Effective Band Ratio of Landsat 8 Images Based on VNIR-SWIR Reflectance Spectra of Topsoils for Soil Moisture Mapping in a Tropical Region

Effective mapping and monitoring of soil moisture content (SMC) in space and time is an expected application of remote sensing for agricultural development and drought mitigation, particularly in the context of global climate change impact, given that agricultural drought is occurring more frequently and severely worldwide. This study aims to develop a regional algorithm for estimating SMC by using Landsat 8 (L8) imagery, based on analyses of the response of soil reflectance, by corresponding L8 bands with the change of SMC from dry to saturated states, in all 103 soil samples taken in the central region of Vietnam. The L8 spectral band ratio of the near-infrared band (NIR: 850–880 nm, band 5) versus the short-wave infrared 2 band (SWIR2: 2110 to 2290 nm, band 7) shows the strongest correlation to SMC by a logarithm function (R2 = 0.73 and the root mean square error, RMSE ~ 12%) demonstrating the high applicability of this band ratio for estimating SMC. The resultant maps of SMC estimated from the L8 images were acquired over the northern part of the Central Highlands of Vietnam in March 2015 and March 2016 showed an agreement with the pattern of severe droughts that occurred in the region. Further discussions on the relationship between the estimated SMC and the satellite-based retrieved drought index, the Normal Different Drought Index, from the L8 image acquired in March 2016, showed a strong correlation between these two variables within an area with less than 20% dense vegetation (R2 = 0.78 to 0.95), and co-confirms the bad effect of drought on almost all areas of the northern part of the Central Highlands of Vietnam. Directly estimating SMC from L8 imagery provides more information for irrigation management and better drought mitigation than by using the remotely sensed drought index. Further investigations on various soil types and optical sensors (i.e., Sentinel 2A, 2B) need to be carried out, to extend and promote the applicability of the prosed algorithm, towards better serving agricultural management and drought mitigation.


Introduction
Soil moisture content (SMC) is a key parameter that needs be monitored in order to provide information for plant growth and crop management, as well as to understand water-associated hazards Soil samples from the northern part of the Central Highlands of Vietnam were first processed into a dry state at SMC = 0% by oven-drying until the weight remained constant, and then refined to a fine grain size (<0.01 mm), to subtract the effects of the grain size difference on soil reflectance. Each soil sample was then sub-sampled into 10 lots of 20 g samples, to which 10 various amounts of water were added, to create a series of soil samples with 10 various SMC levels. A total of 80 soil samples were created from eight samples taken from the northern part of the Central Highlands of Vietnam ( Figure 1b).

SMC Determination and Spectral Measurement
The spectral reflectance of these 80 soil samples were measured in a controlled laboratory environment, using an ASD FieldSpec 3 Max Portable spectroradiometer (ASD. Inc. Boulder, Colorado, USA) with fiber optics and an illumination source (ASD Pro Lamp (14.5 V, 50W), Boulder, At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.
Soil samples from the northern part of the Central Highlands of Vietnam were first processed into a dry state at SMC = 0% by oven-drying until the weight remained constant, and then refined Remote Sens. 2019, 11,716 4 of 18 to a fine grain size (<0.01 mm), to subtract the effects of the grain size difference on soil reflectance. Each soil sample was then sub-sampled into 10 lots of 20 g samples, to which 10 various amounts of water were added, to create a series of soil samples with 10 various SMC levels. A total of 80 soil samples were created from eight samples taken from the northern part of the Central Highlands of Vietnam (Figure 1b). Table 1. Features of soil samples and soil types in the study basing the description of FAO [28].

Kind of Soil
No. of Sample Collection

Representative Color Indicator Features Soil Characteristics
Rhodic ferralsols 1 bands; (3) finally, a discussion is made on the relationship of the estimated SMC and the Normal Difference Drought Index (NDDI) [26] in drought assessment and mapping.

Soil Sampling and Processing
Top soil samples were taken from 21 points distributed over the Nam Can commune of Nghe An Province in April 2018 (13 topsoil samples) and the northern part of the Central Highlands of Vietnam (eight topsoil samples) in April 2017, as shown in Figure 1A,B, respectively. According to the soil map of Vietnam [27], a total of 21 soil samples (Table 1) were taken from six soil types covering both sampling areas, including: ferric acrisols (15 samples), orthic ferralsols (two samples), rhodic ferralsols (one sample), acric ferralsols (one sample), eutric fluvisols (one sample), and dystric gleysols (one sample). Table 1. Features of soil samples and soil types in the study basing the description of FAO [28]. At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.

Kind of Soil
Remote Sens. 2019, 11, x FOR PEER REVIEW 3 of 18 bands; (3) finally, a discussion is made on the relationship of the estimated SMC and the Normal Difference Drought Index (NDDI) [26] in drought assessment and mapping.

Soil Sampling and Processing
Top soil samples were taken from 21 points distributed over the Nam Can commune of Nghe An Province in April 2018 (13 topsoil samples) and the northern part of the Central Highlands of Vietnam (eight topsoil samples) in April 2017, as shown in Figure 1A,B, respectively. According to the soil map of Vietnam [27], a total of 21 soil samples (Table 1) were taken from six soil types covering both sampling areas, including: ferric acrisols (15 samples), orthic ferralsols (two samples), rhodic ferralsols (one sample), acric ferralsols (one sample), eutric fluvisols (one sample), and dystric gleysols (one sample). Table 1. Features of soil samples and soil types in the study basing the description of FAO [28]. At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.

Kind of Soil
Highest iron, strong acidity and low phosphorus, rarely affected by erosion, very few reserves of available minerals, easily lost topsoil organic matter, low resilience and moderate sensitivity Orthic ferralsols 2 Remote Sens. 2019, 11, x FOR PEER REVIEW 3 of 18 bands; (3) finally, a discussion is made on the relationship of the estimated SMC and the Normal Difference Drought Index (NDDI) [26] in drought assessment and mapping.

Soil Sampling and Processing
Top soil samples were taken from 21 points distributed over the Nam Can commune of Nghe An Province in April 2018 (13 topsoil samples) and the northern part of the Central Highlands of Vietnam (eight topsoil samples) in April 2017, as shown in Figure 1A,B, respectively. According to the soil map of Vietnam [27], a total of 21 soil samples (Table 1) were taken from six soil types covering both sampling areas, including: ferric acrisols (15 samples), orthic ferralsols (two samples), rhodic ferralsols (one sample), acric ferralsols (one sample), eutric fluvisols (one sample), and dystric gleysols (one sample). At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.
High iron, strong acidity and low phosphorus, rarely affected by erosion, very few reserves of available minerals, easily lost topsoil organic matter, low resilience and moderate sensitivity Acric ferralsols 1 Remote Sens. 2019, 11, x FOR PEER REVIEW 3 of 18 bands; (3) finally, a discussion is made on the relationship of the estimated SMC and the Normal Difference Drought Index (NDDI) [26] in drought assessment and mapping.

Soil Sampling and Processing
Top soil samples were taken from 21 points distributed over the Nam Can commune of Nghe An Province in April 2018 (13 topsoil samples) and the northern part of the Central Highlands of Vietnam (eight topsoil samples) in April 2017, as shown in Figure 1A,B, respectively. According to the soil map of Vietnam [27], a total of 21 soil samples (Table 1) were taken from six soil types covering both sampling areas, including: ferric acrisols (15 samples), orthic ferralsols (two samples), rhodic ferralsols (one sample), acric ferralsols (one sample), eutric fluvisols (one sample), and dystric gleysols (one sample). At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.
Moderate iron, strong acidity and low phosphorus, rarely affected by erosion, very few reserves of available minerals, easily lost topsoil organic matter, low resilience and moderate sensitivity

Ferric acrisols 15
Remote Sens. 2019, 11, x FOR PEER REVIEW 3 of 18 bands; (3) finally, a discussion is made on the relationship of the estimated SMC and the Normal Difference Drought Index (NDDI) [26] in drought assessment and mapping.

Soil Sampling and Processing
Top soil samples were taken from 21 points distributed over the Nam Can commune of Nghe An Province in April 2018 (13 topsoil samples) and the northern part of the Central Highlands of Vietnam (eight topsoil samples) in April 2017, as shown in Figure 1A,B, respectively. According to the soil map of Vietnam [27], a total of 21 soil samples (Table 1) were taken from six soil types covering both sampling areas, including: ferric acrisols (15 samples), orthic ferralsols (two samples), rhodic ferralsols (one sample), acric ferralsols (one sample), eutric fluvisols (one sample), and dystric gleysols (one sample). At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.
Acid, low base status (< 50% base saturation) and strongly leached, poor in nutrient, easily soil degradation with agricultural practices Eutric fluvisols 1 Remote Sens. 2019, 11, x FOR PEER REVIEW 3 of 18 bands; (3) finally, a discussion is made on the relationship of the estimated SMC and the Normal Difference Drought Index (NDDI) [26] in drought assessment and mapping.

Soil Sampling and Processing
Top soil samples were taken from 21 points distributed over the Nam Can commune of Nghe An Province in April 2018 (13 topsoil samples) and the northern part of the Central Highlands of Vietnam (eight topsoil samples) in April 2017, as shown in Figure 1A,B, respectively. According to the soil map of Vietnam [27], a total of 21 soil samples (Table 1) were taken from six soil types covering both sampling areas, including: ferric acrisols (15 samples), orthic ferralsols (two samples), rhodic ferralsols (one sample), acric ferralsols (one sample), eutric fluvisols (one sample), and dystric gleysols (one sample). At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.
Suitable for intensive agriculture, high resilience and low sensitivity Dystric gleysols 1 Difference Drought Index (NDDI) [26] in drought assessment and mapping.

Soil Sampling and Processing
Top soil samples were taken from 21 points distributed over the Nam Can commune of Nghe An Province in April 2018 (13 topsoil samples) and the northern part of the Central Highlands of Vietnam (eight topsoil samples) in April 2017, as shown in Figure 1A,B, respectively. According to the soil map of Vietnam [27], a total of 21 soil samples (Table 1) were taken from six soil types covering both sampling areas, including: ferric acrisols (15 samples), orthic ferralsols (two samples), rhodic ferralsols (one sample), acric ferralsols (one sample), eutric fluvisols (one sample), and dystric gleysols (one sample). At each sampling point, the soil sample taken at the land top layer was uncovered by vegetation from 0 to 10 cm depth. Approximately 1 kg of topsoil was collected in a soil sample to ensure the representation of the sample. The locations of soil samples were located by a GPS (the Global Positioning System) receiver and added onto the soil map [27] to identify the soil types and the corresponding characteristics of the soil samples.
Most fertile soils, organic matter decrease due to easily eroded, moderate resilience and moderate to low sensitivity.

SMC Determination and Spectral Measurement
The spectral reflectance of these 80 soil samples were measured in a controlled laboratory environment, using an ASD FieldSpec 3 Max Portable spectroradiometer (ASD. Inc. Boulder, CO, USA) with fiber optics and an illumination source (ASD Pro Lamp (14.5 V, 50W), Boulder, CO, USA), which had been corrected for indoor environment. The measurement provided reflectance data every 1 nm from 350 to 2500 nm for each soil sample. After spectral measurement, these samples were determined for the SMC by using the standard method [29], where the moisture content (%) is calculated from the sample weight before and after drying, by Equation (1). Also, the datasets of soil reflectance, R rs (λ), and SMC from these 80 soil samples were curve-fit-analyzed, to construct the model for estimating SMC from the soil surface reflectance: where the SMC units in %, M 2 is the weight of moist soil, and M 1 is the weight of dry soil. Furthermore, 13 soil samples (ferric acrisols soil types) taken from the topsoil layer at 13 points over the Nam Can commune ( Figure 1a) were also measured in-laboratory for reflectance, and the SMC was determined. The datasets of R rs (λ) and SMC from these soil samples were used to examine the performance of the proposed SMC estimation model, derived from the 80 above-mentioned soil Remote Sens. 2019, 11, 716 5 of 18 samples, and to evaluate the atmospheric correction method for the L8 data obtained concurrently on 11 April 2018.

L8 Data Extraction and SMC Mapping
The L8 images used in this study (Table 2) were downloaded from the USGS EarthExplorer website at both levels 1 and 2. L8 level 1 data was geometrically corrected, using ground control points (GCPs) and a digital elevation model (DEM) to correct for terrain and view angles. The Landsat 8 Surface Reflectance data used in this study was directly ordered and downloaded from the USGS Earth Explorer, whereas the surface reflectance, R rs (λ), was retrieved at a 30-m spatial resolution generated from L8 data, through the application of the Landsat Surface Reflectance Code (LaSRC), which uses the Moderate Resolution Imaging Spectroradiometer-Climate Modeling Grid-Aerosol Information as input factors for atmospheric correction [30]. The L8 at level 1 data was radiometrically calibrated into top-of-atmosphere (TOA) reflectance, using the standard method integrated in the ENVI 5.3 routine [31] and then into bottom-of-atmosphere reflectance, R rs (λ), using the dark object subtraction method, DOS [32]. The L8 satellite-retrieved R rs (λ) outputted from these two atmospheric correction methods were then evaluated and selected for mapping SMC over the Central Region of Vietnam. A density slice integrated in ENVI 5.3 was applied to each faction map, using the following intervals: 0 to 10%; 10 to 20%; 20 to 30%; 30 to 40%; 40 to 50%; 50 to 60%; 60 to 70%; 70 to 80%; 80 to 90%; >90%. Only low-density vegetation areas on the scenes were applied to the SMC estimation model, and to this classification method. The process to estimate and map SMC in the Central Region of Vietnam is shown in Figure 2 below, which describes the flowchart of methods used in this work.
Remote Sens. 2019, 11, x FOR PEER REVIEW 5 of 18 SMC was determined. The datasets of Rrs(λ) and SMC from these soil samples were used to examine the performance of the proposed SMC estimation model, derived from the 80 above-mentioned soil samples, and to evaluate the atmospheric correction method for the L8 data obtained concurrently on 11 April 2018.

L8 Data Extraction and SMC Mapping
The L8 images used in this study (Table 2) were downloaded from the USGS EarthExplorer website at both levels 1 and 2. L8 level 1 data was geometrically corrected, using ground control points (GCPs) and a digital elevation model (DEM) to correct for terrain and view angles. The Landsat 8 Surface Reflectance data used in this study was directly ordered and downloaded from the USGS Earth Explorer, whereas the surface reflectance, Rrs(λ), was retrieved at a 30-m spatial resolution generated from L8 data, through the application of the Landsat Surface Reflectance Code (LaSRC), which uses the Moderate Resolution Imaging Spectroradiometer-Climate Modeling Grid-Aerosol Information as input factors for atmospheric correction [30]. The L8 at level 1 data was radiometrically calibrated into top-of-atmosphere (TOA) reflectance, using the standard method integrated in the ENVI 5.3 routine [31] and then into bottom-of-atmosphere reflectance, Rrs(λ), using the dark object subtraction method, DOS [32]. The L8 satellite-retrieved Rrs(λ) outputted from these two atmospheric correction methods were then evaluated and selected for mapping SMC over the Central Region of Vietnam. A density slice integrated in ENVI 5.3 was applied to each faction map, using the following intervals: 0 to 10%; 10 to 20%; 20 to 30%; 30 to 40%; 40 to 50%; 50 to 60%; 60 to 70%; 70 to 80%; 80 to 90%; >90%. Only low-density vegetation areas on the scenes were applied to the SMC estimation model, and to this classification method. The process to estimate and map SMC in the Central Region of Vietnam is shown in Figure 2 below, which describes the flowchart of methods used in this work.

Spectral Response of SMC
By the experiment described in Sections 2.1 and 2.2, the results demonstrated that the SMCs at the saturated state of the soil samples were only slightly different among the soil types, ranging from 44.7% for the rhodic ferralsols soil to 56.7% for the acric ferralsols soil, with an average of 51.9% for all six soil types. Figure 3 presents the spectral change of the soil at SMC = 0%, and at the saturation state corresponding to SMC~51.9%. Accordingly, the mean reflectance, R rs (λ) of soil with 0% SMC was approximately double to three times higher than the soil at the saturation state. Over the spectrum from 400 nm to 2500 nm, all of the soil spectral lines of both these SMC levels clearly showed a broad peak at the regions corresponding to the location of the L8 short-wave infrared band 1 (band 6: B6), and two negative peaks at the regions corresponding to the locations of L8, B5, and B7, whereas the negative peak located within the B7 spectral range is sharper and more isolated from the surrounding regions than the trough located in B5. In both of the soil states, dry and saturation, acric ferralsols has a higher R rs (λ) than the other soil types, which may be a consequence of the higher amount of clayish minerals such as kaolinite consisted in this soil types than the others [28]. The rhodic ferralsols soil samples had the lowest R rs (λ) in the dry state, but it had moderate R rs (λ) at the saturation state, compared to the others. The sample of rhodic ferralsol soil, which was the highest Fe 2 O 3 -containing soil type (approximately 25%) compared to the other soil types [28], was collected in the weathering crust of the eruptive basalt rock that is commonly found in the Central Highlands of Vietnam. In that sense, it can be recognized that it is not only SMC that is affecting the R rs (λ), but also the physical-chemical properties contributing to the change in the R rs (λ) of soil.

Spectral Response of SMC
By the experiment described in Sections 2.1 and 2.2, the results demonstrated that the SMCs at the saturated state of the soil samples were only slightly different among the soil types, ranging from 44.7% for the rhodic ferralsols soil to 56.7% for the acric ferralsols soil, with an average of 51.9% for all six soil types. Figure 3 presents the spectral change of the soil at SMC = 0%, and at the saturation state corresponding to SMC ~ 51.9%. Accordingly, the mean reflectance, Rrs(λ) of soil with 0% SMC was approximately double to three times higher than the soil at the saturation state. Over the spectrum from 400 nm to 2500 nm, all of the soil spectral lines of both these SMC levels clearly showed a broad peak at the regions corresponding to the location of the L8 short-wave infrared band 1 (band 6: B6), and two negative peaks at the regions corresponding to the locations of L8, B5, and B7, whereas the negative peak located within the B7 spectral range is sharper and more isolated from the surrounding regions than the trough located in B5. In both of the soil states, dry and saturation, acric ferralsols has a higher Rrs(λ) than the other soil types, which may be a consequence of the higher amount of clayish minerals such as kaolinite consisted in this soil types than the others [28]. The rhodic ferralsols soil samples had the lowest Rrs(λ) in the dry state, but it had moderate Rrs(λ) at the saturation state, compared to the others. The sample of rhodic ferralsol soil, which was the highest Fe2O3-containing soil type (approximately 25%) compared to the other soil types [28], was collected in the weathering crust of the eruptive basalt rock that is commonly found in the Central Highlands of Vietnam. In that sense, it can be recognized that it is not only SMC that is affecting the Rrs(λ), but also the physical-chemical properties contributing to the change in the Rrs(λ) of soil. The response of soil reflectance, Rrs(λ) to the change of SMC was further examined and explored through charts presented in Figure 4. According to these charts, the higher the SMC, the lower the Rrs(λ) obtained. This trend was demonstrated in all six soil samples corresponding to the six investigated soil types. In the spectral lines of acric ferralsols (Figure 4a  The response of soil reflectance, R rs (λ) to the change of SMC was further examined and explored through charts presented in Figure 4. According to these charts, the higher the SMC, the lower the R rs (λ) obtained. This trend was demonstrated in all six soil samples corresponding to the six investigated soil types. In the spectral lines of acric ferralsols (Figure 4a) and dystric gleysols (Figure 4b), the soil samples rose gradually with the decrease of SMC, and in four other soil-type samples, the spectral lines rose suddenly at various SMC, depending on the soil type, i.e., at SMC = 8.7% of the rhothic ferrasols sample (Figure 4c  In all of the soil type samples, the spectral trough at the B7 location was clearly presented, and it increased inversely with SMC. Similarly, the trough within B5 was not clearly presented in several cases, i.e., in dystric gleysols (Figure 4b) and eutric fluvisols (Figure 4d) samples, but the inverse trend between Rrs(λ) and SMC was also shown in all cases. The spectral peak located with the same location of B6 was also presented clearly in all cases, but the inverse trend between Rrs(λ) and SMC was apparent in several cases, particularly near the saturation and saturation states of dystric gleysols, rhothic ferrasols, and ferric acrisols soil samples (Figure 4b,c,f). The strong correlations of detected SMC with B5, B6, and B7 in each soil sample from this study (Table 3) can help us to understand why In all of the soil type samples, the spectral trough at the B7 location was clearly presented, and it increased inversely with SMC. Similarly, the trough within B5 was not clearly presented in several cases, i.e., in dystric gleysols (Figure 4b) and eutric fluvisols (Figure 4d) samples, but the inverse trend between R rs (λ) and SMC was also shown in all cases. The spectral peak located with the same location of B6 was also presented clearly in all cases, but the inverse trend between R rs (λ) and SMC was apparent in several cases, particularly near the saturation and saturation states of dystric gleysols, rhothic ferrasols, and ferric acrisols soil samples (Figure 4b,c,f). The strong correlations of detected SMC with B5, B6, and B7 in each soil sample from this study ( these bands or spectral regions have been used for estimating SMC from satellite data in the preceding works [12]. To identify the best band and/or band combination for estimating SMC using L8 images, the curve-fit regressions of SMC and the strongest correlated band (B7), and the proposed band ratio and band combinations in preceding studies [9,12,33] were conducted, and results were presented in Table 4, using the total data of 80 soil sub-samples. Accordingly, SMC has the highest correlation, with the ratio of L8 B5 versus B7, B5/B7, (R 2 = 0.73), and the smallest error (RMSE = 8.85), indicating the high potential of this band ratio for estimating SMC, using the L8 image. In most regressions, the logarithm function provides the lowest error compared to other functions, such as exponential, linear, and power. Two commonly used indexes for SMC monitoring purposes based on field radiometric data, the Normalized Multi-band Drought Index (NMDI) [34], and the Normalized Difference Soil Moisture Index (NSMI) [35], were also evaluated by using the L8 spectral bands in this work. NSMI, which used the difference of the reflectance at 1800 nm and 2119 nm, corresponded to the difference of the retrieved reflectance of L8, B6, and B7, (B6 − B7)/(B6 + B7), had a moderate correlation with SMC (R 2 = 0.51). This was similar, to the correlation of NMDI, which used the difference of the reflectance at NIR (860 nm), corresponding to L8 B5, and the differences of the reflectance at 1640 nm and 2130 nm, corresponding to the differences of B6 and B7, [B5 − (B6 − B7)]/[B5 + (B6 − B7)], and SMC (R 2 = 0.34). This result demonstrated the limitation of these two indexes when applying L8 data for quantifying SMC in the tropical region. To further investigate the cross-relationship between the L8 SWIR band, B5/B7, and NSMI and SMC, scatterplots between SMC and these three variables were created, and they are shown in Figure 3. Accordingly, the use of the SWIR band (B7) was only seen in several cases, particularly at a range of SMC that was lower than 30% (Figure 5a). NSMI was proposed for estimating SMC by using surface reflectance at two wavelengths, 1800 nm and 2119 nm. The difference between the L8 B6 center wavelength (1610 nm) and 1800 nm may be a factor that reduces the correlation of NSMI with SMC ( Figure 5b). The fit of the observations with the logarithm curve demonstrated the relationship between SMC and B5/B7 (Figure 5c), providing a scientific base for estimating SMC from the L8 images by the following Equation: SMC = 8.14 + 40.29 × log(B5/B7) (2) where SMC units in %, and B5/B7 is the ratio of reflectance corresponding to L8, B5, and B7.
Remote Sens. 2019, 11, x FOR PEER REVIEW 9 of 18 where SMC units in %, and B5/B7 is the ratio of reflectance corresponding to L8, B5, and B7. ; c) the L8 spectral ratio of the NIR band (B5) versus the SWIR-2 band, B5/B7, using data measured from 80 soil sub-samples. The best fit of the exponential curve to the relationship of SMC and B5/B7, which is (c) indicates the appropriateness of this band ratio for estimating SMC by using L8 data. Figure 6A shows the comparison of SMC estimated from Rrs(λ), corresponding to L8 B5/B7, using Equation (2), hereafter referred to as L8RSM, and the in situ SMCs of 13 soil samples taken in the Nam Can commune, Nghe An Province, in April 2018. The SMC of these 13 soil samples ranged from 10.7% to 33.1%, averaging at 26%, indicating the SMC range of the topsoils over Nam Can commune under normal weather conditions in April. The spectral features of the 13 soil samples present in Figure 6B again confirmed the inverse trend of SMC and Rrs(λ). The small difference between these two datasets (RMSE = 3.21, corresponding to approximately 12% of the mean of in situ SMC) confirmed the high accuracy of L8RSM for estimating SMC ( Figure 6A). In order to apply L8RSM to the estimation of SMC from the L8 satellite data, the image processing method, i.e., the atmospheric correction method, should be focused upon, to ensure the accuracy of the derived reflectance.

Mapping SMC Using L8 Images
Mapping SMC in space and time provides useful information for agriculture management, particularly in areas that are frequently affected by drought, such as the Central Highlands of Vietnam. Passive satellite remote sensing uses signals from the emission and reflection of electromagnetic energy by the soil surface, to estimate SMC. The effects of atmosphere, land-cover, surface roughness, and soil texture on these signals are still challenges for estimating and mapping SMC from satellite data [36]. The use of the spectral band ratio, as in this work, has been proven to  B6+B7)); c) the L8 spectral ratio of the NIR band (B5) versus the SWIR-2 band, B5/B7, using data measured from 80 soil sub-samples. The best fit of the exponential curve to the relationship of SMC and B5/B7, which is (c) indicates the appropriateness of this band ratio for estimating SMC by using L8 data. Figure 6A shows the comparison of SMC estimated from R rs (λ), corresponding to L8 B5/B7, using Equation (2), hereafter referred to as L8R SM , and the in situ SMCs of 13 soil samples taken in the Nam Can commune, Nghe An Province, in April 2018. The SMC of these 13 soil samples ranged from 10.7% to 33.1%, averaging at 26%, indicating the SMC range of the topsoils over Nam Can commune under normal weather conditions in April. The spectral features of the 13 soil samples present in Figure 6B again confirmed the inverse trend of SMC and R rs (λ). The small difference between these two datasets (RMSE = 3.21, corresponding to approximately 12% of the mean of in situ SMC) confirmed the high accuracy of L8R SM for estimating SMC ( Figure 6A). In order to apply L8R SM to the estimation of SMC from the L8 satellite data, the image processing method, i.e., the atmospheric correction method, should be focused upon, to ensure the accuracy of the derived reflectance. where SMC units in %, and B5/B7 is the ratio of reflectance corresponding to L8, B5, and B7. ; c) the L8 spectral ratio of the NIR band (B5) versus the SWIR-2 band, B5/B7, using data measured from 80 soil sub-samples. The best fit of the exponential curve to the relationship of SMC and B5/B7, which is (c) indicates the appropriateness of this band ratio for estimating SMC by using L8 data. Figure 6A shows the comparison of SMC estimated from Rrs(λ), corresponding to L8 B5/B7, using Equation (2), hereafter referred to as L8RSM, and the in situ SMCs of 13 soil samples taken in the Nam Can commune, Nghe An Province, in April 2018. The SMC of these 13 soil samples ranged from 10.7% to 33.1%, averaging at 26%, indicating the SMC range of the topsoils over Nam Can commune under normal weather conditions in April. The spectral features of the 13 soil samples present in Figure 6B again confirmed the inverse trend of SMC and Rrs(λ). The small difference between these two datasets (RMSE = 3.21, corresponding to approximately 12% of the mean of in situ SMC) confirmed the high accuracy of L8RSM for estimating SMC ( Figure 6A). In order to apply L8RSM to the estimation of SMC from the L8 satellite data, the image processing method, i.e., the atmospheric correction method, should be focused upon, to ensure the accuracy of the derived reflectance.

Mapping SMC Using L8 Images
Mapping SMC in space and time provides useful information for agriculture management, particularly in areas that are frequently affected by drought, such as the Central Highlands of Vietnam. Passive satellite remote sensing uses signals from the emission and reflection of electromagnetic energy by the soil surface, to estimate SMC. The effects of atmosphere, land-cover, surface roughness, and soil texture on these signals are still challenges for estimating and mapping

Mapping SMC Using L8 Images
Mapping SMC in space and time provides useful information for agriculture management, particularly in areas that are frequently affected by drought, such as the Central Highlands of Vietnam. Passive satellite remote sensing uses signals from the emission and reflection of electromagnetic energy by the soil surface, to estimate SMC. The effects of atmosphere, land-cover, surface roughness, and soil texture on these signals are still challenges for estimating and mapping SMC from satellite data [36]. The use of the spectral band ratio, as in this work, has been proven to reduce the effects of not only the atmosphere [37], but also the roughness of the topographic surface [38] on the satellite-derived reflectance. Figure 7 presents an improved accuracy of the spectral band ratio compared to the single spectral bands after processing by two atmospheric correction methods: DOS and LaSRC (Figure 7c), using two datasets of in situ R rs (λ) measured at seven soil sampling points in the Nam Can commune, and the retrieved R rs (λ) of seven corresponding free-cloud pixels of the L8 image, acquired concurrently over the commune area on 11 April 2018. Accordingly, L8-retrieved R rs (λ) from single bands, band 5 ( Figure 7a) and band 7 (Figure 7b) were nearly three times higher than in situ R rs (λ), with the RMSE being 0.02 and 0.03, respectively. The L8-retrieved spectral ratio of band 5 versus band 7 (B5/B7) showed a consistent fit to the range of in situ data ( Figure 7c) with a small RMSE (0.20, corresponding to 12% of the mean of in-situ dataset). This implicates that the L8-derived single band reflectance for SMC estimation may be limited due to large uncertainties in the atmospheric correction methods, but the band ratio model is less subject to errors in the atmospheric correction. Therefore, L8R SM is suitable for use in SMC estimation and mapping using the L8 images, which are atmospheric corrected by DOS method.
Remote Sens. 2019, 11, x FOR PEER REVIEW 10 of 18 [38] on the satellite-derived reflectance. Figure 7 presents an improved accuracy of the spectral band ratio compared to the single spectral bands after processing by two atmospheric correction methods: DOS and LaSRC (Figure 7c), using two datasets of in situ Rrs(λ) measured at seven soil sampling points in the Nam Can commune, and the retrieved Rrs(λ) of seven corresponding free-cloud pixels of the L8 image, acquired concurrently over the commune area on 11 April 2018. Accordingly, L8retrieved Rrs(λ) from single bands, band 5 ( Figure 7a) and band 7 ( Figure 7b) were nearly three times higher than in situ Rrs(λ), with the RMSE being 0.02 and 0.03, respectively. The L8-retrieved spectral ratio of band 5 versus band 7 (B5/B7) showed a consistent fit to the range of in situ data ( Figure 7c) with a small RMSE (0.20, corresponding to 12% of the mean of in-situ dataset). This implicates that the L8-derived single band reflectance for SMC estimation may be limited due to large uncertainties in the atmospheric correction methods, but the band ratio model is less subject to errors in the atmospheric correction. Therefore, L8RSM is suitable for use in SMC estimation and mapping using the L8 images, which are atmospheric corrected by DOS method. The estimation of SMC at seven sampling points using the L8 scene, which is taken over Nam Can commune and atmospheric corrected by DOS, and L8RSM was presented in Figure 8. Figure 8a shows the map of sampling point overlaid the estimated SMC result. Figure 8b is the scatterplot showing the validated result of L8RSM for estimating SMC using in situ SMC at seven sampling points in Nam Can commune and estimated SMC from the L8 image presented in Figure 8a. Accordingly, estimated SMC has a highly correlation with in situ SMC (R 2 = 0.68) and a reasonable error (RMSE = 5.80) demonstrating the high possibility of L8RSM and studied methods for quantifying and mapping SMC in a tropical region using L8 images. Figure 7. Comparisons of in situ R rs (λ) and retrieved R rs (λ) corresponding to (a) the L8 NIR band (band 5: B5), (b) the L8 SWIR-2 band (band 7: B7), and (c) the spectral ratio corresponding to B5/B7, obtained via two atmospheric correction methods: DOS and LaSRC. While the retrieved R rs (λ) of B5 and B7 indicates significant biases with respect to in situ R rs (λ), the retrieved B5/B7 (c) demonstrates a smaller dependency of the data on atmospheric correction, and it is consistent with in situ data, and therefore suitable for use in SMC estimation.
The estimation of SMC at seven sampling points using the L8 scene, which is taken over Nam Can commune and atmospheric corrected by DOS, and L8R SM was presented in Figure 8. Figure 8a shows the map of sampling point overlaid the estimated SMC result. Figure 8b is the scatterplot showing the validated result of L8R SM for estimating SMC using in situ SMC at seven sampling points in Nam Can commune and estimated SMC from the L8 image presented in Figure 8a. Accordingly, estimated SMC has a highly correlation with in situ SMC (R 2 = 0.68) and a reasonable error (RMSE = 5.80) demonstrating the high possibility of L8R SM and studied methods for quantifying and mapping SMC in a tropical region using L8 images. Figure 9 presents the maps of SMC, as estimated from L8 images ( Table 2) acquired over the northern part of the Central Highlands of Vietnam in the late dry season of 2015 and 2016, March 2015 and March 2016, using Equation (2). As mentioned, to subtract the effects of vegetation, only the SMCs of uncovered soils and/or soils covered by thinly vegetated areas where NDVI ranges from 0 to 0.4, corresponding to cropland and non-vegetated lands, such as bare soil and built-up landcover types [39], were estimated and mapped from the L8 images. Areas with dense vegetation (NDVI > 0.4) appeared in these two maps as "not observed" areas, and they are filled in green. The drought-affected districts reported by local provincial governments [40][41][42] and non-governmental organizations, NGOs [43,44], present in these two maps are shown as oblique-lined areas.
via two atmospheric correction methods: DOS and LaSRC. While the retrieved Rrs() of B5 and B7 indicates significant biases with respect to in situ Rrs(), the retrieved B5/B7 (c) demonstrates a smaller dependency of the data on atmospheric correction, and it is consistent with in situ data, and therefore suitable for use in SMC estimation.
The estimation of SMC at seven sampling points using the L8 scene, which is taken over Nam Can commune and atmospheric corrected by DOS, and L8RSM was presented in Figure 8. Figure 8a shows the map of sampling point overlaid the estimated SMC result. Figure 8b is the scatterplot showing the validated result of L8RSM for estimating SMC using in situ SMC at seven sampling points in Nam Can commune and estimated SMC from the L8 image presented in Figure 8a. Accordingly, estimated SMC has a highly correlation with in situ SMC (R 2 = 0.68) and a reasonable error (RMSE = 5.80) demonstrating the high possibility of L8RSM and studied methods for quantifying and mapping SMC in a tropical region using L8 images.   (2). As mentioned, to subtract the effects of vegetation, only the SMCs of uncovered soils and/or soils covered by thinly vegetated areas where NDVI ranges from 0 to 0.4, corresponding to cropland and non-vegetated lands, such as bare soil and built-up landcover types [39], were estimated and mapped from the L8 images. Areas with dense vegetation (NDVI > 0.4) appeared in these two maps as "not observed" areas, and they are filled in green. The droughtaffected districts reported by local provincial governments [40][41][42] and non-governmental organizations, NGOs [43,44], present in these two maps are shown as oblique-lined areas. The resultant SMC distribution maps in March 2015 and March 2016 showed that almost all soils in the observed areas were characterized by a low SMC level (SMC < 40%), evidence of the occurrence of severe droughts in the area; the drought disaster in 2016 was more severe than the one in 2015, which was manifested through the expansion of the SMC < 40% areas in the 2016 map (Figure 9a), compared to the 2015 map (Figure 9b). From the maps, soils in the northern part of the Central Highlands were under water stress for crop growth, particularly for the growth of popular local farming plants, such as coffee, rice, maize, peanut, soybeans, and greens. According to the Vietnamese Standard published by the Vietnamese Ministry of Agriculture and Rural Development [45], the optimal SMC for these plants growing is the range from 60% to 85%, and at a SMC lower than 40%, all of these plants were shriveled up ( Figure 10). The SMC distribution maps provided evidence, and they were conformable to local district reports on the effect of drought disasters in these areas for agricultural activities, particularly in areas where the yields were reported as a deadloss [41,42]. The resultant SMC distribution maps in March 2015 and March 2016 showed that almost all soils in the observed areas were characterized by a low SMC level (SMC < 40%), evidence of the occurrence of severe droughts in the area; the drought disaster in 2016 was more severe than the one in 2015, which was manifested through the expansion of the SMC < 40% areas in the 2016 map (Figure 9a), compared to the 2015 map (Figure 9b). From the maps, soils in the northern part of the Central Highlands were under water stress for crop growth, particularly for the growth of popular local farming plants, such as coffee, rice, maize, peanut, soybeans, and greens. According to the Vietnamese Standard published by the Vietnamese Ministry of Agriculture and Rural Development [45], the optimal SMC for these plants growing is the range from 60% to 85%, and at a SMC lower than 40%, all of these plants were shriveled up ( Figure 10). The SMC distribution maps provided evidence, and they were conformable to local district reports on the effect of drought disasters in these areas for agricultural activities, particularly in areas where the yields were reported as a dead-loss [41,42].

Discussion
The effect of soil moisture on soil spectral feature has been long addressed in preceding studies with two highlighted features: more SMC lowers reflectance but does not change the shape of soil spectral line; clear presence of water absorption at 1400 nm, 1900 nm and 2200 nm [48,49]. The use of reflectance at 2.2 µm to identify clay minerals is constrained by the influence of SMC [50]. The L8 SWIR2 band (band 7) data within 2110-2290 nm has been designed to estimate moisture content of soil and vegetation in thin cloudy condition [51]. The reflectance corresponding to the L8 NIR band has been often used to discriminate between the vegetation covered areas and bare soil areas [52] or to infer plant water stress and subsequent reduction of plant productivities [53]. As mentioned, water strongly absorbs in the middle-infrared region (MIR) and is a major factor controlling the spectral properties of surface materials, thus use of reflectance within the MIR region (the L8 bands 6 and 7) is suggested to estimate the water content in surface materials. Many studies show the appropriateness of using the reflectance within 1550-1650 nm (corresponding the L8 band 6) for estimating the leaf water content [53,54], and in which the reflectance near 820 nm (corresponding L8 NIR band) has been used as a reference factor that helps clearly reflect the change in reflectance within 1550-1650 nm, accompanied by change in leaf water content. Figure 4 shows clearly the effect of SMC to the depth of the negative peak of reflectance at L8 band 7, but no peak related to a strong water absorption at the L8 band 6 (1570-1650 nm) was observed. Therefore, the use of spectral band ratio of band 5 versus band 7 (NIR/SWIR2) reflects more sensitively to the change of SMC, which helps more accurately estimate SMC from reflectance data.
Three drought types are commonly recognized; i.e., meteorological, hydrological, and agricultural drought [55], of which agricultural drought is directly affected by the local economy and by social development. Agricultural drought is defined as the reduction of soil moisture availability to plants at the SMC level, as crop growth is adversely affected, and therefore, crop yield, and hence, agricultural production, is decreased [56]. According to this definition, agricultural drought can be proactively predicted in a reasonable manner by monitoring SMC [57]. The traditional in situ-based drought monitoring method (including ground measurements of hydro-climatic variables such as precipitation, temperature, relative humidity, evaporation and SMC, etc.) provides accurate data at point locations, but not the spatial dynamics over a large area. The requirement for spatial impact of drought leads to the wide use of remote sensing-based drought indices for agriculture management. Several agricultural drought indices that use remote sensing, such as NDDI [26], the vegetation health index (VHI) [58], the Shortwave Infrared Water Stress Index (SIWSI) [59], the Normalized Multiband Drought Index (NMDI) [34], and the Vegetation Temperature Condition Index (VTCI) [60], have been proposed. Among these indices, NDDI, as proposed by Gao [26], has been widely used, because it combines vegetation health and surface wetness conditions, and does not require ground-based data [61][62][63][64]. Figure 11 showcases the relationship between the estimated SMC and NDDI, retrieved from L8 images that were taken over the northern part of the Central Highland of Vietnam in March 2015.

Discussion
The effect of soil moisture on soil spectral feature has been long addressed in preceding studies with two highlighted features: more SMC lowers reflectance but does not change the shape of soil spectral line; clear presence of water absorption at 1400 nm, 1900 nm and 2200 nm [48,49]. The use of reflectance at 2.2 µm to identify clay minerals is constrained by the influence of SMC [50]. The L8 SWIR2 band (band 7) data within 2110-2290 nm has been designed to estimate moisture content of soil and vegetation in thin cloudy condition [51]. The reflectance corresponding to the L8 NIR band has been often used to discriminate between the vegetation covered areas and bare soil areas [52] or to infer plant water stress and subsequent reduction of plant productivities [53]. As mentioned, water strongly absorbs in the middle-infrared region (MIR) and is a major factor controlling the spectral properties of surface materials, thus use of reflectance within the MIR region (the L8 bands 6 and 7) is suggested to estimate the water content in surface materials. Many studies show the appropriateness of using the reflectance within 1550-1650 nm (corresponding the L8 band 6) for estimating the leaf water content [53,54], and in which the reflectance near 820 nm (corresponding L8 NIR band) has been used as a reference factor that helps clearly reflect the change in reflectance within 1550-1650 nm, accompanied by change in leaf water content. Figure 4 shows clearly the effect of SMC to the depth of the negative peak of reflectance at L8 band 7, but no peak related to a strong water absorption at the L8 band 6 (1570-1650 nm) was observed. Therefore, the use of spectral band ratio of band 5 versus band 7 (NIR/SWIR2) reflects more sensitively to the change of SMC, which helps more accurately estimate SMC from reflectance data.
Three drought types are commonly recognized; i.e., meteorological, hydrological, and agricultural drought [55], of which agricultural drought is directly affected by the local economy and by social development. Agricultural drought is defined as the reduction of soil moisture availability to plants at the SMC level, as crop growth is adversely affected, and therefore, crop yield, and hence, agricultural production, is decreased [56]. According to this definition, agricultural drought can be proactively predicted in a reasonable manner by monitoring SMC [57]. The traditional in situ-based drought monitoring method (including ground measurements of hydro-climatic variables such as precipitation, temperature, relative humidity, evaporation and SMC, etc.) provides accurate data at point locations, but not the spatial dynamics over a large area. The requirement for spatial impact of drought leads to the wide use of remote sensing-based drought indices for agriculture management. Several agricultural drought indices that use remote sensing, such as NDDI [26], the vegetation health index (VHI) [58], the Shortwave Infrared Water Stress Index (SIWSI) [59], the Normalized Multiband Drought Index (NMDI) [34], and the Vegetation Temperature Condition Index (VTCI) [60], have been proposed.
Among these indices, NDDI, as proposed by Gao [26], has been widely used, because it combines vegetation health and surface wetness conditions, and does not require ground-based data [61][62][63][64]. Figure 11 showcases the relationship between the estimated SMC and NDDI, retrieved from L8 images that were taken over the northern part of the Central Highland of Vietnam in March 2015. Accordingly, the "abnormal dry" level predicted by NDDI approximately corresponded to the SMC range of 30% to 40%, which is a state of water stress in growing plants. The weak point of the NDDI is the low applicability of this index for monitoring of short-term drought [65], due to the lag time between the surface soil and vegetation in the dry state, when a region has been affected by drought. Therefore, direct monitoring and mapping of the estimated SMC from L8 data helps with early forecasting of drought disaster, and management of irrigation activities to mitigate bad impacts from drought. Further investigation of the relationship of SMC and NDDI shows that the estimated SMC is negatively correlated with the NDDI value, particularly in area cover, mostly by NDVI < 0.4, corresponding to medium-to thin vegetation-covered areas (R 2 = 0.95; Figure 11c). Along with the increase of areas covered by thick vegetation (NDVI > 0.4), the correlation of SMC and NDDI tends towards a decrease: R 2 = 0.78 over the area covered by less than 20% thick vegetation (Figure 11a), and R 2 = 0.03 over the area that is covered by more than 40% thick vegetation (Figure 11b). This tendency indicates a lower accuracy of L8R SM in estimating SMC from a pixel, as thick vegetation covers more than a 20% area of the pixel. Figure 11 also indicates that when the estimated SMC is lower than 40% in almost medium-to-thin levels of vegetation cover, the region is affected by severe drought, and furthermore, the richness of vegetation is a vital factor for water retention in the soil.
Remote Sens. 2019, 11, x FOR PEER REVIEW 13 of 18 Accordingly, the "abnormal dry" level predicted by NDDI approximately corresponded to the SMC range of 30% to 40%, which is a state of water stress in growing plants. The weak point of the NDDI is the low applicability of this index for monitoring of short-term drought [65], due to the lag time between the surface soil and vegetation in the dry state, when a region has been affected by drought. Therefore, direct monitoring and mapping of the estimated SMC from L8 data helps with early forecasting of drought disaster, and management of irrigation activities to mitigate bad impacts from drought. Further investigation of the relationship of SMC and NDDI shows that the estimated SMC is negatively correlated with the NDDI value, particularly in area cover, mostly by NDVI < 0.4, corresponding to medium-to thin vegetation-covered areas (R 2 = 0.95; Figure 11c). Along with the increase of areas covered by thick vegetation (NDVI > 0.4), the correlation of SMC and NDDI tends towards a decrease: R 2 = 0.78 over the area covered by less than 20% thick vegetation (Figure 11a), and R 2 = 0.03 over the area that is covered by more than 40% thick vegetation (Figure 11b). This tendency indicates a lower accuracy of L8RSM in estimating SMC from a pixel, as thick vegetation covers more than a 20% area of the pixel. Figure 11 also indicates that when the estimated SMC is lower than 40% in almost medium-to-thin levels of vegetation cover, the region is affected by severe drought, and furthermore, the richness of vegetation is a vital factor for water retention in the soil. Figure 11. Map of the retrieved NDDI-overlaid estimated SMC from L8 images acquired over the northern part of the Central Highlands of Vietnam in March 2015, which demonstrates the spatial relationship between these two variables: (A) within the area, 20-30% is dense vegetation coverage; (B) within the area, more than 40% is dense vegetation coverage; (C) within the area with lower than 20% dense vegetation coverage.
As mentioned, agricultural drought is closely connected to soil moisture and crop water deficit, and therefore SMC is the most direct and important indicator for assessing and monitoring an agricultural drought event [66]. Remote sensing-based drought indices are mostly developed by using vegetation health conditions such as NDVI, VHI, TVDI, VI, NDDI, NDTI, the vegetation condition index, VCI [58], etc. However, vegetation-based drought indices are reported to fail as early warnings of agricultural drought [65,67], because the time that needs to elapse before leaf greenness is destroyed is much longer than the elapsed time for a decrease in the SMC to a level at which plants are under water stress (SMC < 40%) [68]. The L8RSM proposed in this work helps to directly and frequently estimate SMC, in order to identify the areas in which plants are under water stress, for early warning and proactive planning, in order to respond and mitigate agricultural drought.
L8RSM was also exploited to predict agricultural drought in previous studies through the use of several indices; i.e., Moisture Stress Index [54], Normalized Difference Infrared Index [69], and Figure 11. Map of the retrieved NDDI-overlaid estimated SMC from L8 images acquired over the northern part of the Central Highlands of Vietnam in March 2015, which demonstrates the spatial relationship between these two variables: (A) within the area, 20-30% is dense vegetation coverage; (B) within the area, more than 40% is dense vegetation coverage; (C) within the area with lower than 20% dense vegetation coverage.
As mentioned, agricultural drought is closely connected to soil moisture and crop water deficit, and therefore SMC is the most direct and important indicator for assessing and monitoring an agricultural drought event [66]. Remote sensing-based drought indices are mostly developed by using vegetation health conditions such as NDVI, VHI, TVDI, VI, NDDI, NDTI, the vegetation condition index, VCI [58], etc. However, vegetation-based drought indices are reported to fail as early warnings of agricultural drought [65,67], because the time that needs to elapse before leaf greenness is destroyed is much longer than the elapsed time for a decrease in the SMC to a level at which plants are under water stress (SMC < 40%) [68]. The L8R SM proposed in this work helps to directly and frequently estimate SMC, in order to identify the areas in which plants are under water stress, for early warning and proactive planning, in order to respond and mitigate agricultural drought.
L8R SM was also exploited to predict agricultural drought in previous studies through the use of several indices; i.e., Moisture Stress Index [54], Normalized Difference Infrared Index [69], and Surface Water Index [70]; these were proposed for the observation of leaf moisture in densely vegetated areas. In a sense, this spectral ratio was demonstrated to be suitable for estimating the moisture content of the top-surface materials by using satellite data, despite differences in the soil components of various soil types (six soil types as mentioned in this study) or the canopy values of the leaves. More investigations and analyses on the hyperspectral features of the leaves at various moisture contents, and/or hyperspectral features of SMC in other top-soil types should be carried out, to evaluate the performance of this spectral ratio for estimating the moisture content of the land surfaces, and for prediction of drought effects.

Conclusions
Literature review indicates the rare use of optical satellite data, such as Landsat, for directly estimating SMC in the topsoil, instead of indirectly estimating the moisture of other land cover types (i.e., the leaf moisture content, the moisture of dense vegetation cover), by proposing several vegetation and/or water indices. An addition to the challenge comes from the diversity of the landcover types, and knowledge of the principle responses of the reflectance spectra to various SMC is still limited. This study demonstrates the appropriateness of the spectral ratio of the L8 band NIR-to-SWIR2 for estimating the SMC, L8R SM , through analyses of the spectral response corresponding L8 band data, with various SMC over various soil types, which are obtained from in-laboratory measurements of R rs (λ) and SMC (R 2 = 0.73 and RMSE~12%). The hyperspectral features of a total of 103 soil samples taken from the Central Region of Vietnam clearly show the adverse effects of SMC on R rs (λ), particularly within the NIR-SWIR region. The appropriateness of the spectral ratio of the L8 band NIR-to-SWIR2 for estimating the SMC observed in this study provides a physical base and explains why this ratio has been used in previous studies for the estimation of surface moisture content when using remote optic data. The resultant maps of the spatial distribution of the estimated SMCs from L8 images acquired in March 2015 and March 2016 over the northern part of the Central Highlands of Vietnam were conformable to the practical droughts occurring in the region at the same time. The spatial relationship between the estimated SMCs and the retrieved NDDI, a widely used remote sensing index for agricultural drought monitoring, from L8 images in March 2016, presents a strong correlation for these two variables, when they are obtained within areas with less than 20% dense vegetation coverage. Directly estimating SMC from the L8 images is better for the management of irrigation during the drought occurrence period, compared to NDDI, because of the large time-lag between the decrease of SMC and decrease of leaf greenness. With a view towards effectively monitoring SMC by using optical satellite data, more investigations into the response of R rs (λ) to different SMC over various soil types or landcovers should be carried out in the near future. Other optical data, such as Sentinel 2A and 2B, should also be exploited to shorten the observation period, in order to perform better drought disaster mitigation.