Data and code for machine learning analysis of aeolian transport processes

Description

This repository provides representative datasets and analysis codes supporting a study on the application of machine learning methods to aeolian transport processes. The dataset includes example and processed data illustrating the structure, variables, and feature representations used in the analysis, such as wind-related parameters and derived predictors relevant to aeolian transport. These data are representative subsets of larger field measurement datasets. The code includes core scripts for data preprocessing, feature construction, model training, evaluation, and interpretability analysis. The machine learning workflow encompasses supervised classification and regression models, together with explainable machine learning techniques used to assess feature contributions.

Steps to reproduce

All datasets used in this study are provided in Apache Feather (.feather) format and stored in the data directory. Each monitoring site corresponds to a single feather file. The analysis was conducted using Python. Required libraries include numpy, pandas, scipy, matplotlib, seaborn, scikit-learn, xgboost, shap, and statsmodels. All scripts assume relative paths and read the feather files directly from the data directory without any additional preprocessing. The code repository is organized into four main parts: feature engineering, raw feature–based analysis, statistics-based analysis, and fluctuation and turbulence analysis. Scripts in the feature engineering section compute wind and temperature distributions, sand flux statistics, correlation metrics, and atmospheric stability parameters such as the Monin–Obukhov stability parameter (z/L). The remaining scripts implement machine-learning models and statistical analyses corresponding to the results reported in the manuscript. To reproduce the results, users only need to select the desired feather file, adjust the station name if necessary, and modify key parameters such as the time window size or feature combinations at the beginning of each script. All figures and quantitative results in the paper can be regenerated using the provided code without requiring any external data sources.

Institutions

• Shijiazhuang Tiedao University

  Hebei, Shijiazhuang

Categories

Funders

• National Natural Science Foundation of China

  Beijing, China

  Grant ID: 52278511

Licence