3. REMOTE SENSING OF SOIL MOISTURE
3.4 Downscaling Techniques
measurements, the AMSR-E product gives reasonable results in terms of correlation and the ASCAT product was unstable.
iv. A study conducted by Albergel et al. (2012) evaluated the ASCAT and SMOS products against in-situ soil moisture observations from over 200 stations across the world for the year 2010. A similar study was conducted by Brocca et al. (2011), which evaluated the ASCAT and AMSR-E satellite-based soil moisture products around Europe. The main purpose was to evaluate the potential of different ASCAT and AMSR-E products in obtaining reliable estimates of soil moisture. The study concluded that the AMSR-E LPRM provided the best results.
The consistent theme, with regards to downscaling procedures, which is identified in these key publications, is the combined use of passive microwave data with fine-scale optical data (surface temperature and vegetative indexes). The overall aim of these downscaling techniques are to provide soil moisture estimates at the same accuracy as the input remote sensing soil moisture product, but at the spatial resolution of the optical data used. These key downscaling publications are evaluated and summarized in Table 3.1 below.
Table 3.1 Evaluation of downscaling techniques
Author Merlin et al Piles et al Merlin et al Merlin et al
Year 2012 2011 2010 2009
Region Yanco, South-eastern Australia (2010)
Yanco, South-eastern Australia (2010)
Yanco, South-eastern Australia (2006)
Yanco, South-eastern Australia (2006)
Input data MODIS (LST, emissivity, NDVI and Albedo)
MODIS VIS/IR data ( LST and NDVI)
MODIS (LST, red and infrared reflectance and NDVI)
Wind speed, MODIS data (surface
temperature, NDVI), ASTER (radiometric surface temperature)
Product SMOS SMOS SMOS Simulated SMOS
Output
resolution (km)
1 10 and 1 4 0.5
Methodology sequential model Build model between NDVI, LST and soil moisture
Relationship between soil moisture and soil evaporative efficiency
sequential model
Results R2 = 0.70-0.85 in summer, however very poor results obtained in winter
R2 of 0.14-0.21, RMSE is between 0.9-0.17
Mean slope between simulated and observed is 0.94, with an error of 0.012
R2 of 0.68, RMSE of 0.062, bias of 0.045
The evaluation of the abovementioned methods highlights significant limitations, which hinder the successful application of the various downscaling procedures that have been developed. These include (a) observational days have to be cloud-free, to avoid obscurities in data retrieval; (b) the presence of vegetation interferes with land surface temperature retrieval; (c) there is a difference in the input data sensing depth; and (d) the model assumptions may not be valid in heterogeneous areas (Merlin et al., 2008; 2009; 2010; Piles et al., 2011).
From the evaluation of these downscaling techniques, several comparisons can be made.
Firstly, all of the downscaling research studies were conducted in the same region; this simplifies the evaluation process between the different techniques used. It can be noted that the technique used by Piles et al. (2011) requires the least data input, while the technique used by Merlin et al. (2009) requires the most input data. All the techniques use SMOS data or a simulated SMOS data set, as the SMOS satellite is the most recent soil moisture satellite.
The evaluation between the methods is enhanced due to the common aspects of the research studies.
The differences in the evaluated techniques can be seen in their methodologies, output resolutions and results. Merlin et al. (2012) had an output resolution of 1 km and showed good results in summer; however, the technique performed very poorly in winter, when compared to in-situ soil moisture data. Piles et al. (2011) had an output resolution of 1 and 10 km, but performed poorly, when compared to in-situ soil moisture data. The Merlin et al.
(2010) resulted in the best correlation between observed and simulated soil moisture. Merlin et al. (2009) showed good correlation and had the finest resolution.
The limitations of the downscaling process can be summarized as follows:
i. The accuracy of the soil moisture product may decrease as the spatial resolution increases (there is a trade-off between obtaining accurate data and obtaining fine resolution products);
ii. The process requires input data, which may not be freely or readily available;
iii. The technique may be site-specific;
iv. The complexity of the algorithms used; and
v. Cloud-free images are required, if MODIS products are used.
In recent years, the need for fine resolution soil moisture products has resulted in numerous downscaling research studies being conducted. The main publications in this field have been studies conducted by Merlin et al. (2009; 2010; 2012) and Piles et al. (2011). In addition to these, more recent studies have been conducted, which are based on the same principles of the key techniques. However, there are slight variations, in order to improve and build upon these key approaches. The new studies include:
i. A study conducted by Zhao and Li (2013) aimed to develop a downscaling method to improve the spatial resolution of the AMSR-E derived soil moisture product. The approach was based upon the conventional method of the microwave-optical synergistic technique, which uses LST, vegetative indexes and albedo. This approach replaces LST with two temperature temporal variation parameters. The study was conducted in the Iberian Peninsula for the year 2007. The results showed an improvement in the approach (R2 increased by 0.08), when the new approach was compared to the conventional method.
ii. A recent study by Ruiz et al. (2014) was conducted on combining SMOS visible and near infrared satellite data for high resolution soil moisture over a two-year period in the REMEDHUS network in Spain. The study used a new downscaling algorithm, based on a relationship between LST, NDVI and brightness temperature. The study aimed to obtain a downscaled image with the same accuracy of SMOS, but at the spatial resolution of MODIS (500 m). The best result obtained was a correlation with the in-situ measurements of 0.72.