A framework to nowcast soil moisture with NASA SMAP level 4 data using in-situ measurements and deep learning

Abstract

Study Region: Southeast Texas, USA. Study Focus: NASA's Soil Moisture Active Passive (SMAP) product, particularly the Level 4 (SMAPL4) data, provides high-resolution and extensive coverage of surface and root zone soil moisture (SM), essential for weather and climate research. However, a latency of 2.5–4.0 days in SMAPL4 data limits its real-time hydrologic and weather prediction applications. To address this, we developed a model integrating deep learning (DL) techniques (Long Short-Term Memory, Fully Connected Neural Network) with Principal Component Analysis (PCA) to nowcast SM data in real-time. The model is trained on multi-source SM observations, including near real-time in-situ and satellite data, and deployed over a 56,000+ km² area in southeast Texas. New Hydrological Insights for the Region: Our DL methodology nowcasts SM accurately in both time and space through real-time assimilation of multi-source data, mitigating SMAP's latency and offering near real-time soil moisture estimates. The nowcasted SM aligns closely with actual SMAPL4 data, capturing spatial and temporal variations. SMAP underestimates the spatio-temporal variability of soil moisture compared to in-situ data, highlighting the necessity for diverse data integration. The proposed framework can improve the real-time flood and drought monitoring and offers insights for various hydrological applications. Nowcasting error mapping identifies regions with higher uncertainties, guiding future model improvements.

Keywords

• Soil moisture
• SMAP, In-situ Data, Deep Learning
• Nowcasting
• Spatiotemporal