Hybrid landslide susceptibility modelling for rainfall-triggered landslides in forested slopes: a review for bridging physics-based and machine learning methods

Abstract

Background: Managing the risk of shallow landslides is important for any land use in steep terrain. This is especially true for plantation forestry in New Zealand, where construction of forest roads or clear-fell harvesting can exacerbate both frequency and scale of rainfall-induced shallow landslides. In forestry, landslide susceptibility mapping has typically been completed using heuristic field-based methods. Geospatial empirical and statistical models, mainly based on past storm events have predictive performance at scale but are limited by their lack of physical grounding. Conversely, deterministic physics-based slope stability methods, while interpretable, require detailed geotechnical data that are not available at operational scales. Recent advances in hybrid models offer a pathway to combine the mechanistic transparency of physics-based methods with the flexibility of data-driven geospatially based algorithms.

Methods: This review evaluates recent research on hybrid landslide susceptibility modelling for rainfall-triggered shallow landslides, with a particular focus on applications relevant to plantation forestry. Using a structured literature search across major databases from 2010 to 2024, 94 peer-reviewed studies were used to inform the development of six recurring categories of hybrid integration: physically derived features, physics-guided loss or constraint functions, physics-inspired or interpretable network architectures, physics-guided sample selection, physics-driven metaheuristic optimisation, and joint susceptibility-intensity modelling. The characteristics, strengths, opportunities and limitations of each category were synthesised based on the reported model structure, type of physical integration, interpretability, scalability, and relevance to operational forestry applications.

Results: Hybrid models demonstrate a strong potential to enhance slope-stability assessment in forestry operations by reducing data demands, improving interpretability, and enabling physically consistent, large-area mapping. The unique nature, strengths, opportunities and limitations of each hybrid model category are outlined and compared.

Conclusions: The findings highlight the opportunity to transition towards interpretable neural-network modelling approaches that learn relationships between environmental variables while embedding physically meaningful geotechnical and hydrological processes within the model structure. Successful development would support forestry land managers identify localised shallow landslide risk to support the operational planning.