Scholarlink

<p>In precision agriculture, the delineation of management zone (MZ) fundamentally relies on identifying spatial variabilities across agricultural fields and applying resources differentially. The key to precision agricultural strategies lies in the integration of heterogeneous data sets. This is crucial for establishing MZs, which are essentially uniform agricultural blocks distinguished by unique soil spatial characteristics, though some variability probably remains within each block. Current research lacks the analysis of the feasibility of using the Rasch model and neural networks for data dimensionality reduction in MZ delineation, as well as the comparative evaluation of different clustering model composites. This study analyzed two field-scale agricultural regions in China and evaluated the effectiveness of eight different clustering model composites in enhancing MZ delineation techniques using a unified optimal interpolation strategy. Each composite combined a dimensionality reduction technique, such as the Rasch model, Principal Component Analysis (PCA), AutoEncoder (AE), or Variational AutoEncoder (VAE), with a clustering algorithm (K-means or Gaussian Mixture Models). This study is the first to employ neural network-based AE and VAE for cluster analysis to delineate MZs. The results revealed that Ordinary Kriging interpolation method performed relatively well in interpolation (compared with inverse distance weight, empirical Bayesian Kriging and radial basis function). The result of Silhouette Coefficient scores and variance reduction coefficient indicated that the division into two MZs was the most effective. Utilizing the Rasch model for dimension reduction, the soil attributes of the two study areas were classified into three rating scales, and it was found that organic matter (OM) has the greatest impact on soil fertility. Compared with traditional PCA methods, when the number of MZ is 2, AE (Silhouette Coefficients: 0.57 and 0.59) and VAE (0.73 and 0.64) exhibited superior clustering performance. However, the integration of the Rasch model into clustering composites led to increased fragmentation in the management zoning (MZ) delineation maps. Furthermore, in specific research areas, clustering composites that employed Gaussian Mixture Models consistently showed lower evaluation metrics than those using K-Means. It is recommended to use multi-source data as input for MZ delineation to fully leverage the advantages of dimensionality reduction based on neural networks. Furthermore, these clustering model composites should be flexibly deployed to adapt to diverse data environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing management zone delineation through advanced dimensionality reduction models and clustering algorithms

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Yuefan Wang,
  • Yifan Yuan,
  • Fei Yuan,
  • Xiaojun Liu,
  • Yongchao Tian,
  • Yan Zhu,
  • Weixing Cao,
  • Qiang Cao

摘要

In precision agriculture, the delineation of management zone (MZ) fundamentally relies on identifying spatial variabilities across agricultural fields and applying resources differentially. The key to precision agricultural strategies lies in the integration of heterogeneous data sets. This is crucial for establishing MZs, which are essentially uniform agricultural blocks distinguished by unique soil spatial characteristics, though some variability probably remains within each block. Current research lacks the analysis of the feasibility of using the Rasch model and neural networks for data dimensionality reduction in MZ delineation, as well as the comparative evaluation of different clustering model composites. This study analyzed two field-scale agricultural regions in China and evaluated the effectiveness of eight different clustering model composites in enhancing MZ delineation techniques using a unified optimal interpolation strategy. Each composite combined a dimensionality reduction technique, such as the Rasch model, Principal Component Analysis (PCA), AutoEncoder (AE), or Variational AutoEncoder (VAE), with a clustering algorithm (K-means or Gaussian Mixture Models). This study is the first to employ neural network-based AE and VAE for cluster analysis to delineate MZs. The results revealed that Ordinary Kriging interpolation method performed relatively well in interpolation (compared with inverse distance weight, empirical Bayesian Kriging and radial basis function). The result of Silhouette Coefficient scores and variance reduction coefficient indicated that the division into two MZs was the most effective. Utilizing the Rasch model for dimension reduction, the soil attributes of the two study areas were classified into three rating scales, and it was found that organic matter (OM) has the greatest impact on soil fertility. Compared with traditional PCA methods, when the number of MZ is 2, AE (Silhouette Coefficients: 0.57 and 0.59) and VAE (0.73 and 0.64) exhibited superior clustering performance. However, the integration of the Rasch model into clustering composites led to increased fragmentation in the management zoning (MZ) delineation maps. Furthermore, in specific research areas, clustering composites that employed Gaussian Mixture Models consistently showed lower evaluation metrics than those using K-Means. It is recommended to use multi-source data as input for MZ delineation to fully leverage the advantages of dimensionality reduction based on neural networks. Furthermore, these clustering model composites should be flexibly deployed to adapt to diverse data environments.