Upscaling sparse, irregularly spaced in situ soil moisture measurements for calibration and validation of SMAP soil moisture products

Friday, 18 December 2015
Poster Hall (Moscone South)
Jane Whitcomb1, Daniel Clewley2, Mahta Moghaddam3, Ruzbeh Akbar3 and Agnelo Rocha da Silva3, (1)University of Southern California, Electrical Engineering, Los Angeles, CA, United States, (2)Plymouth Marine Laboratory, Remote Sensing Group, Plymouth, United Kingdom, (3)University of Southern California, The Ming Hsieh Dept. of Electr. Eng., Los Angeles, CA, United States
There is a large difference in the footprints over which remote sensing instruments, such as the Soil Moisture Active Passive (SMAP) mission, retrieve soil moisture and that of in situ networks. Therefore a method for upscaling in situ measurements is required before they can be used to validate remote sensing instruments. The upscaling problem is made more difficult when measurements are sparse and irregularly spaced within the footprint. To address these needs, we have developed a method for producing upscaled estimates of soil moisture based on a network of in situ soil moisture measurements and airborne P-band SAR data, and utilizing a Random Forests-based regression algorithm. Sites within the SoilSCAPE network, for which the technique was developed, typically contains sensors at ~30 locations, with each location sampled at multiple depths. Measurements are taken at 20 minute intervals and averaged over a selectable time interval, thereby supporting near-real time generation of soil moisture maps. The collected measurements are automatically uploaded to a central database from which they can be accessed for use in the regression algorithm. Our regression-based approach works well with irregularly-spaced sensors by incorporating a set of data layers that correlate well with soil moisture. The layers include thematic land cover, elevation, slope, aspect, flow accumulation, clay fraction, air temperature, precipitation, and P-Band HH, VV, and HV backscatter. Values from these data layers are extracted for each sensor location and applied to train the Random Forests algorithm. The decision trees generated are then applied to estimate soil moisture at a 100 m spacing throughout the network region, after which the evenly-spaced values are averaged to accord with the 3-, 9-, and 36-km SMAP measurement grids. The resulting set of near-real time soil moisture estimates suitable for SMAP calibration and validation is placed online for use by the SMAP Cal/Val team. Accuracy of the algorithm has been verified against three independent soil moisture sensors from Oregon State University, with RMS errors ranging from 0.034 to 0.063 volumetric soil moisture.