Integrated Ensemble Learning, Feature Selection, and Hyperparameter Optimization for Accurate Mapping of Typical Vegetation on Landslides in the Upper Reaches of the Yellow River Using Gaofen-2 Imagery

Abstract

Vegetation plays a crucial role in terrestrial ecosystems, especially in its influence on shallow landslide occurrences. However, in landslide-prone regions, such as the Xiazangtan area in the upper reaches of the Yellow River, sparse and fragmented vegetation presents classification challenges. This study explores vegetation composition utilizing very-high-resolution Gaofen-2 satellite imagery and three feature selection (FS) methods: ReliefF, recursive feature elimination, and logistic regression, optimized with tree-structured Parzen estimator. Five machine learning algorithms were evaluated, including random forest (RF), support vector machine (SVM), extreme gradient boosting, light gradient boosting machine (LightGBM), and category boosting. The results show that the integration of FS with ensemble learning consistently achieves an overall accuracy of over 0.8. The ReliefF-LightGBM combination outperforms others, showcasing superior efficiency. Notably, the boosting algorithm in ensemble learning combined with FS outperforms traditional SVM and RF. This study demonstrates the effectiveness of combining hyperparameter optimization, FS methods, and the boosting algorithm of ensemble learning for vegetation classification purposes. This approach provides a reliable means of accurately identifying and detecting vegetation in the monitoring area, thereby offering an effective strategy for mitigating shallow landslides in the surrounding regions.

Keywords

• Ensemble learning
• feature selection (FS)
• Gaofen-2
• hyperparameter optimization
• Xiazangtan landslide