A GIS-based Spatial Prediction of Landslide Hazard Zones and Mapping in an Eastern Himalayan Hilly Region Using Large Scale Soil Mapping and Analytical Hierarchy Process

Srinivasan, R.; Vasu, Duraisamy; Suputhra, S. Amar; Hegde, Rajendra; Ray, S.K.; Singh, Sanjay

doi:10.1007/s12524-022-01579-8

Cited by 2 publications

(1 citation statement)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Detailed maps based on high-resolution sampling can act as stepping stones for many conservations policies at the national and international levels. Various attempts have been made to construct soil maps using digital and conventional methods at the global level (Arrouays et al, 2020; Grunwald, 2009; Hengl et al, 2014, 2015; Hengl, Leenaars, et al, 2017; Poggio et al, 2021) national-level (Purushothaman et al, 2022; Reddy et al, 2021) and at the local level (Arora et al, 2021; Dharumarajan et al, 2020, 2021, 2022; Kalambukattu et al, 2018; Kaushal et al, 2021; Santra et al, 2017; Srinivasan et al, 2022). However, most of the maps are at low resolution levels.…”

Section: Introductionmentioning

confidence: 99%

Characterization and classification of fine-resolution soil profile for precision agriculture using random forest and self-organizing map

Elias,

Sharma,

Goel

2024

Preprint

View full text Add to dashboard Cite

The availability of high throughput soil profile information is an important component in precision agriculture to perform efficient soil management for sustainable production. We collected 14 soil physiochemical features from Nagpur, Pune, and Haveri, representing target environments of safflower cultivation, and also from our experiment station at Delhi, at fine resolution and created graphical maps to depict the variability. Additionally, we evaluated the predictive ability of two statistical learning models, random forest (RF) and self-organizing maps (SOM) against multinomial regression models for correctly classifying the soil profile. Clustering was performed around the medoids produced from the dissimilarity matrices of these models using partitioning around medoids (PAM) model. The robustness, versatility, and predictive ability of models in correctly classifying the soil profile to clusters were then tested using cross-validation which was repeated 100 times. This study was performed using training data with proportionate size varying from 60 to 95%, and increasing the unit area of observation up to nine times (or decreasing the total number of observations up to a ninth). RF model was found to be the best performing with average prediction accuracy above 85% in all settings which reached close to 100% in some settings. The predictive ability of all the models was maintained even when only the most influencing six variables were used for classification. The optimal training population size for prediction was found to be 70 to 80%. Based on our study, it is recommended to i) collect fine resolution edaphic features from a marginal farm before crop season, ii) use RF or SOM model to identify the most influencing features distinguishing the soil samples iii) expand the area of sample collection, find values for the most influencing features, and use RF model to correctly predict the class to which the new set of the soil belongs to.

show abstract