Missing data reconstruction and evaluation of retrieval precision for AMSR2 soil moisture

Abstract: Quality and precision of soil moisture data derived from a newly launched sensor, advanced microwave scanning radiometer 2 (AMSR2), needs to be evaluated before its quantitatively application in such fields as hydrological and energy cycle, agricultural management and so on. AMSR2 level 3 soil moisture data estimated based on the Land Parameter Retrieval Model (LPRM) has the highest spatial resolution of 10 km. The purpose of the paper was to reconstruct missing data and evaluate its retrieval precision of AMSR2 level 3. The missing data reconstruction was based on a penalized least square regression with three-dimensional discrete cosine transform (DCT-PLS) method. In order to evaluate the feasibility of the method, AMSR2 data in 2013 from 2 randomly selected locations (Yugan county of Jiangxi Province and Pingdu of Shandong Province) with continuous data originally were used and 20% pixels were given NaN for DCT-PLS-based missing data reconstruction. The reconstructed result was compared with the original one to evaluate the DCT-PLS method. Meanwhile, the AMSR2 data for the whole China on June 1, 2013 was used for the method validation in reconstructing missing data (eg. Anhui, Shanxi, Taiwan, and East China). Moreover, MODIS products within Shanxi area including land surface temperature (MOD11A2), vegetation index (MOD13A2) and surface albedo (MCD43B3) were combined to downscale AMSR2 level 3 soil moistures obtained on Jun. 1, 2013 and Nov.1, 2012 by using a statistical regression method. The retrieved soil moisture with the spatial resolution of 10 and 1 km were evaluated based on the field measurement of 38 stations as well as temperature vegetation drought index (TVDI) computed based on MODIS products. The results showed that: 1) the correlation coefficient (r) of reconstructed and original data was 0.9834 and 0.9557 (P<0.001) in both locations. In spatial distribution, the reconstructed data for the whole China had high correlation with the original data (r=0.9255, P<0.001). The reconstructed data could reveal a reasonable continuous distribution. Hence, the DCT-PLS method was reliable for missing data reconstruction; 2) Though the spatial distribution of AMSR2 soil moisture at the two different spatial scales was consistent with that of the field measurements, the correlations were low on Nov. 1, 2012 and Jun 1, 2013 respectively and they were improved by downscaling. There was a negative correlation between the AMSR2 soil moisture and TVDI at 10 km resolution. Compared with 10 km, the 1 km resolution showed reasonable spatial distribution of soil moisture. The correlation between the soil moisture and TVDI at 10 km resolution was -0.1869, -0.4720 on Jun 1, 2013 and Nov. 1, 2012 respectively, and it increased to -0.5389, -0.8984 at 1 km resolution. Moreover, the downscaled data at 1 km resolution could reflect more spatio-temporal distribution details of soil moisture, which could reduce unreliable estimation for higher soil moisture at 10 km resolution. Therefore, the 1 km downscaled data was more reliable. In sum, the DCT-PLS reconstruction combined with downscale method is good for soil moisture retrieval with a higher spatial resolution for AMSR2 level 3 data.