Feature Engineering of Geohazard Susceptibility Analysis Based on the Random Forest Algorithm: Taking Tianshui City, Gansu Province, as an Example

In this paper, Feature Engineering (FE) was applied to Landslide Susceptibility Mapping (LSM), while the most suitable conditioning feature dataset and analysis method were tested and analyzed. Tianshui city was taken as the study area, three types of geohazard (collapse, landslide, and unstable slo...

Full description

Saved in:
Bibliographic Details
Published in:Remote sensing (Basel, Switzerland) Switzerland), 2022-11, Vol.14 (22), p.5658
Main Authors: Ling, Xiao, Zhu, Yueqin, Ming, Dongping, Chen, Yangyang, Zhang, Liang, Du, Tongyao
Format: Article
Language:English
Subjects:
Accuracy
Algorithms
Classification
Conditioning
feature engineering
Feature selection
Geological hazards
landslide
landslide susceptibility mapping
Landslides
Maximum entropy method
Neural networks
Optimization
random forest algorithm
Regression analysis
Rivers
Slope stability
Slopes
Standardization
Statistical analysis
Statistical methods
Support vector machines
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this paper, Feature Engineering (FE) was applied to Landslide Susceptibility Mapping (LSM), while the most suitable conditioning feature dataset and analysis method were tested and analyzed. Tianshui city was taken as the study area, three types of geohazard (collapse, landslide, and unstable slopes) were used, while a total of twenty-three conditioning features were generated; two dimensionless methods (normalization and standardization) were tested afterward. Four Random-Forest-based (RF-based) feature selection methods using different indicators (Gini Impurity, GI; Out of Bag Accuracy, OOBA) were proposed and tested separately. The LSMs of four models were carried out under the guidance results of FE, namely Classification and Regression Tree (CART), Random Forest (RF), Logistic Regression (LR), and Support Vector Machine for Classification (SVC). For feature enhancement, standardization had significant advantages over normalization. All RF-based methods were proven effective, lifting the AUC by 0.01~0.02. The RF model achieved the highest LSM accuracies, respectively, 0.949 (landslide), 0.957, and 0.949 (unstable slopes), improved by 0.008 (landslide), 0.005 (collapse), and 0.013 (unstable slopes). This proved that the FE helped to improve LSM and can help to decide the dominant conditioning factors for regional geohazards.
ISSN:2072-4292
2072-4292
DOI:10.3390/rs14225658