Machine learning and deep learning for classifying source samples in the sediment source fingerprinting research in Central Asia

Despite the high capabilities of data-based smart models for classification and clustering purposes, to the best of our knowledge, no previous research has explored their application in sediment tracking. In this respect, here, we explore the abilities of three classic machine learning (ML) models (e.g., k-means, k-means++ and k-medoids) and eight deep learning (DL) models (e.g., auto-encoder based data clustering (AEC), deep clustering network (DCN), deep embedded clustering (DEC), improved deep embedded clustering (IDEC), dip-based deep embedded clustering with k-estimation (DipDECK), DipEncoder, deep k-means (DKM) and N2D) for the clustering and classifying of 132 potential source samples for loess deposits in the Ili Basin, Central Asia. A stepwise discriminant function analysis was applied to calculate the source samples classification accuracy, based on different classification schemes presented by ML and DL models. The main goal is comparing the performance of the ML and DL models with the conventional source samples classification technique. The reclassification rate of the source samples classification for the different models was as follows: DipDECK and DKM > DEC, k-means and k-means++ > AEC, IDEC, N2D, and k-medoids > DCN > DipEncoder. Overall, all classic ML and DL models exhibited an accuracy above 90% (apart from DipEncoder with an accuracy of 89.4%), while all these models outperformed the conventional tests (DFA-based reclassification rate: 84%) used in the sediment source tracking studies. Two DL models (DipDECK and DKM) were selected as the most accurate for source samples classification (reclassification rate: 99.2%), opening a new era in sedimentation studies.