Comparison of Different Machine Learning Approaches for Monthly Satellite-Based Soil Moisture Downscaling over Northeast China

Abstract

Although numerous satellite-based soil moisture (SM) products can provide spatiotemporally continuous worldwide datasets, they can hardly be employed in characterizing fine-grained regional land surface processes, owing to their coarse spatial resolution. In this study, we proposed a machine-learning-based method to enhance SM spatial accuracy and improve the availability of SM data. Four machine learning algorithms, including classification and regression trees (CART), K-nearest neighbors (KNN), Bayesian (BAYE), and random forests (RF), were implemented to downscale the monthly European Space Agency Climate Change Initiative (ESA CCI) SM product from 25-km to 1-km spatial resolution. During the regression, the land surface temperature (including daytime temperature, nighttime temperature, and diurnal fluctuation temperature), normalized difference vegetation index, surface reflections (red band, blue band, NIR band and MIR band), and digital elevation model were taken as explanatory variables to produce fine spatial resolution SM. We chose Northeast China as the study area and acquired corresponding SM data from 2003 to 2012 in unfrozen seasons. The reconstructed SM datasets were validated against in-situ measurements. The results showed that the RF-downscaled results had superior matching performance to both ESA CCI SM and in-situ measurements, and can positively respond to precipitation variation. Additionally, the RF was less affected by parameters, which revealed its robustness. Both CART and KNN ranked second. Compared to KNN, CART had a relatively close correlation with the validation data, but KNN showed preferable precision. Moreover, BAYE ranked last with significantly abnormal regression values.

Can Terrestrial Water Storage Dynamics be Estimated From Climate Anomalies?

“…As two of the most distinguished ensemble learning algorithms, the RF and XGB have been reported strongly outperforming many other classic machine learning algorithms, such as support vector machine, k-nearest neighbors, decision trees, and artificial neural network (ANN), in many applications. Successful stories could be found in precipitation and soil moisture mapping and downscaling (Jing et al, 2016;Liu et al, 2018), land cover mapping (Cracknell & Reading, 2014), surface temperature rescaling (Hutengs & Vohland, 2016), and geological mapping (Cracknell & Reading, 2014). Their flexibility (e.g., no required feature normalization and simple parameters) and tree-based structure make them easier and more efficient to be used and interpreted than support vector machine, k-nearest neighbors, and ANN.…”

A New Approach for Soil Moisture Downscaling in the Presence of Seasonal Difference

“…Soil moisture is a key factor in agriculture research, but the spatial resolution of current products is insufficient to meet research needs. The variation of SM is a complex and synthetic process, which is impacted by numerous factors [19]. In recent years, more attention has been paid to the influence of spatial heterogeneity rather than temporal heterogeneity in the research on improving the spatial resolution of soil moisture remote sensing products.…”

“…The ground data stations record soil relative humidity. Due to the lack of SM monitoring measurements at 5 cm, we chose soil moisture relative humidity data at a depth of 10 cm, as in past research [16,19,40]. Although the measurement depths are inconsistent, there is a strong correlation between the SM values of the two continuous soil layers [16,41].…”

“…Machine learning methods can express well this regression correlation [16]. The machine learning algorithms used for downscaling include random forest (RF), support vector machines (SVM), K-nearest neighbor (KNN) [16][17][18][19][20]. These algorithms have their own characteristics, so it is necessary to choose the most suitable algorithm for the research area [19].…”

“…Such seasonal differences can also be reflected in downscaling studies. Yangxiaoyue et al [19] and Jovan et al [37] all mentioned that the accuracy of downscaled products varied with seasons, and the accuracy of downscaled products was lower in rainy seasons. That is to say, the downscaling model established by continuous time series products cannot capture well the seasonal difference.…”