MULTI-SENSOR LONG-TERM GLOBAL SOIL MOISTURE MAPPING WITH ENHANCED SPATIO-TEMPORAL COVERAGE

T. Zhao¹

¹Aerospace Information Research Institute, Chinese Academy of Sciences

Soil moisture is a fundamental variable governing the exchanges of water, energy, and carbon cycles between the land surface and the atmosphere. Its accurate monitoring at a global scale is therefore critical for advancements in weather forecasting, climate change research, agricultural management, and drought early warning systems. While satellite-based passive microwave remote sensing provides a unique capability for all-weather, large-scale soil moisture observation, the creation of a consistent long-term data record is significantly challenged by the multitude of sensors with differing orbital characteristics, retrieval algorithms, and service periods. These inconsistencies lead to substantial gaps and disparities in spatio-temporal coverage, limiting the utility of the data for robust trend analysis and process studies.

This study introduces a novel two-stage paradigm that synergistically integrates physics-based retrieval with machine learning to overcome these multi-sensor inconsistencies and generate a harmonized, long-term global soil moisture dataset. In the first stage, we leverage the superior sensitivity of L-band observations from the Soil Moisture Active Passive (SMAP) mission. A sophisticated physical retrieval algorithm, the Multi-Channel Collaborative Algorithm (MCCA), is employed to generate high-fidelity soil moisture estimates. This MCCA SMAP product serves as a physically anchored and high-quality training target for the machine learning model, ensuring the foundation of our approach is rooted in robust radiative transfer principles.

In the second stage, we harness the extensive historical archive of brightness temperature observations from a constellation of key passive microwave sensors: the Tropical Rainfall Measuring Mission’s Microwave Imager (TMI, 1997-2015), the Advanced Microwave Scanning Radiometer for EOS (AMSR-E, 2002-2011), the Global Precipitation Measurement mission’s Microwave Imager (GMI, 2014-present), and the Advanced Microwave Scanning Radiometer 2 (AMSR2, 2012-present). These multi-sensor observations are first cross-calibrated to ensure inter-sensor consistency. A global grid-specific Long Short-Term Memory (LSTM) neural network is then designed and trained to learn the complex, non-linear mapping between these multi-frequency brightness temperatures and the target MCCA SMAP soil moisture values.

The trained LSTM model is subsequently applied to the full historical record of each sensor, producing four harmonized soil moisture data streams. These are then fused into a single, seamless daily global product, named MCCA-ML, providing a continuous record from 1997 to 2023 at a 25 km spatial resolution. Extensive validation against a comprehensive suite of in-situ measurements from dense soil moisture networks across the globe demonstrates that the MCCA-ML product successfully captures the spatial patterns and seasonal dynamics of soil moisture. The product shows remarkable consistency with the physical benchmarks set by MCCA SMAP while dramatically improving upon the spatio-temporal coverage of any single sensor. Notably, the daily global land coverage frequently exceeds 80% during periods when two or more satellites were operational, far surpassing the coverage of existing merged products.

This work demonstrates the powerful synergy between physical modeling and machine learning for Earth observation. The resulting MCCA-ML dataset effectively bridges the gaps between multiple satellite missions, mitigating the limitations of individual sensors and traditional fusion techniques. It provides the research community with a long-term, high-coverage, and consistent soil moisture data record that is invaluable for studying long-term hydrological and climatic trends, validating land surface models, and monitoring global environmental change.