ESTIMATION OF PHYSICAL AND CHEMICAL SOIL PROPERTIES BY ARTIFICIAL NEURAL NETWORKS

ABSTRACT Soil physical and chemical analyses are relatively high-cost and time-consuming procedures. In the search for alternatives to predict these properties from a reduced number of soil samples, the use of Artificial Neural Networks (ANN) has been pointed out as a great computational technique to solve this problem by means of experience. This tool also has the ability to acquire knowledge and then apply it. This study aimed at using ANNs to estimate the physical and chemical properties of soil. The data came from the physical and chemical analysis of 120 sampling points, which were submitted to descriptive analysis, geostatistical analysis, and ANNs training and analysis. In the geostatistical analysis, the semivariogram model that best fitted the experimental variogram was verified for each soil property, and the ordinary kriging was used as an interpolation method. The ANNs were trained and selected based on their assertiveness in the mapping of considered standards, and then used to estimate all soil properties. The mean errors of ordinary kriging estimates were compared to those of ANNs and then compared to the original values using Student's t-Test. The results showed that the ANN had an assertiveness compatible with ordinary kriging. Therefore, such technique is a promising tool to estimate soil properties using a reduced number of soil samples.

摘要：土壤理化分析通常成本高昂且耗时较长。为探索通过少量土壤样本即可预测土壤属性的替代方案，人工神经网络（Artificial Neural Networks，ANN）被认为是一种极具潜力的经验驱动型计算技术，可用于解决该问题。该工具还具备知识获取与应用的能力。本研究旨在利用人工神经网络估算土壤理化属性。研究数据源自120个采样点的理化分析结果，所有样本均经过描述性分析、地统计分析以及人工神经网络的训练与分析。在地统计分析环节，针对每一项土壤属性，筛选出与实验变异函数拟合度最优的半变异函数模型，并采用普通克里金法作为插值方法。人工神经网络以其在目标指标映射中的预测准确性为依据完成训练与筛选，随后被用于全部土壤属性的估算。通过学生t检验（Student's t-Test），将普通克里金法估算结果的平均误差与人工神经网络的平均误差进行对比，并将二者与原始实测值进行比对。结果表明，人工神经网络的预测精度与普通克里金法相当。因此，该技术是一种极具应用前景的方法，可通过少量土壤样本实现土壤属性的估算。