Authors

Abstract

Soil nitrogen content plays a critical role in fruit tree growth and fruit quality. To achieve high-precision estimation of the three-dimensional spatial distribution of soil nitrogen in peach orchards, this study proposes a hybrid model integrating a Backpropagation (BP) neural network with Regression Kriging (RK). The model is designed to overcome the limitations of traditional single-method approaches when addressing complex spatial nonlinear problems. Using soil data from peach orchards in Fujian Province as the research subject, the predictive performance of five models—BP neural network, Ordinary Kriging (OK), Regression Kriging (RK), Co-Kriging (CK), and the proposed RK-BP hybrid model—was systematically compared. The results indicate that the RK-BP hybrid model outperforms all baseline models, achieving the highest coefficient of determination (R² = 0.97) and the lowest root mean square error (RMSE = 7.0). Compared with the BPNN and RK models, the proposed model improved prediction accuracy by 74.17% and 34.95%, and enhanced precision by 77.29% and 77.46%, respectively. The RK-BP model successfully integrates nonlinear relationship modeling with spatial structural analysis, achieving complementary advantages. This study confirms the effectiveness and great potential of hybrid modeling frameworks combining machine learning and geostatistics for 3D digital soil mapping, offering significant theoretical value and promising prospects for precision agriculture.

Abstract in Chinese

土壤中的氮含量对果树的生长和果实品质起着至关重要的作用。为了实现对桃园中土壤氮元素的三维空间分布高精度估计，本研究提出了一种将反向传播（BP）神经网络与回归克里金（RK）方法相结合的混合模型。该模型旨在克服传统单一方法在处理复杂空间非线性问题时的局限性。以福建省桃园的土壤数据作为研究对象，对五个模型——BP 神经网络、普通克里金（OK）、回归克里金（RK）、协同克里金（CK）以及所提出的 RK-BP 混合模型——的预测性能进行了系统比较。结果表明，RK-BP 混合模型优于所有基准模型，实现了最高的决定系数（R² = 0.97）和最低的均方根误差（RMSE=7.0）。与 BPNN和 RK模型相比，所提出的模型将预测准确度提高了74.17%和34.95%，精度分别提高了 77.29% 和 77.46%。RK-BP 模型成功地将非线性关系建模与空间结构分析相结合，实现了互补优势。本研究证实了将机器学习和地质统计学相结合的混合建模框架在三维数字土壤制图中的有效性和巨大潜力，为精准农业提供了重要的理论价值和广阔的发展前景。