基于近红外光谱的土壤营养指标多层次定量检测及其模型稳定性分析

Soil is an important part of agricultural sustainable development, concentrate delivers the agricultural important elements and many ecological environment problems. Quantitative determination of soil nutrients, successively improvement of the soil fertility for crop planting, is the key to promote the development of precise agriculture. In this project, we propose the multi-level indicator quantitative technology for soil nutrition quality based on near-infrared spectrometry, to establish stable spectroscopic analytical models for multi-component simultaneous quantitativeness, to achieve the simple, rapid determination of the soil contents (e.g. organic matter, nitrogen, phosphorus, potassium and etc.). Actually, in-depth study of several chemometric methods will be performed for quantitative analysis. Several core comprehensive indicators for model evaluation are constructed at the multiple levels of single-, duel-, or multi-components. The integrated modes for the combined optimization of waveband selection and data pretreatment are investigated, aiming to establish near-infrared analytical models with an appreciate prediction accuracy. Furthermore, the modeling parameters should be re-analyzed and circle-optimized by considering model stability, so that the models can be used to determine the standard range of near-infrared predictive values of each component for soil nutrition contents. Near-infrared spectrometry with multi-level indicator quantitativeness is expected to become a new technology for rapid detection of soil nutritional quality. Compared with the independent determination of single indicator, the technology of simultaneous quantitativeness of multi-level indicators is in more accordance with the actual demand. It is expected to have broad application prospects.

土壤是农业可持续发展的重要组成部分，是农业重要要素和诸多生态环境问题的集中体现者。通过对土壤营养成分的定量检测，进而改善作物种植的土壤肥力，是促进精密农业发展的关键。本项目基于近红外光谱，研究土壤营养指标多层次定量分析方法，建立直接快速、稳定的光谱定标模型，实现对土壤营养成分（有机质、氮素、磷、钾等）的多成分同时测定。深入研究若干计量学方法，通过在成分独立、双成分组合和多成分融合的多个层次上构建综合模型指标，确定光谱预处理与信息子波段联合筛选的综合优化模式，建立具有较高预测精度的近红外分析模型；进一步结合模型稳定性分析进行模型参数优选，从而确定土壤营养指标的近红外多层次预测范围，使近红外光谱的指标多层次分析成为一种能够快速检测土壤营养的新方法。与单一指标独立检测相比，指标多层次同时定量分析更符合实际需求，具有广阔的应用前景。

土壤是农业可持续发展中最重要的组成部分，是现代农业环境的重要要素和诸多生态问题的集中体现者。实现对土壤成分的实时快速定量检测是发展精准农业的必要手段。本项目研究了基于近红外（NIR）光谱的土壤营养指标多层次定量检测的计量学方法，建立直接快速的NIR光谱定标检测模型，实现对土壤各营养成分（有机质、总氮等）的指标多层次定量测定，并分析了模型稳定性。一方面，研究了Savitziky-Golay滤波、Whittaker平滑、竞争自适应加权抽样（CARS）、移动窗口（Moving Window）等技术，确定光谱预处理与信息子波段联合筛选的综合优化模式，实现数据降噪和特征提取；另一方面，研究了若干线性和非线性的光谱建模计量学方法，如偏最小二乘（PLS）、支持向量机（SVM）、随机森林（RF）、BP神经网络（BPNN）等，实现了多重计量学方法组合的步进网格式参数优化，并提出了基于网络深度学习模式的特征变量筛选方法，从各成分独立、双成分组合和多成分融合的多个层次上对土壤或以土壤为载体的农作物的营养成分指标实现NIR定量分析，建立具有较高预测精度的、多层次适用的NIR定标模型；进一步利用实验重复性和建模不确定性理论进行NIR定标模型的稳定性分析，找到了待测营养成分的近红外多层次定量预测的标定范围。使近红外光谱指标多层次定量分析技术成为能够快速检测土壤营养成分的农业信息化手段，为土壤质量在线监控提供技术参考。.本项目取得比较丰富的研究成果。在国内外学术期刊上发表论文13篇（其中SCI、EI收录10篇）；参加全国性和国际性会议并发表会议论文1篇；申请国家发明专利3项；获得软件著作权1项；累计培养硕士研究生7名。