Downscaling of AMSR-E Soil Moisture over North China Using Random Forest Regression

Hongyan Zhang, Shudong Wang, Kai Liu, Xueke Li, Zhengqiang Li, Xiaoyuan Zhang, Bingxuan LiuISPRS Int. J. Geo Inf 2022 · DOI

TL;DR

This paper tackles the challenge of coarse resolution soil moisture data from satellites by developing a Random Forest (RF) regression model to downscale AMSR-E data from 25 km to 1 km resolution over North China. The study demonstrates that this RF-based approach, incorporating auxiliary data like elevation, temperature, and vegetation indices, significantly improves soil moisture estimation accuracy compared to the original coarse data and other traditional downscaling methods.

What Problem Does This Paper Address?

Accurate soil moisture data is crucial for understanding hydrological processes, agricultural productivity, and drought monitoring. Satellite-based soil moisture products, like those from AMSR-E, offer broad coverage but often suffer from coarse spatial resolution (e.g., 25 km). This low resolution limits their utility for local and regional studies, especially in complex landscapes like North China, where rapid industrialization and water consumption exacerbate drought conditions. Existing downscaling methods have limitations, including data requirements, regional specificity, or difficulty in handling complex terrain and diverse land use. This paper aims to bridge this gap by developing a robust and scalable downscaling method for AMSR-E soil moisture data over North China.

Methods

The core of this study is the application of a Random Forest (RF) regression model for downscaling. RF is chosen for its ability to handle non-linear relationships and its robustness to outliers. The authors propose a novel framework that integrates five auxiliary datasets with RF:

  • Auxiliary Data: Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI), Surface Albedo (ALB), Digital Elevation Model (DEM), and Precipitation. These are selected for their known influence on soil moisture and their availability at higher spatial resolutions.
  • Spatiotemporal Window Strategy: An adaptive window size determination method is employed to balance the risk of overfitting with the need to capture local variations.
  • Upscaling: To train the RF model, the higher-resolution auxiliary data are upscaled to match the 25 km resolution of the AMSR-E data.
  • Downscaling: The trained RF model then uses the 1 km resolution auxiliary data to predict soil moisture at a finer scale.

This approach differs from traditional methods by leveraging the ensemble learning capabilities of RF and incorporating a comprehensive set of auxiliary variables with a sophisticated windowing strategy.

Datasets

The study utilizes the following datasets:

  • AMSR-E Soil Moisture: Level 3 data from 2009, with a spatial resolution of 25 km.
  • Auxiliary Data (Higher Resolution):
    • MODIS Land Surface Temperature (LST): MOD11A1 product, 1 km resolution.
  • MODIS Normalized Difference Vegetation Index (NDVI): MOD13A2 product, 1 km resolution (16-day composite).
  • MODIS Surface Albedo (ALB): MCD43C3 product, 0.05° resolution.
  • SRTM DEM: 90 m resolution.
  • Precipitation: China Meteorological Forcing Dataset (CMFD), 0.1° resolution.
  • In Situ Soil Moisture Observations: Field measurements from 7 ground stations in North China for 2009, used for validation.

The study area is North China, covering approximately 14,900 km², encompassing diverse climates and land use types.

Key Insights and Ideas

  • DEM as a Dominant Factor: Digital Elevation Model (DEM) emerged as the most important explanatory factor for downscaling, highlighting its crucial role in capturing spatial variations in soil moisture across complex terrain.
  • Synergistic Effect of Auxiliary Data: The study strongly suggests that combining multiple auxiliary variables (DEM, Albedo, NDVI, LST, and precipitation) significantly improves downscaling accuracy compared to using fewer variables.
  • RF's Superiority: The Random Forest model demonstrated superior performance over traditional methods like Multiple Linear Regression (MLR) and Support Vector Regression (SVR), likely due to its ability to model complex non-linear relationships and its inherent resistance to overfitting.
  • Data Range Compression: Machine learning methods, including RF, tend to reduce the range of the target variable. This was observed as an underestimation in high soil moisture areas and overestimation in low soil moisture areas, a common characteristic that needs consideration.

Key Results

  • Improved Accuracy: The downscaled AMSR-E soil moisture data achieved a correlation coefficient (R) of 0.19 with in situ measurements, a significant improvement over the original AMSR-E data (R=0.02).
  • Reduced Errors: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and unbiased RMSE (ubRMSE) were reduced from 0.20, 0.15, and 0.19 m³/m³ (AMSR-E) to 0.18, 0.14, and 0.16 m³/m³ (downscaled), respectively.
  • Comparison with Other Methods: The RF method outperformed MLR and SVR, achieving higher R values and lower RMSE/MAE. For instance, the RF method yielded an R of 0.261 and RMSE of 0.181 m³/m³, compared to MLR (R=0.217, RMSE=0.201) and SVR (R=0.063, RMSE=0.213).
  • Spatial Pattern Preservation: The downscaled soil moisture maps retained the general spatial patterns of the original AMSR-E data while providing finer details.

Limitations and Future Work

  • Limited Ground Stations: The study acknowledges that the verification results have limitations due to the relatively few ground stations available in North China. Future work could benefit from collecting more in situ data.
  • Original Data Quality: The quality of the original AMSR-E data can influence downscaling results. Exploring other satellite soil moisture products or higher-quality data could be beneficial.
  • Validation Metrics: While standard metrics were used, the authors suggest incorporating more advanced validation metrics like Empirical Orthogonal Functions (EOF) and Spatial Efficiency metric (SPAEF) in future studies.
  • Ensemble Methods: The study primarily relies on RF. Future research could explore ensemble machine learning algorithms for potentially even better performance and robustness.
  • Broader Application: The authors suggest applying the downscaling method to other regions and combining it with other data sources or directions to expand its applicability.

Summary updated

BibTeX
@article{zhang2022downscaling,
  author = {Hongyan Zhang and Shudong Wang and Kai Liu and Xueke Li and Zhengqiang Li and Xiaoyuan Zhang and Bingxuan Liu},
  title = {{Downscaling of AMSR-E Soil Moisture over North China Using Random Forest Regression}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2022},
  doi = {10.3390/ijgi11020101},
}
Figure