DataSheet1_Estimating and mapping the soil total nitrogen contents in black soil region using hyperspectral images towards environmental heterogeneity.docx

Introduction: Fast and accurate estimation and spatial mapping of soil total nitrogen (TN) content is important for the development of modern precision agriculture, such as soil fertility monitoring and land reclamation decision-making. Hyperspectral remote sensing has been demonstrated to be an accurate real-time technique for rapid estimation and mapping of soil TN content.
Methods: To solve the problem of poor accuracy and generalization of estimation models caused by soil environmental heterogeneity in estimating and mapping soil TN content using hyperspectral images, 502 soil samples were collected from a typical black soil area in Yushu City, Jilin Province, China, as a test area, and three sample grouping strategies were established by soil environmental variables (soil type, thickness of the black soil layer, and topographic factors), and Pearson correlation coefficient and competitive adaptive reweighted sampling algorithm were used to determine the TN characteristic bands of each sample set under different strategies. Based on the data characteristics of the sub-sample set, the local regression estimation model based on sample grouping was constructed using the CatBoost algorithm, and the estimation and distribution mapping of soil TN content was carried out.
Results and Discussion: The results showed that after dividing the samples according to the differences in soil environmental factors, the characteristic information of the samples is more targeted, with more abundant numbers and distribution ranges of TN characteristic bands. Compared to the global regression estimation with all samples, the local regression based on the grouping of soil environment differences showed improved accuracy, with the local regression estimation model constructed with the ST-G strategy exhibiting the highest estimation accuracy (Rp2 = 0.839). The results can provide a reference for large-area soil properties mapping, and technical support for soil quality digitization and precision fertilization.

引言：快速精准估算并空间制图土壤全氮（Total Nitrogen, TN）含量，对现代精准农业的发展具有重要意义，例如土壤肥力监测与土地复垦决策等场景。高光谱遥感（hyperspectral remote sensing）已被证实是一种可用于快速估算并绘制土壤全氮含量分布的精准实时技术。
方法：针对利用高光谱图像估算并绘制土壤全氮含量时，因土壤环境异质性导致估算模型精度与泛化能力欠佳的问题，本研究以中国吉林省榆树市典型黑土区为试验区，采集了502份土壤样本。基于土壤环境变量（土壤类型、黑土层厚度与地形因子）构建了三种样本分组策略，并采用皮尔逊相关系数（Pearson Correlation Coefficient）与竞争自适应重加权采样算法（Competitive Adaptive Reweighted Sampling Algorithm），确定了不同策略下各样本集的全氮特征波段。基于子样本集的数据特征，本研究利用CatBoost算法构建了基于样本分组的局部回归估算模型，并开展了土壤全氮含量的估算与空间分布制图工作。
结果与讨论：研究结果表明，依据土壤环境因子差异对样本进行分组后，样本的特征信息更具针对性，全氮特征波段的数量与分布范围也更为丰富。相较于基于全部样本的全局回归估算，基于土壤环境差异分组的局部回归估算精度得到了提升，其中采用ST-G策略构建的局部回归估算模型估算精度最高（决定系数Rp²=0.839）。本研究结果可为大尺度土壤属性制图提供参考，同时为土壤质量数字化与精准施肥提供技术支撑。