基于深度学习的土壤近红外光谱定量分析模型研究

This proposal innovatively combines traditional spectral analysis and deep learning algorithm, and aims to overcome the bottleneck problems of BP neutral network in the construction of nonlinear prediction model using soil’s near-infrared spectroscopy (NIRS) such as the limitations in training parameters and over-fitting induced by huge amounts of samples. The research contents in this proposal include: the exploration of the spatial-temporal context representation method of the soil’s near-infrared spectra, the simplification of the matrix model of the soil’s NIR spectra based on compressed sensing, the analysis and prediction of soil’s properties in combination of deep learning and the construction of soil’s NIRS prediction and analysis system based on deep learning. Finally, a novel deep-learning-based quantitative analysis method with the use of soil’s NIR spectra is proposed. .Using the self-developed Vehicle-mounted Soil Nutrient and Heavy Metal Rapid Detection Device and the collected a large amount of soil standard samples, the NIRS prediction model for soil’s major fertility parameters such as organic matter, total nitrogen, pH value and moisture was constructed based on deep learning. Owing the application of deep learning, the model’s prediction accuracy and robustness will be effectively enhanced, and the model’s time-series prediction ability for a same plot while the model’s spatial generation ability for different plots will also be significantly improved. Moreover, the feasibility of the construction of qualitative and quantitative models of soil’s other properties will be explored in this proposal. The research results in this proposal can provide the technical supports for the rapid, on-the-site and quantitative determination with the use of soil’s NIR spectra and finally achieve a large-scale engineering application.

传统光谱分析与深度学习算法交叉创新，突破BP神经网络土壤近红外光谱非线性分析预测建模训练参数受限与海量样本过拟合的瓶颈问题，围绕土壤近红外光谱的时空上下文表示方法、基于压缩感知的土壤近红外光谱大数据约减方法、基于深度学习的土壤近红外光谱分析预测模型、基于深度学习的土壤近红外光谱分析系统构建等四个方面，建立一套基于深度学习的土壤近红外光谱定量分析方法。.利用项目团队研制完成的“车载土壤养分与重金属快速检测装置”以及项目合作积累的大量土壤测试标样，针对砂姜黑土有机质、总氮、pH值、含水量等土壤综合肥力主要参数，建成基于深度学习的土壤光谱分析预测模型，有效提高模型预测精度与鲁棒性，显著提升同一地块模型时序预测与不同地块空间泛化能力；同时探索基于深度学习方法构建土壤其它成分定性与定量分析模型的可行性；为利用近红外光谱开展土壤快速、原位、定量检测，实现大规模工程应用提供技术支撑。

农业部的测土配方施肥、国土资源部复垦土地综合评定等工作均需要开展大量的土壤养分与重金属检测工作。然而，目前广泛使用的化学检测方法存在检测时间长、费用高、检测密度低等问题。近红外光谱分析法有望实现直接、快速、无污染、低成本的土壤养分定量分析。.本项目的主要研究集中在传统光谱分析与深度学习算法交叉创新，试图突破BP神经网络在土壤近红外光谱非线性分析预测建模训练参数受限与海量样本过拟合的瓶颈问题，以土壤中有机质、总氮、pH值、含水量等近红外光谱分析预测模型的构建及其预测精度与鲁棒性提升为总体目标，实现以下突破：.一、土壤近红外光谱维数约简方法。研究了大尺度范围基于土壤近红外光谱的养分预测问题中，线性建模方法、波长优选等方法的预测能力，分析了线性方法的预测能力及局限性。.二、基于压缩感知的土壤近红外光谱特征提取。提出了稀疏表示的土壤近红外光谱分析预测、基于梅尔域和卷积神经网络的土壤近红外光谱分析预测模型等基于压缩感知的定量分析方法。.三、基于深度学习的土壤近红外光谱分析预测。提出了基于改进自动编码器光谱法预测土壤有机质含量等级分类、基于卷积神经网络的土壤有机碳预测模型、基于多任务学习的土壤近红外光谱非线性特征提取、基于Elman神经网络模型的土壤重金属含量预测方法等深度学习分析模型，形成了一套基于深度学习的土壤近红外光谱分析体系。.四、相关领域的应用研究。开展了土壤样本前处理自动化装备试制、太赫兹光谱、毛细管电泳方法检测方法的预研，对近红外光谱检测土壤能力的形成有效补充。.本项目的实施，丰富和发展土壤近红外分析与理解中的基本理论与方法，推动近红外光谱分析技术在未来大规模土壤多类成分的现场、原位、实时、定量检测及相关领域规模化应用也发挥推动作用。