Downscaling of SMAP Soil Moisture Data by Using a Deep Belief Network

Abstract

The spatial resolution of current soil moisture (SM) products is generally low, consequently limiting their applications. In this study, a deep belief network-based method (DBN) was used to downscale the Soil Moisture Active Passive (SMAP) L4 SM product. First, the factors affecting soil surface moisture were analyzed, and the significantly correlated ones were selected as predictors for the downscaling model. Second, a DBN model was trained and used to downscale the 9 km SMAP L4 SM to 1 km in the study area on 25 September 2019. Validation was performed using original SMAP L4 SM data and in situ measurements from SM and temperature wireless sensor network with 34 sites. Finally, the DBN method was compared with another commonly used machine learning model-random forest (RF). Results showed that (1) the downscaled 1 km SM data are in good agreement with the original SMAP L4 SM data and field measured data, and (2) DBN has a higher correlation coefficient and a lower root mean square error than those of RF. The coefficients of determination for fitting the two models with the measured data at the site were 0.5260 and 0.4816, with relative mean square errors of 0.0303 and 0.0342 m3/m3, respectively. The study also demonstrated the applicability of the DBN method to AMSR SM data downscaling besides SMAP. The proposed method can provide a framework to support future hydrological modeling, regional drought monitoring, and agricultural research.

Keywords

• deep learning
• deep belief network
• soil moisture
• SMAP L4
• Shandian River Basin