Transferability of predictive models to map susceptibility of ephemeral gullies at large scale

Hamid Mohebzadeh, Asim Biswas, Ben DeVries, Ramesh Rudra and Prasad Daggupati ()
Additional contact information
Hamid Mohebzadeh: University of Guelph
Asim Biswas: University of Guelph
Ben DeVries: University of Guelph
Ramesh Rudra: University of Guelph
Prasad Daggupati: University of Guelph

Natural Hazards: Journal of the International Society for the Prevention and Mitigation of Natural Hazards, 2024, vol. 120, issue 5, No 21, 4527-4561

Abstract: Abstract Ephemeral gully erosion is one of the main sources of soil loss from agricultural landscapes. Various tools including predictive models with machine learning (ML) algorithms have shown promise to identify susceptible areas. However, ML models have two limitations: (1) a trained model in one area may not be applicable in another area and (2) their application for susceptibility mapping of ephemeral gullies at large-scale areas presents a challenge due to the small size of these features and the need for digitization of all gullies for accurate susceptibility mapping. To overcome these limitations, a novel approach was introduced in the current study for comprehensive validation of ML models and prepare a susceptibility map of ephemeral gullies using an areal transfer of calibration–validation relations. Five ML models were evaluated in Northern Lake Erie Basin as a large-scale region. First, the region was divided into three zones based on the most effective factors of gully formation, and a total of eight watersheds were selected in Zone 1 and Zone 2 (hereafter study area). Zone 3 was not considered, because no gullies were observed in this zone. All the ML models were compared using a new validation approach, including local (trained and validated in the same area) and transferred (trained in one area and tested in other areas). Results showed that random forest (RF) was the most accurate local model in both Zone 1 (accuracy = 0.8833, AUC = 0.8830, sensitivity = 0.9239, and specificity = 0.8537) and Zone 2 (accuracy = 0.8606, AUC = 0.8608, sensitivity = 0.8987, and specificity = 0.8381), while gradient boosting decision tree (GBDT) was the most accurate transferred model (accuracy = 0.7298, AUC = 0.7297, and sensitivity = 0.7826). From the results of the current study, it can be concluded that (1) zonation technique supports the prediction of ephemeral gullies by dividing the study area into the small zones that carry similar topographical and morphological characteristics and (2) the local-transferred validation technique is a helpful method for finding the ML model that can be trained in a small watershed and scale up to the larger area without further calibration.

Keywords: Ephemeral gully; Susceptibility mapping; Machine learning; Spatial modeling (search for similar items in EconPapers)
Date: 2024
References: View references in EconPapers View complete reference list from CitEc
Citations:

Downloads: (external link)
http://link.springer.com/10.1007/s11069-023-06377-0 Abstract (text/html)
Access to the full text of the articles in this series is restricted.

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:spr:nathaz:v:120:y:2024:i:5:d:10.1007_s11069-023-06377-0

Ordering information: This journal article can be ordered from
http://www.springer.com/economics/journal/11069

DOI: 10.1007/s11069-023-06377-0

Access Statistics for this article

Natural Hazards: Journal of the International Society for the Prevention and Mitigation of Natural Hazards is currently edited by Thomas Glade, Tad S. Murty and Vladimír Schenk

More articles in Natural Hazards: Journal of the International Society for the Prevention and Mitigation of Natural Hazards from Springer, International Society for the Prevention and Mitigation of Natural Hazards
Bibliographic data for series maintained by Sonal Shukla () and Springer Nature Abstracting and Indexing ().