Toward Developing a Generalizable Pedotransfer Function for Saturated Hydraulic Conductivity Using Transfer Learning and Predictor Selector Algorithm
Overview
Research
Identity
Additional Document Info
View All
Overview
abstract
AbstractPedotransfer functions (PTFs) are being instrumental in saturated hydraulic conductivity (Ks) estimation. Despite various advancements, the performance of existing generic Ks predicting PTFs need augmentation. This study developed a robust Ks predicting PTF using a machine learning (ML) algorithm and exhaustive data set for 324 soils with 28 properties sampled over a tropical savanna region of India. Four ML algorithms were evaluated for this purpose, and random forest (RF) outperformed all others. A substantial improvement to the prediction by RFbased PTF was achieved through predictor selection using a hybrid wrapperembedded algorithm. The predictor selection algorithm selected eight pertinent predictors (HIDS): S, Si, C, FSF, Cu, GMD, D60, and D10. The mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), and NashSutcliffe efficiency (NSE) obtained as average tenfold crossvalidation scores for RF algorithm training with HIDS were 0.87, 1.47, 0.94, and 0.94, respectively. The developed PTF (RFHIDS) was evaluated alongside the recently published PTFs by Araya and Ghezzehei (2019, https://doi.org/10.1029/2018wr024357), within and outside the study region. In that process, it was observed that the RFHIDS possessed superior prediction proficiency compared to the recently published and commonly used PTFs in both cases. These findings mark RFHIDS as the most robust generalizable PTF, which may further be evaluated in different parts of the world. Moreover, looking at the performance of the eight selected predictors within and outside the study region, they can be considered for experiment design globally to make Ks estimation accurate and costeffective.
