Sampling Method Based on Fuzzy Membership for Computing Negative Sample Credibility and Its Applications
Current sampling methods do not provide effective quantitative assessment mechanisms for evaluating the intrinsic credibility of negative samples. This impedes the systematic quantification of the effect of misselection of geologically predisposed areas (i.e., potential landslide zones) as negative...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-07-01
|
| Series: | Applied Sciences |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2076-3417/15/14/7646 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Current sampling methods do not provide effective quantitative assessment mechanisms for evaluating the intrinsic credibility of negative samples. This impedes the systematic quantification of the effect of misselection of geologically predisposed areas (i.e., potential landslide zones) as negative samples on the accuracy of landslide susceptibility evaluation models. To overcome this challenge, this study proposes a fuzzy membership-based sampling method for assessing negative sample credibility in the Liangshan Yi Autonomous Prefecture, where credibility is defined as the confidence level of stable nonlandslide samples. Subsequently, negative samples were sampled across stratified credibility thresholds to construct a frequency ratio–random forest coupled model. The influence of negative sample credibility on model performance was then systematically evaluated using various metrics, including the F1-score (metrics for evaluating classification performance), area under the receiver operating characteristic curve (AUC), and actual landslide distribution ratio (landslide proportion) in high-susceptibility zones. The results are as follows: (1) Increasing the credibility threshold progressively improves model precision while inducing systematic overestimation bias in regional susceptibility assessment; (2) Integrated analysis of model performance and landslide distribution characteristics (where recall, F1-score, and AUC values initially increase then decrease) confirms the optimal effectiveness when selecting negative samples within a credibility threshold range of 0.7–1.0. This study innovatively achieves quantitative optimization of negative samples and provides a universal solution for improving the performance of diverse models reliant on negative sampling strategies. |
|---|---|
| ISSN: | 2076-3417 |