Effects of different sources and spatial resolutions of environmental covariates on predicting soil organic carbon using machine learning in a semi-arid region of Iran

Garosi, Younes; Ayoubi, Shamsollah; Nussbaum, Madlene; Sheklabadi, Mohsen (2022). Effects of different sources and spatial resolutions of environmental covariates on predicting soil organic carbon using machine learning in a semi-arid region of Iran Geoderma Regional, 29, e00513. Elsevier 10.1016/j.geodrs.2022.e00513

Text 1-s2.0-S2352009422000335-main.pdf - Published Version Restricted to registered users only Available under License Publisher holds Copyright. Download (5MB) | Request a copy

The main goal of this study was to consider and compare the effects of different spatial resolutions of covariates from different sources on predicting SOC in a semi-arid region located in the west of Iran. For this purpose, 67 topsoil samples (0–30 cm) with the measured SOC contents were used as the dependent variable. The covariates controlling the SOC content from different sources were provided in two scenarios. For the first scenario (scenario I), six covariate sets with spatial resolution ranging from 2 to 30 m, and original and aggregated pixel sizes were prepared using the digital elevation models (DEMs) and remote sensing data to predict SOC. In the second scenario (scenario II), the available legacy data, including geology, land use and soil texture maps, were prepared with compatible spatial resolution and added to each covariate set provided for scenario I. After feature selection analysis, the modelling processes were performed using two machine learning models, namely, Random Forest (RF) and Support Vector Machine (SVM). The results of performance analysis, as obtained by leave one out cross validation (LOOCV), showed that the RF and covariate set B (with 10 m spatial resolution) in scenario I, with R2 = 0.21, CCC = 0.41, MAE = 0.26 and RMSE = 0.34%, and also, in scenario II, with R2 = 0.32, CCC = 0.51, MAE = 0.24, and RMSE = 0.32%, had a better performance in predicting SOC. In addition, the remote sensing data were identified as the most important variables controlling the spatial distribution of SOC. Finally, by using the RF model as the superior model, the SOC map provided by the covariate set B in scenario II, which was the combination of the three types of covariates (DEM, remote sensing data and legacy data), was shown to have the lowest uncertainty in comparison to the SOC provided by the covariate set B in scenario I. In general, our results showed that the model type, source, resolution and the combination of these variables could greatly influence the prediction outputs. In fact, the SOC map provided with the combination of parsimonious variables at the optimal pixel size could help decision-making in environmental resources management.

Item Type:	Journal Article (Original Article)
Division/Institute:	School of Agricultural, Forest and Food Sciences HAFL School of Agricultural, Forest and Food Sciences HAFL > Agriculture School of Agricultural, Forest and Food Sciences HAFL > Agriculture > Soils and Geoinformation
Name:	Garosi, Younes; Ayoubi, Shamsollah; Nussbaum, Madlene0000-0002-6808-8956 and Sheklabadi, Mohsen
Subjects:	G Geography. Anthropology. Recreation > GA Mathematical geography. Cartography S Agriculture > S Agriculture (General)
ISSN:	23520094
Publisher:	Elsevier
Language:	English
Submitter:	Madlene Nussbaum
Date Deposited:	02 Nov 2022 09:47
Last Modified:	02 Nov 2022 09:47
Publisher DOI:	10.1016/j.geodrs.2022.e00513
Uncontrolled Keywords:	Covariate, Pixel size, Different data sources, Entisols, Inceptisols, Uncertainty
ARBOR DOI:	10.24451/arbor.17883
URI:	https://arbor.bfh.ch/id/eprint/17883