Landslides are one of the most destructive natural hazards; they can drastically alter landscape morphology, destroy man-made structures, and endanger people's life. Landslide susceptibility maps (LSMs), which show the spatial likelihood of landslide occurrence, are crucial for environmental management, urban planning, and minimizing economic losses. To date, the majority of research into data mining LSM uses small-scale case studies focusing on a single type of landslide. This paper presents a data mining approach to producing LSM for a large, heterogeneous region that is susceptible to multiple types of landslides. Using a case study of Piedmont, Italy, a Random Forest algorithm is applied to produce both susceptibility maps and classification maps. These maps are combined to give a highly accurate (over 85% classification accuracy) LSM which contains a large amount of information and is easy to interpret. This novel method of mapping landslide susceptibility demonstrates the efficacy of Random Forest to produce highly accurate susceptibility maps for a large heterogeneous region without the need for multiple susceptibility assessments.

No abstract

Soil bulk density (D_b) is a major contributor to uncertainties in landscape-scale carbon and nutrient stock estimation. However, it is time consuming to measure and is, therefore, frequently predicted using surrogate variables, such as soil texture. Using this approach is of limited value for estimating landscape-scale inventories, as its accuracy beyond the sampling point at which texture is measured becomes highly uncertain. In this paper, we explore the ability of soil landscape models to predict soil D_b using a suite of landscape attributes and derivatives for both topsoil and subsoil. The models were constructed using random forests and artificial neural networks.

Using these statistical methods, we have produced a spatially distributed prediction of D_b on a 100 m × 100 m grid, which was shown to significantly improve topsoil carbon stock estimation. In comparison to using mean values from point measurements, stratified by soil class, we found that the gridded method predicted D_b more accurately, especially for higher and lower values within the range. Within our study area of the Midlands, UK, we found that the gridded prediction of D_b produced a stock inventory of over 1 million tonnes of carbon greater than the stratified mean method. Furthermore, the 95% confidence interval associated with total C stock prediction was almost halved by using the gridded method. The gridded approach was particularly useful in improving organic carbon (OC) stock estimation for fine-scale landscape units at which many landscape–atmosphere interaction models operate

This paper investigates the use of expert knowledge as a resource for digital soil mapping. To do this, three models of topsoil soil bulk density (D b ) were produced: (i) a random forest model formulated and cross-validated with the limited data available (which served as the benchmark), (ii) a naïve Bayesian network (BN) where the conditional probabilities that define the relations between D b and explanatory landscape variables were derived from expert knowledge rather than data and (iii) a 'hierarchical' BN where model structure was also defined by expert knowledge. These models were used to generate spatial predictions for mapping topsoil D b at a landscape scale. The results show that expert knowledge-based models can identify the same spatial trends in soil properties at a landscape scale as state-of-the-art mapping algorithms. This means that they are a viable option for soil mapping applications in areas that have limited empirical data.

Soil bulk density (<i>D</i>_b) is a major contributor to uncertainties in landscape-scale carbon and nutrient stock estimation. However, it is time consuming to measure and is, therefore, frequently predicted using surrogate variables, such as soil texture. Using this approach is of limited value for estimating landscape scale inventories, as its accuracy beyond the sampling point at which texture is measured becomes highly uncertain. In this paper, we explore the ability of soil landscape models to predict soil <i>D</i>_b using a suite of landscape attributes and derivatives for both topsoil and subsoil. The models were constructed using random forests and artificial neural networks. <br><br> Using these statistical methods, we have produced a spatially distributed prediction of <i>D</i>_b on a 100m × 100m grid which was shown to significantly improve topsoil carbon stock estimation. In comparison to using mean values from point measurements, the error associated with predictions was over three times lower using the gridded prediction. Within our study area of the Midlands, UK, we found that the gridded prediction of <i>D</i>_b produced a stock inventory of nearly 8 million tonnes of carbon less than the mean method. Furthermore, the gridded approach was particularly useful in improving organic carbon (OC) stock estimation for fine-scale landscape units at which many landscape-atmosphere interaction models operate