Multi-variety monitoring of potato late blight severity using UAV data with improved SMOTE-CS for small sample modeling and deep feature learning

文献类型： 外文期刊

第一作者： Sun, Heguang

作者： Sun, Heguang;Mai, Huanming;Deng, Xiaoling;Feng, Ziheng;Feng, Haikuan;Yang, Guijun;Song, XiaoYu;Mao, Yanzhi;Li, Qingquan;Guo, Mei;Guo, Wei

作者机构：

关键词： Potato late blight; Remote sensing; SMOTE-CS; Deep learning; Transfer learning

期刊名称：EUROPEAN JOURNAL OF AGRONOMY （ 影响因子：5.5； 五年影响因子：5.9 ）

ISSN： 1161-0301

年卷期： 2025 年 169 卷

页码：

收录情况： SCI

摘要： Accurate and non-destructive monitoring of potato late blight (PLB) using unmanned aerial vehicle (UAV) remote sensing data is of great significance for field management. However, during disease outbreaks, there is a lack of universally applicable rapid monitoring models. On the one hand, different varieties exhibit varying levels of resistance and disease monitoring progression, which can be attributed to genetic and environmental factors. On the other hand, the heterogeneity, imbalance, and noise in spectral and textural data across regions pose significant challenges for disease monitoring. To address these issues, this study first improves upon the noise problem in Synthetic Minority Over-sampling Technique (SMOTE) by employing an enhanced feature selection algorithm based on the Feature Selection with Compactness and Separability (FS-CS) principle. Subsequently, the feature ranking is then used with the Importance-Ordered Weighted Averaging (IOWA) operator to calculate the induced Minkowski OWA distance (IMOWAD), replacing the nearest neighbor distance used in SMOTE. This refinement emphasizes the boundaries of synthetic sample regions and mitigates noise-related issues. This improved method is referred to as SMOTE-CS. Secondly, nine models were constructed to evaluate the effectiveness of FS-CS in feature selection when integrating multiple datasets. Compared to mRmR and ReliefF, FS-CS achieved higher accuracy with a smaller number of features. Finally, to address varietal and environmental differences, modeling was conducted using a shallow transfer learning 1D-CNN model and a deep DRSN model incorporating nonlinear soft thresholding processing, respectively. The results indicate that the 1D-CNN model achieved overall accuracies (OA) of 0.99 and 0.93 on the two datasets, respectively. However, its performance was affected by the poor interpretability of the boundary between the synthetic source and target domain samples. The integration of nonlinear soft-thresholding into the DRSN model enhanced its feature extraction capability and noise suppression. It demonstrated strong performance on the two datasets, achieving an OA of 0.91 and a Kappa coefficient of 0.86. Compared to the original SMOTE version, the proposed approach exhibited superior generalization ability. The results of this study provide new insights into the problems of small sample imbalance, noise, and technical support for multi-species PLB monitoring in different regions.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序