Rebuilding Long Time Series Global Soil Moisture Products Using the Neural Network Adopting the Microwave Vegetation Index

• Main Authors: Panpan Yao, Jiancheng Shi, Tianjie Zhao, Hui Lu, Amen Al-Yaari
• Format: Article
• Language: English
• Published: MDPI AG 2017-01-01
• Series: Remote Sensing
• Subjects: soil moisture; neural network; long time series; microwave vegetation index
• Online Access: http://www.mdpi.com/2072-4292/9/1/35

This study presents a back propagation neural network (BPNN) method to rebuild a global and long-term soil moisture (SM) series, adopting the microwave vegetation index (MVI). The data used in our study include Soil Moisture and Ocean Salinity (SMOS) Level 3 soil moisture (SMOSL3sm) data, the Advanced Microwave Scanning Radiometer-Earth Observing System (AMSR-E), and Advanced Microwave Scanning Radiometer 2 (AMSR2) Level 3 brightness temperature (TB) data and L3 SM products. The BPNNs on each grid were trained over July 2010–June 2011, and the entire year of 2013, with SMOSL3sm as a training target, and taking the reflectivities (Rs) of the C/X/Ku/Ka/Q bands, and the MVI from AMSR-E/AMSR2 TB data, as input, in which the MVI is used to correct for vegetation effects. The training accuracy of networks was evaluated by comparing soil moisture products produced using BPNNs (NNsm hereafter) with SMOSL3sm during the BPNN training period, in terms of correlation coefficient (CC), bias (Bias), and the root mean square error (RMSE). Good global results were obtained with CC = 0.67, RMSE = 0.055 m3/m3 and Bias = −0.0005 m3/m3, particularly over Australia, Central USA, and Central Asia. With these trained networks over each pixel, a global and long-term soil moisture time series, i.e., 2003–2015, was built using AMSR-E TB from 2003 to 2011 and AMSR2 TB from 2012 to 2015. Then, NNsm products were evaluated against in situ SM observations from all SCAN (Soil Climate Analysis Network) sites (SCANsm). The results show that NNsm has a good agreement with in situ data, and can capture the temporal dynamics of in situ SM, with CC = 0.52, RMSE = 0.84 m3/m3 and Bias = −0.002 m3/m3. We also evaluate the accuracy of NNsm by comparing with AMSR-E/AMSR2 SM products, with results of a regression method. As a conclusion, this study provides a promising BPNN method adopting MVI to rebuild a long-term SM time series, and this could provide useful insights for the future Water Cycle Observation Mission (WCOM).