A dynamic ensemble selection algorithm based on three machine learning models for retrieving surface soil freeze/thaw from satellite across China

As surface soil freeze/thaw (FT) cycle serves as a "switch" of land surface processes, accurate retrieval of surface FT dynamics based on satellite passive microwave remote sensing is critical for studies on climate change and dynamics of cryosphere. In this study, we aim to improve the FT retrieval accuracy by developing a new FT retrieval algorithm based on the k-Nearest Oracle Union (KNORA-UNION) dynamic ensemble selection algorithm (DESA) that can optimally integrate three machine learning models, random forests, extra trees, and extreme gradient boosting, on a grid cell scale. We applied our developed DESA retrieval algorithm to retrieve daily surface FT states of China from 2009 to 2020 based on multi-band SSMIS brightness temperatures. We then evaluated our DESA by comparing the observations of 2398 Chinese stations and three other existing FT algorithms, including the modified seasonal threshold algorithm (MSTA), dual-index algorithm (DIA), and decision tree algorithm (DTA). The results show that DESA has the highest retrieval accuracy and the lowest biases across these Chinese stations among the four algorithms. The mean classification accuracy of DESA is 89% and 84% for the PM and AM overpasses, respectively. MSTA is also generally in a good agreement with the surface observations, though it has large biases in the northeastern stations, while DTA tends to underestimate the timings of primary fall freezing due to uncertainty resulted from a single brightness temperature threshold. This study clearly demonstrates that the KNORA-UNION dynamic ensemble selection model can provide satisfactory estimates of surface FT states. This study also provides a valuable multi-decadal record for daily FT states.