Rill detachment capacity (D-c) is a key factor of the overall erosion process on steep and long hillslopes of deforested areas. Accurate predictions of this factor are essential for land managers, in order to adopt the most effective soil conservation techniques against erosion. However, no previous studies have evaluated how much machine learning algorithms (such as Random Forest, RF) and multiple regression models (such as Multiple Linear Regression, MLR, and Partial Least Square Regression, PLS-R) are accurate to predict D-c in forest and deforested sites and with or without application of treatments for soil conservation. This study evaluates the D-c prediction capacity of RF, MLR or PLS-R algorithms in forestlands of Northern Iran. The models have been applied to 250 observations of the changes in D-c between sites covered by natural tree species or subjected to reforestation and five soil treatments, and deforested sites can be predicted using. The results of model applications showed that: (i) in natural and planted forests, both MLR and PLS-R gave accurate predictions of changes in D-c (coefficient of efficiency of Nash and Sutcliffe, NSE, over 0.80); (ii) in the same sites, the RF algorithm was much less accurate when applied to natural forests (NSE = 0.46), and its performance was poor (NSE = - 0.51) in reforested sites; (iii) in deforested and treated sites, the D-c predictions by both MLR and PLS-R were poor (NSE < 0), while the RF algorithm was more accurate (NSE = 0.49), but not its prediction capacity was not optimal. Since both MLR and PLS-R are accurate in predicting rill erosion in natural and planted forests, the use of these models may be suggested to landscape managers, in order to estimate the changes in rill erodibility due to deforestation or reforestation works. Moreover, modelers may adopt the parameters of the linear regression equations setup in this study for predictions of rill erosion in deforested areas with similar characteristics as the experimental sites. The low reliability of all tested models for D-c predictions in sites deforested and treated with soil conservation techniques suggests more research, for instance testing the accuracy of Artificial Neural Networks or multiple regression algorithms with different mathematical structure under those conditions.

Exploring the accuracy of Random Forest and Multiple Regression models to predict rill detachment in soils under different plant species and soil treatments in deforested lands / Parhizkar, Misagh; Lucas-Borja, Manuel Esteban; Zema, Demetrio Antonio. - In: MODELING EARTH SYSTEMS AND ENVIRONMENT. - ISSN 2363-6203. - 10:2(2024), pp. 2533-2546. [10.1007/s40808-023-01919-8]

Exploring the accuracy of Random Forest and Multiple Regression models to predict rill detachment in soils under different plant species and soil treatments in deforested lands

Zema, Demetrio Antonio
2024-01-01

Abstract

Rill detachment capacity (D-c) is a key factor of the overall erosion process on steep and long hillslopes of deforested areas. Accurate predictions of this factor are essential for land managers, in order to adopt the most effective soil conservation techniques against erosion. However, no previous studies have evaluated how much machine learning algorithms (such as Random Forest, RF) and multiple regression models (such as Multiple Linear Regression, MLR, and Partial Least Square Regression, PLS-R) are accurate to predict D-c in forest and deforested sites and with or without application of treatments for soil conservation. This study evaluates the D-c prediction capacity of RF, MLR or PLS-R algorithms in forestlands of Northern Iran. The models have been applied to 250 observations of the changes in D-c between sites covered by natural tree species or subjected to reforestation and five soil treatments, and deforested sites can be predicted using. The results of model applications showed that: (i) in natural and planted forests, both MLR and PLS-R gave accurate predictions of changes in D-c (coefficient of efficiency of Nash and Sutcliffe, NSE, over 0.80); (ii) in the same sites, the RF algorithm was much less accurate when applied to natural forests (NSE = 0.46), and its performance was poor (NSE = - 0.51) in reforested sites; (iii) in deforested and treated sites, the D-c predictions by both MLR and PLS-R were poor (NSE < 0), while the RF algorithm was more accurate (NSE = 0.49), but not its prediction capacity was not optimal. Since both MLR and PLS-R are accurate in predicting rill erosion in natural and planted forests, the use of these models may be suggested to landscape managers, in order to estimate the changes in rill erodibility due to deforestation or reforestation works. Moreover, modelers may adopt the parameters of the linear regression equations setup in this study for predictions of rill erosion in deforested areas with similar characteristics as the experimental sites. The low reliability of all tested models for D-c predictions in sites deforested and treated with soil conservation techniques suggests more research, for instance testing the accuracy of Artificial Neural Networks or multiple regression algorithms with different mathematical structure under those conditions.
2024
Soil hydrology
Log response ratio
Soil properties
Hydrological modelling
File in questo prodotto:
File Dimensione Formato  
Parhizkar_2024_MESE_Exploring_editor.pdf

non disponibili

Descrizione: Versione editoriale
Tipologia: Versione Editoriale (PDF)
Licenza: Copyright dell'editore
Dimensione 1.24 MB
Formato Adobe PDF
1.24 MB Adobe PDF   Visualizza/Apri   Richiedi una copia
Parhizkar_2024_MESE_Exploring_post-print.pdf

embargo fino al 28/12/2024

Descrizione: Documento in post-print
Tipologia: Documento in Post-print
Licenza: Tutti i diritti riservati (All rights reserved)
Dimensione 1.2 MB
Formato Adobe PDF
1.2 MB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12318/151507
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 1
  • ???jsp.display-item.citation.isi??? 1
social impact