关键词: 样本优化, 滑坡易发性, 空间相关性, 地理环境相似性, 非线性合成过采样, 多粒度级联森林

Abstract:

In the domain of landslide susceptibility evaluation, the prevalent issue of sample category imbalance often skews evaluation outcomes in favor of the majority class, thereby undermining the accuracy of landslide forecasts. Sample optimization emerges as a pivotal solution to mitigate these biases. Traditional approaches to sample optimization predominantly concentrate on distinguishing the characteristic disparities between positive and negative samples within the feature space, overlooking the critical aspects of geographical disparities among samples and the intricate nonlinear interrelations between characteristic factors and landslide occurrences. Such oversight may result in a biased and oversimplified representation of sample characteristics. Addressing this gap, the present study introduces an advanced landslide susceptibility evaluation methodology that integrates sample optimization with a nuanced consideration of spatial and feature dynamics. Initially, the method employs an undersampling strategy based on geographical environment similarity criteria that incorporate spatial correlation. Subsequently, it innovates a nonlinear synthetic oversampling technique to augment sample diversity. The analytical process culminates in the application of a multi-grained cascade forest model for the prediction of landslide susceptibility. Focusing on Yibin city as the empirical case study, the efficacy of the proposed method is rigorously validated through statistical metrics, encompassing both model precision verification and susceptibility zoning analysis. Comparative evaluation against nine established methodologies reveals that our proposed framework consistently surpasses in predictive accuracy across varied scenarios of positive sample deficiency, thereby offering susceptibility zones that exhibit a higher degree of concordance with the real-world spatial distribution of landslide incidents.

Key words: sample optimization, landslide susceptibility, spatial correlation, geographical environment similarity, nonlinear synthetic oversampling, multi-grained cascade forest