*2.4. AVHRR Data*

The advanced very high resolution radiometer (AVHRR) utilizes National Oceanic and Atmospheric Administration (NOAA) polar-orbiting satellites to provide four- to six- band multispectral global data [31]. The AVHRR is used to remotely detect cloud cover and the Earth's surface temperature (NOAA satellite information system, 2013). Prior to MODIS data, AVHRR's 5 km CMG NDVI data were used for long-term surface ground measurements [11]. In this study, daily NDVI data (AVH13C1) from AVHRR from 1981 to 1999 were used. The quality of AVHRR data after this time period is inadequate due to satellite drifting and, therefore, data after 2000 was not used in this study [11] (Table 1).


**Table 1.** Description of the data products used in the downscaling process including their spatial and temporal resolutions and data availability.

#### **3. Methodology**

In this research, the downscaling algorithm and methodology used were developed in Fang et al., 2018 [32]. Similar to the study by Lakshmi and Fang (2015) of the Little Washita Watershed, this study assumes that LST is a linear combination of soil and vegetation temperature [33]. We assume the top soil moisture layer is a function of soil evaporation efficiency and field capacity. It is assumed that the soil moisture at a certain time during the day is inversely proportional to the daily temperature change for the same day, and that the presence of vegetation (NDVI) will influence the soil moisture–temperature change relationship. We also assume that the thermal inertia relationship between temperature difference and soil moisture within a 25 km domain has no spatial variability. Additionally, the assumption is made that the field capacity of each NLDAS pixel is homogenous and does not account for variation at the 1 km scale [32].

In this study, we applied an algorithm developed by Fang et al., 2018, based on soil moisture, LST, and NDVI, to create 1 km soil moisture maps [32]. The methodology of this algorithm is outlined in Figure 2. Due to the effects of vegetation cover on soil moisture estimation, the algorithm applied

here uses a vegetation-based lookup table to relate microwave polarization to soil moisture estimates. As soil becomes more wet its heat capacity increases. The soil moisture at a given time is inversely proportional to the change in temperature 12 hours beforehand, which corresponds with SMAP AM and PM overpasses. Soil moisture daily values were negatively related to the daily temperature difference under varying vegetation conditions. The following equation represents the linear relationship between soil moisture and temperature difference for a specific NDVI (single month):

$$
\Delta\Theta(\mathbf{i}, \mathbf{j}) = \mathbf{a}\mathbf{o} + \mathbf{a}\_1 \Delta T\_\theta(\mathbf{i}, \mathbf{j}) \tag{1}
$$

where θ(i, j) is GLDAS soil moisture gridded to match SMAP overpasses and ΔTs(i, j) is the GLDAS 12 h temperature difference closest and prior to SMAP overpasses. This equation uses data at the GLDAS spatial resolution for soil moisture and surface temperature for single months, beginning in 1981. Using the nearest neighbor method, daily NDVI from AVHRR was aggregated to corresponding GLDAS pixels. The NDVI data were categorized into classes from 0 to 1 with increments at 0.1. Classes with less than 8 data points were not included because a sample size smaller than this will not yield valid and statistically significant results from linear regression fitting. Soil moisture at 1 km resolution was calculated from 1 km MODIS LST difference at the corresponding NDVI class. We applied the linear regression fit equation between θ and ΔTs, which was built at 25 km resolution, to all the 1 km MODIS grids within the 25 km GLDAS grid. We assumed that the thermal inertia relationship between temperature difference and soil moisture within the 25 km domain had no spatial variability. The following equation represents the correction of the 1 km soil moisture pixel from the MODIS LST products, acquired by removing the difference between SMAP and MODIS derived soil moisture:

$$\Theta^{\rm corr}(\mathbf{i}, \mathbf{j}) = \Theta(\mathbf{i}, \mathbf{j}) + \left[\Theta - \frac{1}{\mathbf{n}} \sum\_{i=1}^{n} \Theta\_{\mathbf{i}}\right] \tag{2}$$

where θcorr(i, j) is the corrected 1 km soil moisture, n is the number of 1 km soil moisture pixels that are in each SMAP 9 km pixel, Θ is the original SMAP 9 km soil moisture estimate, and θ<sup>i</sup> is the number of uncorrected 1 km SMAP soil moisture pixels that fall in the original 9 km SMAP grid Θ. The value of n is ideally 81, but it may be less due to cloud contaminated data. The corrected soil moisture was characterized by the soil moisture and daily temperature relationship, which changed under different vegetation conditions. Since visualizing rainfall is essential to determining the response of the soil moisture, rainfall from GPM IMERG was used in this study to analyze the wetting and dry-down patterns after a significant rainfall event. One limitation of this methodology occurred when the 9 km original SMAP was biased. Then that bias was passed onto the corrected 1 km soil moisture. Another limitation was the difficulty to recover cloud-contaminated data, which resulted in spatial inconsistencies in the 1 km corrected soil moisture maps.

**Figure 2.** Workflow for building downscaling model and executing the algorithm.

The algorithm used in this study was validated using in situ measurements in the CONUS region, by Fang et al., 2018, for soil moisture estimates from AMSR2 between 2015 and 2017. Their validation showed variability in seasonal performance and stronger correlations in the soil moisture–temperature change relationship during summer months. Also, the remotely sensed soil moisture and downscaled estimates both underestimated in situ soil moisture during precipitation events. It is important to note the effects of precipitation on soil moisture retrieval; the microwave sensing depth is reduced. An additional validation of this algorithm was performed in the Walnut Gulch Experimental Watershed (WGEW) and indicated that downscaled soil moisture had better validation metrics than the original SMAP [32]. The R<sup>2</sup> of the 1 km soil moisture ranged from 0.189 to 0.697, whereas the 9 km SMAP ranged from 0.003 to 0.597. The slope values for the 1 km are higher than those for the 9 km SMAP. Additionally, the 1 km soil moisture RMSE values and biases improved compared to the original SMAP data. There were no consistent soil moisture measurements in the Lower Mekong Basin, and this presents a formidable challenge to validation. However, future work may be able to carry out validation by comparison of the 1 km soil moisture to outputs from hydrological models.

#### **4. Results**

#### *4.1. Rainfall Variation in the Lower Mekong Basin*

Variations in rainfall patterns result in changes in soil moisture. Precipitation has a direct impact on the wetting and drying of soils and, therefore, must be examined alongside soil moisture. In the LMB, the annual wet season (April–September) results in more vegetation growth and cloud cover compared to the dry season. Therefore, the ability to measure soil moisture via remote sensing is affected during these months. Daily precipitation data from GPM IMERG Final Precipitation L3 1 day 0.1◦ by 0.1◦ V05 (GPM\_3IMERGDF) were aggregated for monthly accumulation for April through September from 2015 to 2018, to correspond with the downscaled soil moisture in order to examine the monthly variations (Figure 3) [34].

**Figure 3.** Bar plot of monthly average precipitation for April 2015–September 2018 in the LMB.

In this study, precipitation patterns varied in the wet season months, with July and August generally accumulating the most rainfall and April and May receiving the least (Figure 3). Additionally, precipitation varied from year to year over the LMB, with certain years being more dry or wet than others due to regulation by monsoons. For example, comparing the year 2016 to 2018 in Figure 3

shows 2016 as a much dryer year, especially in the wettest month of the year, July, which received over 100 mm of rainfall. This pattern can also be seen by comparing the monthly maps from 2016 and 2018 (Figure 4). Figure 4 shows the spatial distribution of accumulated precipitation over the LMB for each month, corresponding to the 1 km soil moisture estimates. Rainfall patterns varied significantly between countries in the LMB. Areas in Laos and Cambodia receive the greatest amounts of precipitation annually (over 2800 mm), while the Thailand plateau only received a third or less of that amount. Here, precipitation from IMERG was used to detect and observe the dry-down patterns of soil moisture after a large rainfall event.

**Figure 4.** Monthly rainfall accumulation from GPM IMERG for April 2015–September 2018 in the LMB.
