Integrated modelling for mapping spatial sources of dust in central Asia - An important dust source in the global atmospheric system

Spatial mapping of dust sources in arid and semi-arid regions is necessary to mitigate on-site and off-site impacts. In this study, we apply a novel integrated modelling approach including leave one feature out (LOFO) – as a technique for feature selection -, deep learning (DL) models (gcForest and...

Full description

Saved in:

Bibliographic Details
Published in:	Atmospheric pollution research Vol. 12; no. 9; p. 101173
Main Authors:	Gholami, Hamid, Mohammadifar, Aliakbar, Malakooti, Hossein, Esmaeilpour, Yahya, Golzari, Shahram, Mohammadi, Fariborz, Li, Yue, Song, Yougui, Kaskaoutis, Dimitris G., Fitzsimmons, Kathryn Elizabeth, Collins, Adrian L.
Format:	Journal Article
Language:	English
Published:	Elsevier B.V 01-09-2021
Subjects:	Central asia Deep learning Dust spatial mapping Effective factors Gaussian copula model Hybrid copula-gcForest model Model sensitivity Deep learning Hybrid copula-gcForest model Dust spatial mapping Model sensitivity Gaussian copula model Effective factors Central asia
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Spatial mapping of dust sources in arid and semi-arid regions is necessary to mitigate on-site and off-site impacts. In this study, we apply a novel integrated modelling approach including leave one feature out (LOFO) – as a technique for feature selection -, deep learning (DL) models (gcForest and bidirectional long short-term memory (Bi-LSTM)), game theory (GT) and a Gaussian copula-based multivariate (GCBM) model for mapping dust sources in Central Asia (CA). Eight factors (precipitation, cation exchange capacity, bulk density, wind speed, slope, silt content, lithology and coarse fragment content) were selected by LOFO as effective for controlling dust emissions, and were used in the novel modelling process. Six statistical indicators were utilized to assess the performance of the two DL models and a hybrid copula-gcForest model, while a sensitivity analysis of the models was also carried out. The hybrid copula-gcForest model was identified as the most accurate, predicting that 16%, 7.1%, 9.5% and 67.4% of the study area is grouped at low, moderate, high and very high susceptibility classes for dust emissions, respectively. Based on permutation feature importance measure (PFIM) and Shapely Additive exPlanations (SHAP), bulk density, precipitation and coarse fragment content were evaluated as the three most important factors with the highest contributions to the predictive model output. The study area suffers from intense wind erosion and the generated spatial maps of dust sources may be helpful for mitigating and controlling dust phenomena in CA. •An new integrated modeling approach used to map dust source spatially.•Leave one feature out (LOFO) applied for the feature selection.•Game theory used to interpretability of the predictive models output.•The hybrid copula-gcForest model was identified as the most accurate.•Bulk density has the highest contribution to the predictive model's output.
ISSN:	1309-1042 1309-1042
DOI:	10.1016/j.apr.2021.101173