Mapping Soil Moisture as an Indicator of Wildfire Risk Using Landsat 8 Images in Sri Lanna National Park , Northern Thailand

Severely dry climate plays an important role in the occurrence of wildfires in Thailand. Soil water deficits increase dry conditions, resulting in more intense and longer burning wildfires. The temperature vegetation dryness index (TVDI) and the normalized difference drought index (NDDI) were used to estimate soil moisture during the dry season to explore its use for wildfire risk assessment. The results reveal that the normalized difference wet index (NDWI) and land surface temperature (LST) can be used for TVDI calculation. Scatter plots of both NDWI/LST and the normalized difference vegetation index (NDVI)/LST exhibit the triangular shape typical for the theoretical TVDI. However, the NDWI is more significantly correlated to LST than the NDVI. Linear regression analysis, carried out to extract the maximum and minimum LSTs (LSTmax, LSTmin), indicate that LSTmax and LSTmin delineated by the NDWI better fulfill the collinearity requirement than those defined by the NDVI. Accordingly, the NDWI-LST relationship is better suited to calculate the TVDI. This modified index, called TVDINDWI-LST, was applied together with the NDDI to establish a regression model for soil moisture estimates. The soil moisture model fulfills statistical requirements by achieving 76.65% consistency with the actual soil moisture and estimated soil moisture generated by our model. The relationship between soil moisture estimated from our model and leaf fuel moisture indicates that soil moisture can be used as a complementary dataset to assess wildfire risk, because soil moisture and fuel moisture content (FMC) show the same or similar behavior under dry conditions.


Introduction
Severely dry climate plays an important role in the occurrence of wildfires.In Thailand, forest wildfires are particularly prevalent during the dry season and are especially damaging because of forest loss and degradation.During the dry season, the number of wildfires in Thai conserved forest areas were 4207, 4982, and 6685 in 2014, 2015, and 2016, respectively (Forest Fire Control Division, 2016).These numbers indicate that the number of wildfires appears to be increasing because Thailand has been experiencing longer dry seasons and under dry conditions, wildfires can ignite easily, as fuel sources are readily available.Fuel availability, which drives wildfire occurrences and directly affects wildfire behavior, depends on fuel characteristics, which are fuel load (influencing fire intensity) and fuel moisture content (influencing both fire ignition and spread).It appears that recurring dry seasons foster fuel availability and reduce fuel moisture content, resulting in potentially more damaging high-intensity fires, which may spread rapidly during extremely dry conditions.Soil moisture, defined as the volumetric water content of soil (Eller & Denoth, 1996), is an important indicator of dry conditions and is linked to wildfire occurrence.The reduction of water in soil increases dry conditions, resulting in more intense and longer burning fires (Kozlowski & Pallardy, 2002;Chmura et al., 2011).Previous studies pointed out that soil moisture affects wildfire occurrence.For example, Krueger et al. (2015) showed that large growing-season wildfires occurred exclusively under conditions of low soil moisture.Yebra et al. (2013) suggested that improving wildfire assessments involves using soil moisture as a representative for fuel moisture, which is a key factor for ignition and spread of wildfires.Therefore, surface measurements of soil moisture may provide opportunities for improving estimates of fuel moisture (Qi, Dennison, Spencer, & Riano, 2012), because both are physically linked through soil-plant interactions (Hillel, 1998).
Remote sensing techniques have been extensively used for the analysis of soil moisture, and have provided alternative tools for obtaining rapid estimates of soil moisture on large spatial scales (Goward, Xue, & Czajkowski, 2002;Sandholt, Rasmussen, & Andersen, 2002;Ishimura, Shimizu, Rahimzadeh, & Omasa, 2011).Vegetation indices (VIs), which are mathematical combinations of different spectral bands from satellite remotely sensed data, have been utilized to estimate soil moisture (Z.Gao, W. Gao, & Chang, 2011;Chen et al., 2015).The normalized difference vegetation index (NDVI) is the normalized reflectance difference between the near-infrared (NIR) and visible red (R) bands (Rouse, Haas, Deering, Schell, & Harlan, 1974;Tucker, 1979), which measures changes in chlorophyll content.As a result, it is considered a function of vegetation strength, which changes as vegetation interacts with soil moisture.The normalized difference water index (NDWI) is a more recent satellite-derived index from the NIR and short-wave infrared (SWIR) channels that reflects changes in both water content and spongy mesophyll in vegetation canopies (Gao, 1996).This index has been employed for the determination of vegetation water content and stress (Ceccato, Gobron, Flasse, Pinty, & Tarantola, 2002), and is therefore expected to be linked to soil moisture due to its impact on vegetation water stress.Moreover, land surface temperature (LST) can rise rapidly with water stress (Goetz, 1997), which is directly related to soil moisture.Accordingly, LST is also widely used as a soil moisture indicator (Carlson, 2007).
The relationship between VI and LST has been investigated to evaluate evapotranspiration rates.The VI-LST relationship normally shows a negative correlation, resulting in triangular-shaped VI-LST plots at different spatial scales (Nemani, Pierce, Running, & Goward, 1993;Goetz, 1997).Based on the VI-LST correlation, the temperature vegetation dryness index (TVDI), computed from the NDVI-LST relationship has become a widely used dryness index to estimate surface soil moisture (Sandholt, Rasmussen, & Andersen, 2002;Mallick, Bhattacharya, & Patel, 2009;Patel, Anapashsha, Kumar, Saha, & Dadhwal, 2009).For example, Wang, Qu, Zhang, Hao, and Dasgupta (2007) applied NDVI-LST produced from moderate resolution imaging spectroradiometer (MODIS) data to investigate the correlation with soil moisture determined by field measurements.The results revealed that NDVI-LST is strongly correlated with soil moisture and can be used to generate soil moisture estimates.Chen et al. (2015) used the TVDI (NDVI-LST) derived from Landsat-5 TM data to estimate soil moisture and found that the TVDI can reflect the soil moisture status under different tree species.In this study, we propose a new application of the NDWI-LST relationship, which could enhance the efficiency of the TVDI calculation.Additionally, the normalized difference drought index (NDDI), which combines information about both greenness and water obtained from the NDVI and the NDWI (Gu, Brown, Verdin, & Wardlow, 2007), has been applied in numerous studies to evaluate drought and it was found that it is an appropriate indicator for the dryness of a particular area (Renza, Martinez, Arquero, & Sanchez, 2010;Gouveia, Bastos, Trigo, & DaCamara, 2012).The NDDI appears to respond to soil moisture based on drought conditions, and was used in this study to determine soil moisture.
The objectives of this study are to estimate the spatial distribution of soil moisture using VIs based on Landsat 8 OLI/TIRS data and to evaluate the use of soil moisture data for wildfire risk assessment.Specifically, this paper includes: (1) soil moisture estimates for mapping the spatial distribution of soil moisture by combining TVDI and NDDI based on a regression approach.We propose a possible adaptation and application of NDWI and LST for constructing a TVDI based on the similar design of the triangular NDVI-LST space.We then compare the efficiencies of NDVI-LST and NDWI-LST for calculating the TVDI.(2) An investigation of the relationship between estimated soil moisture and fuel moisture measured in the field to assess the suitability of the simulated soil moisture data for wildfire prediction.(3) We hypothesize that (i) the NDWI-LST relationship performs as well as or better than the NDVI-LST relationship and can be applied for calculating TVDI, and (ii) that estimated soil moisture derived from our model is directly related to fuel moisture, influencing wildfire occurrence.In this study, we used the Landsat 8 TIRS and MODIS products for calculating LST and the Landsat 8 OLI product for determining TVDI and NDDI.
This study could also be used as an approach to enhance the efficiency of wildfire assessment using soil moisture as a surrogate for fuel moisture, identifying areas prone to wildfire across different landscapes.Until now, Thailand has not widely applied remote sensing to wildfire management.Using soil moisture measured by remote sensing as a complementary dataset for wildfire management may have the unique potential to predict wildfire danger for Thailand's forest areas and enhance the effectiveness of planning and decision-making in the area of wildfire management.jas.ccsenet.Hong, Qin, & Zhu, 2013).Each soil sample was placed in a plastic container and sealed tightly for further laboratory analysis.For the gravimetric analysis of soil moisture, we first weighed the soil samples (wet weight in grams) using a standard laboratory scale and then placed them in a drying oven at 105 °C for 48 hours (Gardner, 1986).After drying, we weighed the dried soil samples (dry weight in grams).The percentage of gravimetric soil moisture was calculated using Equation 1:

Study
Five soil moisture measurements from each of the five subplots within each sample plot were averaged to obtain representative soil moisture for each 30-m 2 site, corresponding to the spatial resolution of the Landsat 8 images.
The averaged soil moisture data from 34 sample plots were used for both training (80%) and validation (20%) data.

Leaf Fuel Moisture Measurements
Leaf fuel was collected for fuel moisture measurements, which were used for analyzing the relationship with simulated soil moisture.We specifically focused on dead leaves on the ground surface, because those represent the largest fuel component.A small sample of leaf litter was randomly collected from each 1-m 2 subplot and then placed into a sealed envelope for further laboratory analysis.In the laboratory, leaf litter samples were weighed and oven-dried at 80 °C for 48 hours, then weighed again to calculate the fuel moisture content (FMC) in percent following the procedure described by Desbois, Deshayes and Beudoin (1997).The most common FMC calculation is the ratio of water to dry weight as expressed by Equation 2. The FMC values for the five subplots were averaged to obtain a representative FMC for each 30-m 2 sample plot.(2)

Remotely Sensed Data and Preprocessing
We used cloud-free Landsat 8 OLI/TIRS and MODIS eight-day composite LST datasets at a spatial resolution of 30 m and 1000 m, respectively, as primary data (Table 1 The LST is the temperature of the Earth's surface as derived from remotely sensed thermal infrared data (Weng, Fu, & Gao, 2014).It depends on the albedo, vegetation cover, and soil moisture.The Landsat 8 LST was computed by fusing images of MODIS LST and Landsat 8 brightness temperature (Tb), provided by Hazaymeh and Hassan (2015).Generating Landsat 8 LST was based on the linear relationship between MODIS LST and Landsat 8 Tb, which were obtained almost simultaneously and under similar atmospheric conditions.
A scatter plot of remotely sensed LST and VI often results in a triangular shape (Price, 1990;Carlson, Gillies, & Perry, 1994) and the "dry" and "wet" edges of the triangle can be used to obtain information on soil moisture content.Figure 2 shows the conceptual TVDI based on the NDVI-LST triangle, where LST is plotted as a function of NDVI.The linear combination of NDVI-LST typically shows a strongly negative relationship and the TVDI can be estimated from the dry and wet edges of the triangle.
jas.ccsenet.The NDDI was computed from the NDVI and NDWI values according to the definition proposed by Gu, Brown, Verdin, and Wardlow (2007).The combination of information about both vegetation (NDVI) and water (NDWI) conditions can be used to determine vegetation drought conditions, which reflect the effects of soil moisture.Due to the variation of the NDVI and NDWI within a range from -1 to +1, these values were converted to 8 bits (0-255) for the calculation of the NDDI, which ranges between -1 and +1.Higher NDDI values indicate more severe drought and lower soil moisture.The NDDI is computed as:

Soil Moisture Model and Validation
We established a soil moisture estimation model based on a collection of field sampling and remote sensing data.A stepwise multiple regression approach was used to assess the relationship between field soil moisture data and remote sensing data, i.e., TVDI and NDDI were used as independent variables.The model can be computed by a regression formula as follows: Where, the estimated soil moisture is given as a percentage (%), and a, b, and b' are the coefficients of the regression lines of the TVDI and NDDI.
The model was validated by ground and remote sensing data.We used the actual soil moisture from the field measurements to evaluate the accuracy of the predictive model by statistical inference: (i) adj-R 2 , (ii) root mean squared error (RMSE), (iii) absolute average difference (AAD), and (iv) the precision of the model.The precision (%) of the model is calculated as follows: Where, Yi is the actual soil moisture of the field samples (%), Y'i is the estimated soil moisture from remotely sensed data (%), and N is the sample size.
Finally, the validated model was applied to a Landsat 8 image acquired on 19 February 2015 in Sri Lanna National Park (dipterocarp and deciduous forests) in order to estimate and map the spatial soil moisture distribution during the dry season.

Analysis of the Relationship between Estimated Soil Moisture and Leaf Fuel Moisture
To investigate the relationship between soil moisture estimated from our model and FMC, we performed a correlation analysis using the Pearson correlation and linear regression methods.Estimated soil moisture was extracted from the model at the same locations as were used to measure leaf fuel moisture in the field to determine correlation.We then explored the possibility of applying estimated soil moisture from our model to the prediction of wildfire occurrences.

Results and Discussion
Scatter plots of the relationships between NDVI-LST and NDWI-LST are shown in Figure 3. Compared to the NDVI-LST plot, the NDWI-LST relationship shows a clearer triangular shape, following the theoretical triangle of the TVDI.We determined LST max (dry edge) and LST min (wet edge) to highlight linear trends.A comparison of pixels representing LST max and LST min extracted from the NDVI-LST and the NDWI-LST plots indicates a stronger relationship between these pixels in the NDWI-LST space.Based on Figure 3, the LST max , representing the dry edge, shows a strong negative correlation between the NDWI and LST (adj-R 2 = 0.84, p-value < 0.01), and the LST min , representing the wet edge, shows a negative correlation between the NDWI and LST with adj-R 2 = 0.63 at a significant level for p < 0.01.In contrast, NDVI has a lower correlation with LST, with LST max at adj-R 2 = 0.62 (p-value > 0.05) and LST min at adj-R 2 = 0.47 (p-value < 0.01).The results of the collinearity requirement indicate that the NDWI has a stronger negative correlation with the LST than the NDVI, which is why the NDWI was used to calculate TVDI. jas.ccsenet.
Figure 3.  .We sture, mated T and jas.ccsenet.The model allows to remotely determine the spatial distribution of soil moisture as a complementary dataset for identifying wildfire-prone areas, which is a fundamental step toward involving soil moisture in the assessment of wildfire risk.We therefore recommend soil moisture estimation by remotely sensed model as another indicator for monitoring wildfire risks and intensity.Furthermore, the demonstrated NDWI-LST relationship provides another option for researchers studying soil moisture when the established TVDI based on the NDVI-LST relationship is insufficient.Future studies should address soil moisture as one of the factors used for enhancing estimates of FMC, as soil moisture is shown to be correlated with FMC.
Figure 2. A