Downscaling Soil Moisture with Simple Machine Learning Ensembles

Soil moisture is a key factor that influences the productivity and energy balance of ecosystems and biomes. Global soil moisture measurements have coarse native resolutions of 36km and infrequent revisits of around three days. However, these limitations are not present for many variables connected to soil moisture such as land surface temperature and evapotranspiration. For this reason many previous studies have aimed to discern the relationships between these higher resolution variables and soil moisture to produce downscaled soil moisture products.

In this study, we test four ensembles of simple machine learning models for this downscaling task. These ensembles use a dataset of over 1,000 sites across the US to predict soil moisture at sub-km scales. We find that all ensembles, particularly one with a very simple structure, can outperform SMAP on  a cross-fold analysis of the 1,000+ sites. This ensemble has an average ubRMSE of 0.058 vs SMAPs 0.065 and an average R of 0.639 vs SMAPs 0.562. However, not all ensembles are beneficial, with some architectures performing better with different training weights than with ensemble averaging. Additionally, we find that although general improvements over SMAP are observed, there appears to be difficulty in consistently doing so in cropland regions with high clay and low sand content.

Key Points:

• Ensembles of simple ML architectures can downscale SWC predictions to sub 1km resolutions
• Simpler architectures can outperform these ensembles and may be further enhanced with an improved weighting scheme during training
• Training the models on temporally padded data provides more benefits than drawbacks in terms of overall performance.
• The top performing ensemble is unreliable on croplands with higher than average clay and lower than average sand content.