Bias Correction of GLDAS ‐Derived Daily Minimum Soil Temperature ( DMST ) in Shallow and Deeper Soil Profiles Using Supervised Machine Learning Algorithm

This study investigates the accuracy of GLDAS‐Noah‐2.1 daily minimum soil temperature (DMST) at 0–10 cm (shallow) and 40–100 cm (deeper) depths in Khorasan Razavi, Iran, using 5 years (2008–2012) of data from 13 synoptic stations. Initial evaluations using statistical metrics revealed significant discrepancies, with GLDAS tending to overestimate DMST. The initial raw GLDAS data showed a considerable systematic error, with a bias ranging from 1.96°C to 6.0°C in the shallow profile and from 0.58°C to 6.62°C in the deeper soil profile. On average, the shallow layer performed poorly, yielding an RMSE of 4.82°C and an average bias of 3.85°C, while the deeper layer showed an average RMSE of 2.72°C and a bias of 2.14°C. To mitigate these biases, a K‐nearest neighbors (KNN) supervised machine learning algorithm was employed and optimized through grid search. The KNN model dramatically enhanced performance for both layers. For the shallow depth, the average RMSE was reduced to 2.56°C (a ~47% reduction), and the average bias was reduced to 0.06°C. For the deeper layer, the average RMSE was reduced to 1.30°C (a ~52% reduction), and the average bias was reduced to 0.00°C. Furthermore, the Nash–Sutcliffe efficiency (NSE) improved from an initial average of 0.75 and 0.86–0.92 and 0.97 for the shallow and deeper layers, respectively. Post‐correction, the model achieved a “Very Good” performance rating for all stations, with the average percent bias (Pbias) falling to 0.37% (shallow) and −0.08% (deeper). The results underscore the efficacy of machine learning‐based bias correction in enhancing the reliability of GLDAS datasets for regional climate and agricultural applications.