数字农科院2.0

A Feature-Optimized and Performance-Weighted Ensemble Learning for Estimating Soil Salinity Using UAV Imagery and Soil Auxiliary Information

文献类型： 外文期刊

作者： Wang, Li;Yang, Jing;Wu, Shangrong;Xia, Lang;Lu, Miao;Li, Wenjuan;Wu, Wenbin;Zha, Yan;Yang, Peng

作者机构：

期刊名称： JOURNAL OF REMOTE SENSING

ISSN：

年卷期： 2026 年 6 卷

页码：

收录情况： EI（2025版） ; ; 北大核心（2023版） ; ; CSCD（2025-2026年度） ; ; 科技核心（2024版） ; ; 农林核心（2024版）ESCI（2025版）

摘要： Reliable estimation and mapping of soil salinity content (SSC) offer crucial information for sustainable agricultural development and food safety. Ensemble learning utilizing unmanned aerial vehicle (UAV) imagery has proven to be effective in estimating soil properties. Nevertheless, due to the high dimensionality of data and the variability in performance among different base learners within ensemble learning, effective methods that can identify sensitive features and optimize base learners remain to be explored. Furthermore, the mechanism by which soil auxiliary information, such as soil organic matter (SOM) and pH, influences the accuracy of SSC estimation is not yet fully understood. In this study, we proposed a novel ensemble learning framework, named feature-optimized and performance-weighted ensemble learning (FOPW-EL), for estimating SSC using UAV imagery and soil auxiliary information. This framework incorporated the hybrid embedded feature selection (HEFS) to extract sensitive features and the CRITIC (Criteria Importance Through Intercriteria Correlation)-Entropy combined weighting (CECW) to identify the optimal base learners. Our results demonstrate that the FOPW-EL model substantially outperformed all other models, with average validation metrics of the coefficient of determination (R2) of 0.752, ratio of prediction to deviation (RPD) of 1.744, and linear correlation consistency coefficient (LCCC) of 0.792. Additionally, the FOPW-EL model exhibited low prediction uncertainty, with the global uncertainty index consistently below 0.39. We identified SOM, salinity index (SI2), normalized difference salinity red edge index (NDSIREG), and pH as the most sensitive feature variables. The feature importance assessment indicates that spectral indices played a key role in SSC modeling, accounting for 58.94% of the contribution. The analysis of partial dependence plots revealed a strong positive correlation between SI2 and SSC, with a correlation coefficient of 0.76. In contrast, SOM and pH values showed significant negative correlations with SSC, with correlation coefficients of -0.42 and -0.22, respectively. These findings indicate that this proposed method holds considerable promise for providing accurate, cost-effective, and efficient SSC estimation, which could greatly aid in soil quality monitoring.

分类号：

  • 相关文献
作者其他论文 更多>>