A machine learning-based dynamic ensemble selection algorithm for microwave retrieval of surface soil freeze/thaw: A case study across China

Abstract

ABSTRACT The surface soil freeze/thaw (FT) cycle serves as a “switch” for 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 the cryosphere. This study aims to improve FT retrieval accuracy by developing a new FT retrieval algorithm that applies the K-Nearest Oracle Union (KNORA-UNION) dynamic ensemble selection algorithm. This algorithm can optimally integrate three machine learning models on a grid cell scale, namely Random Forests, Extra-Trees, and Extreme Gradient Boosting. We applied our developed freeze/thaw dynamic ensemble selection retrieval algorithm (FT-DESA) to retrieve China’s daily surface FT states from 2009 to 2020 based on multiband Special Sensor Microwave Imager/Sounder (SSMIS) brightness temperatures. We then evaluated our FT-DESA results by comparing the observations of 2398 stations and three other existing FT algorithms, including the modified seasonal threshold algorithm (MSTA), decision tree algorithm (DTA), and dual-index algorithm (DIA) across China. Our results show that FT-DESA has the highest retrieval accuracy and the lowest biases across China among the four algorithms. The mean classification accuracy for the PM and AM overpasses of FT-DESA is 89% and 84%, respectively. The evaluations further indicate that some of the existing algorithms do not reflect the temporal and spatial heterogeneity in selecting thresholds for FT classification. This study demonstrates that the freeze/thaw dynamic ensemble selection algorithm can provide daily estimates of surface FT states across China, improve FT states’ retrieval accuracy, and provide a valuable multi-decadal record for daily FT states.