maps were not assessed in terms of accuracy or correlation with soil moisture. For small geographical areas, the moderate resolution data of Landsat showed to be more accurate than MODIS data [14]. The use of either Landsat or MODIS data would depend on calibration and validation of indices using ground data.
The use of active remote sensing data of RADARSAT II to monitor soil moisture was also investigated in Jordan and showed accurate predictions [15]. However, several factors might restrict the use of such data sources including the low temporal resolution, the shallow soil depth represented by these data and the relatively high cost of moderate resolution data. Therefore, this study aims to assess different remotely sensed indices, derived from passive data, and their correlation with soil moisture for the purpose of agricultural drought monitoring in the north of Jordan. In addition, the study compares maps of drought generated from the different indices in terms of spatial distribution and rainfall gradient along the study area.

Study Area
The study was carried out in an area of 210 km 2 inside the Yarmouk River Basin The study area inside YRB was selected to cover rainfed crop of barley across the rainfall gradient inside the basin. The study area was selected as the rainfed areas inside this basin are under the threat of drought and desertification [15] [16].
The mean annual rainfall in the basin ranges between 150 mm in the east to 400 mm in the west, while the selected area has an annual rainfall of 230 -350 mm.
The study area has mean annual minimum and maximum temperatures of 12.3˚C and 23.1˚C, respectively.
The annual potential evaporation at the basin's level ranges from 1500 to 2150 mm/year [16]. The study area is flat and forms part of Horan plains with an altitude range of 560 -600 m. Soils of the study area are deep in the west and shallow (depth < 70 cm) in the east with moderate to high clay contents. Both of Figure 1. Location of the study area in the Yarmouk River Basin [11]. In the eastern parts of the basin and the study area, a transhumant system of cultivation is practiced where barley is usually cultivated for straw rather than for grain, to support the grazing herds of sheep, especially when drought is detrimental to grain yield.

Methods
The study methodology included a sequence of steps and procedures of data collection and analysis, as summarized in Figure 2. The data used in the study included satellite images of Landsat 8-Operational Land Imager (OLI), Sentinel 2-Multispectral Instrument (MSI) and climatic data from Jordan Meteorological Department (JMD) and Ministry of Water and Irrigation (MWI), in addition to ground data collected during several field visits. The following subsections include detailed description of these steps and procedures.

Image Processing
Landsat 8-OLI and Sentinel 2-MSI images for the 2016/2017 seasons were downloaded and processed to derive different drought indices that were correlated with soil moisture. The 30 m spatial resolution of the OLI data was considered as high resolution for drought mapping. However, due to cloud cover in some periods during the winter season, the MSI data with 10 m resolution was used for these periods. The two were characterized by having high correlated spectral data for atmospherically corrected images, i.e. data of spectral reflectance at ground level [17]. Since the study was based on multi-temporal images, atmospheric corrections were carried out using ground readings for reference objects collected by a hand held radiometer for one image and carrying out an image-to-image correction for the remaining images [10].  Table 1).

Derivation of Drought Indices
The processed data of OLI and MSI was used to derive four indices selected for drought monitoring as summarized in the following points.

1) NDVI
This is the most commonly used index to monitor vegetation. The index was calculated by using red and NIR bands as follows [6]: where NIR and R are the reflectance of near-infrared red bands, respectively.

2) PVI
The perpendicular vegetation index (PVI) is distinct; orthogonal to the soil line and can be used for computing the maximum signals of the green vegetation taking into account the effect of soil background. The index was calculated as follows [18]: where the subscripts s and v refer to the soil and vegetation reflectance, respectively. The soil line was generated by extracting the values of red and NIR bands for the study area to derive the slope and intercept. Locations of the sampled sites, in terms of red and NIR reflectance, were plotted on this cure to study the variations in vegetation cover among the sampled sites and to determine which of these sites would be highly affected by possible drought or soil moisture stress.
where M is the slope of the soil line in the Red-NIR spectral feature space. The value (1.30) was derived from using the linear relationship between R and NIR [14].

4) MPDI
The range of f v is 0 (bare soil) to 1 (full vegetation cover). After deriving the f v layer, the MPDI was calculated as follows [8]:

Ground Data Collection
Soil moisture samples were collected from ten sites (fields) in the study area during the 2016/2017 season. The sites were selected to represent an east-west transect that covers the rainfall gradient in the study area, with a distance of 3 -4 km between the field and the other ( Figure 3). Also, the selection included the fields that were planted on the same data (10th of December 2016). These included eight rainfed fields cultivated with barley, one field cultivated with barley under supplemental irrigation and one bare field that included natural grasses used for grazing. The sites were characterized by flat topography and fine to medium soil texture that had high water holding capacity.   Table 2. In addition to soil sampling, measurements were made for plant heights inside the 10 sites using a diagonal transect inside each site during January-May. Gravimetric soil samples were collected in the days coincided with the satellite overpass. Samples were taken by an auger from three depths extending from surface down to 50 cm (0 -10, 10 -30 and 30 -50). Physical characterization for particle size distribution, bulk density and volumetric water contents at field capacity and permanent wilting point was also carried out for the collected samples.
The volumetric soil moisture content (θ v ) was measured for each soil sample and the equivalent soil moisture content (W) was then calculated for each depth and for each sampling site. Soil moisture stress (K) was calculated using the following equation [21]: where, W is the equivalent soil moisture content for a particular soil depth or layer, W h and W p are the Field Capacity (FC) and Permeant Wilting Points (PWP), respectively, for that soil depth or layer. When the value of K approaches zero then plant will not suffer from water stress, while values closer to one indicate conditions of water stress resulting from low levels of soil moisture.

Drought Severity Mapping
A map of drought severity was produced for the month of March using the maps of drought level based on PDI, MPDI, PVI and K maps and using the ranges  shown in Table 3. The NDVI was excluded as drought classification using this index would require the use of long historical record and the use of seasonal deviations from the long term mean assuming normal distribution of data [12]. The indices used for drought mapping, on the other hand, did not require the assumption of normal distribution and the map of drought could be carried out for a single image [21]. The month of March was selected as growth of rainfed crops would reach its peak during this month [12] [14].

Evaluation of Results
The derived remote sensing indices were assessed in terms of correlation with soil moisture stress for the different soil layers and for the root zone. A correlation matrix was used to analyze the extent of interdependence between different drought indices. In addition, other statistical parameters of minimum, maximum, mean, standard deviation, Root Mean Square Error (RMSE) were also derived and summarized for the derived indices ( Table 4). The data of vegetation indices was also plotted for the sampling sites during the growing season using    Table 5).

Assessment of Vegetation and Drought Indices in the Sampled Sites
The maximum values of PVI were observed in the MSI image of the 12th of March with higher values for sites of 1, 6, 8 and 10 than other sites (Figure 4). Similarly, the maximum values of NDVI were observed during the same period. Both of PVI and NDVI agreed with the field observations and measurements collected for Plant Height (PH) inside the sampling sites (see Table 5). As such, significant correlations (p < 0.05) between plant height and PVI and NDVI were observed in the sites cultivated with barley, where the coefficient of determination (R 2 ) was 0.51 for PH-PVI and 0.85 for PH-NDVI relationships.
The relatively low value of R 2 for the case of PVI was attributed to its high sensitivity to the soil brightness [24]. Data from both OLI and MSI detected the variations in soil moisture for bare soils throughout the growing season ( Figure  4). This was reflected on the spectral space of the red-NIR that showed high reflectance for dry soils than wet soils due to absorbance of electromagnetic radiation in the red and NIR wavelength. Among the examined indices, PDI was able to detect these variations during the growing season especially when soil was bare.
Plotting the data of PDI for dry and wet periods for both bare and vegetated soils, respectively, showed that PDI was parallel to the soil line and responded to soil moisture levels during the season ( Figure 5). The line L, which dissected the coordinate origin and is vertical to the soil line, delineated in Figure 5 and formulated from the soil line expression [7] would form the basis for the PDI concept.    Results showed significant correlations among some indices and no correlations among others ( Table 6). The correlation coefficient reflected the nature of the remotely sensed indices which could be grouped into two groups that included PVI and NDVI in one group and PDI and MPDI in another group. The PVI and NDVI indices were mainly reflecting seasonal vegetation condition and showed higher correlations with the seasonal soil moisture stress. Previous work in Jordan [6] showed that NDVI was highly correlated cumulative rainfall rather than monthly or single rainfall event. The PDI and MPDI, on the other hand, were highly correlated and showed positive correlation with soil moisture. Both indices were low in January and February, indicating drought during this period when most rainfall occurred. This trend was not detected by PVI which indicated severe drought conditions during January and February and in other periods, except in March.
In general, PDI and MPDI reflected the conditions of drought in the study area with better response to rainfall than PVI. The values of NDVI, on the other hand, that maximum vegetation growth occurred during March and early April (at most), as indicated by the NDVI profile plotted for the sampled site ( Figure   6). Since most of rainfall occurred during December-March, then this period could be considered as the critical period for monitoring drought in this Mediterranean environment. These results were also reported by previous research in Jordan [12] [13] [14] [25] which indicated that peak vegetation growth is reached by the end of March.

Correlation among Indices and Soil Moisture
Results showed variations in the degree of correlation between remotely sensed drought indices and soil moisture stress (K) during the growing season and for short-term drought conditions (Table 7). For the period January-April, the relationship between remote sensing indices and K was insignificant for NDVI, while low values of R 2 were observed for the relationships between other indices   Generally, MPDI showed better correlations with soil moisture stress when compared with PDI. High correlations between MPDI and subsurface soil moisture was also indicated by Ghulam et al. [8] who proposed MPDI for real time drought monitoring.
The significant correlations with relatively high R 2 between NDVI, PVI and K could be explained by the fact that vegetation growth reached its peak during March and early April when soil moisture was depleted by plants. In this particular period, strong correlations were observed for subsurface soil moisture and remote sensing indices. The strong correlation between NDVI and soil moisture during March-April would also indicate the response of NDVI to cumulative rainfall during the season, as confirmed by previous research [6]. The relationships between drought indices and K were linear with PDI and MPDI being proportionally correlated with K, and NDVI and PVI being inversely correlated with K. An example on these relationships is shown in Figure 7 for MPDI and NDVI during March-April. The inverse relationships between NDVI and K might be confusing when  this stage without reflecting drought severity. Also, the NDVI is well known for its sensitivity to soil background, soil wetness and saturation in its values for areas with dense vegetation [6] [26]. Therefore, MPDI would be recommended for drought monitoring during the periods of soil moisture stress in March and April as its values would reflect the severity of drought in relation to plant conditions. In terms of R 2 , results were close to the values obtained by a previous study that correlated MPDI and PDI with soil moisture [14].
The results and relationships obtained from this study, however, would be more advantageous when compared with correlations made directly with soil moisture as the level of moisture stress would differ among the different soil types at the same soil moisture levels. Although both of PDI and MPDI would show similar spatial patterns (Figure 8), however, MPDI may perform better at stages of dense vegetation cover [27], like site 10 which had denser cover than other rainfed sites. In general, MPDI would reflect drought severity in relation to soil moisture stress regardless of soil type. Since the values of R 2 were slightly higher for K-MPDI relationship than for K-PDI relationship, the use of MPDI to monitor drought would be recommended.

Drought Severity Mapping
Maps of drought severity based on PDI, MPDI, PVI and K for March (see   Figure 10). The map showed that half of the study area was exposed to moderate and severe drought while the eastern parts were suffering from extreme drought.
The map agreed with field observations collected during the growing season (     sensing drought indices of NDVI, PDI and MPDI could explain drought 56% (R 2 = 0.56) of variations in drought spatial pattern.
Adding K to these indices would increase the model redundancy as shown by the VIF values that increased when the four variables were used to explain the variations in drought severity map, as the VIF value reached 6.9 which became close to 7.5; the level at which model redundancy would start [23]. These results could indicate that K or remote sensing indices highly correlated K should form the basis for monitoring drought and mapping its severity.

Conclusion
Results showed that remote sensing indices showed different responses to soil moisture stress (K) with similarities in trends of correlations, which were positive for MPDI and PDI and negative for NDVI and PVI. The correlations were significant with relatively high values of R 2 for the MPDI, NDVI, PVI and K

Data Availability
Some or all data, models, or code generated or used during the study are proprietary or confidential in nature and may only be provided with restrictions (e.g. maps and climatic data).