Downscaling Satellite Retrieved Soil Moisture Using Regression Tree‐Based Machine Learning Algorithms Over Southwest France

Abstract

Satellite retrieved soil moisture (SM) shows great potential in hydrological, meteorological, ecological, and agricultural applications, while the coarse resolution limits its utilization in regional scale. The regression tree-based machine learning algorithms reveal promising capability in SM downscaling. However, it lacks systematic study dedicated to intercomparisons of algorithms to explicitly illuminate their characteristics. In this study, comparisons are made to systematically evaluate performances of classification and regression tree (CART), random forest (RF), gradient boost decision tree (GBDT), and extreme gradient boost (XGB) in Soil Moisture Active Passive (SMAP) SM downscaling in southwest France. The results show that the four algorithms downscaled SM are capable of capturing spatial distribution features of the original SMAP SM. The downscaled regions with favorable accuracy are mostly situated in the dominant Mediterranean climate zone with moderate vegetation coverage and mild topography variation. The best results are obtained by GBDT in grassland with R value of 0.77 and ubRMSE value of 0.04 m 3 /m 3. The RF and XGB also achieve good performances. On the whole, the GBDT approach is robust and reliable, which could downscale SM with superior correlation and smaller bias than the others. Besides, it achieves higher accuracy than the original SMAP in grassland and shrubland. The feature importance index of each explainable variable fluctuates regularly among different seasons and models. This study proves the outstanding performance of GBDT in SMAP SM downscaling and is expected to act as a valuable reference for studies focusing on SM scale conversion algorithms.

Enhancing Spatial Resolution of SMAP Soil Moisture Products through Spatial Downscaling over a Large Watershed: A Case Study for the Susquehanna River Basin in the Northeastern United States

“…In general, the performance evaluation results indicated that the RF models developed at both spatial scales for all the experiments showed higher performance measures for the condition of our study area, which was a warm summer humid climate with a predominance of vegetation. This agreed quite well with the results reported by previous studies conducted in different regions of the world under varying climatic and land surface conditions using the RF model, e.g., the subtropical continental monsoon [72], semi-humid temperate monsoon [25], South Asian monsoon [23,62], and Mediterranean climate zone [73].…”

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

“…The overestimated phenomenon is located in the central and northern areas. This is consistent with earlier studies [1,46,59], which illustrate that machine learning methods tend to reduce the range of the target data. Therefore, the method will reduce the value of the high soil moisture area in the original AMSR-E data and increase the value of the low soil moisture area.…”

Machine learning based estimation of field-scale daily, high resolution, multi-depth soil moisture for the Western and Midwestern United States

“…The other studies which outperformed ours used alternative model calibration and validation strategies ( Table S1 ). For example, random division of calibration and validation datasets was adopted by studies that reported more promising model fits (R 2 > 0.8) ( Liu et al, 2020 ; Greifeneder, Notarnicola & Wagner, 2021 ; Zhang et al, 2021 ). It should be noted that validation against randomly selected data points, even if stratified by study years, can generate much better model fits compared to independent sites because of the use of autocorrelated time-series data for both model training and validation ( Meyer et al, 2018 ; Ploton et al, 2020 ).…”