基于激光诱导击穿技术的中国典型农田土壤的原子光谱特征及其应用

The mass information about soil fertility, varied in time and in space, is required for modern precision agriculture. Although the conventional chemical analysis method has played a major role in the soil fertility evaluation, but it seems difficult to meet the needs of modern agricultural production due to too large input, and the technique of atomic spectroscopy provides an alternative option for the fast acquisition of soil fertility. Laser induced breakdown technology will be involved in this study, and the light source will be optimized by high power laser and dual beam. Then the typical farmland soils with temporal and spatial variation from all over China will be collected for the atomic spectra recording. The factors affecting the determination of soil atomic spectra, including sample preparation method, spectral range, resolution, scanning sites and laser ablation depth, will be investigated and optimized, and the determination method of soil atomic spectra will be built in the first time. The obtained soil atomic spectra will be treated by Gauss function and data conversion (wavelet analysis and deconvolution), so as to reduce the interference of environmental noise and sample matrix, and obtain representative soil atomic spectra with high ratio of signal to noise. Based on above results, typical farmland soils will be characterized by soil atomic spectra, integrated utilization of mathematical tools, i.e. principal component analysis, partial least squares regression and artificial neural network, will be made for spectral modeling, and the target information about soil fertility will be separated; then the attempts will be initiated for elucidating the mechanism of the formation and evolution of soil fertility from the level of atomic composition, which will provide theoretical support for fast soil fertility evaluation and soil digitalization.

现代精准农业要求采集随时空变化而变化的海量土壤肥力信息，虽然传统化学分析法发挥了重大作用，但由于投入大而难以满足现代农业生产的需求，而原子光谱法则为土壤肥力信息的快速获取提供了可能手段。本研究将利用激光诱导击穿技术，优化配置高功率和双光束激光光源，以我国不同时空分布的典型农田土壤样本为材料，获得土壤原子光谱，分析测定的影响因素 (样本制备方法、波段范围、分辨率、扫描位点和激光烧蚀深度等)，率先构建土壤原子光谱的测定方法；然后利用高斯函数和数据转换 (小波分析和反卷积)，对土壤原子光谱进行前处理，以减少系统环境干扰和样本基体效应，获取代表性的高信噪比土壤原子光谱；进而分析我国典型农田土壤的原子光谱特征，综合利用主成分分析、偏最小二乘法回归和人工神经网络等数学工具进行光谱建模，提取目标土壤肥力信息，并从原子组成的角度解析土壤肥力形成和演变特征，为土壤肥力评估提及土壤数字化提供理论支撑。

传统的土壤理化分析方法具有成本高、操作繁琐、破坏性强等局限性，难以满足精准农业海量的信息需求。现代光谱技术具有简单、快速、无损检测、数据量大等优点，能够作为精准农业中土壤信息快速获取的重要手段。激光诱导击穿光谱作为一种原子光谱技术，其设备简单、测试成本低廉、可实现无损检测，且可以原位表征土壤的元素组成信息。本项目研究了LIBS仪器栅极宽度、扫描次数、扫描深度、土壤相对含水率、土壤粒径以及土壤紧实度对土壤LIBS光谱质量的影响，并优化了这些参数。结果表明，土壤LIBS光谱质量随栅极宽度的增加先增加后降低；不同类型土壤LIBS光谱质量随扫描次数的变化不尽相同，结合欧式距离、主成分分析（PCA）、偏最小二乘判别分析（PLS-DA）、支持向量机（SVM）、随机森林（RF）、概率神经网络（PNN）模型，实现了土壤类型鉴定。构建了土壤属性与土壤LIBS光谱的偏最小二乘回归（PLSR）模型，评估模型对多种土壤属性的同时预测能力。结果表明，土壤LIBS全谱（200~1000 nm）对土壤pH、阳离子交换量（CEC）、SOM、TN、全钾（TK）有卓越；使用土壤LIBS特征发射峰建模提高了SOM的预测精度，降低了TN、TP、TK的预测精度；土壤LIBS全谱和特征发射峰对AP和AK含量的预测能力均较差。采用变量投影重要性法（VIP）、无用信息变量消除（UVE）、竞争性自适应重加权抽样（CARS）、随机蛙跳（RFG）四种变量筛选算法对土壤LIBS光谱波段进行筛选。结果表明，四种算法均能提高SOM的预测精度，其中CARS算法不仅能较好地消除SOM的预测偏差，还能克服校正集过拟合问题，更适用于SOM的预测。采用PCA和红绿蓝（RGB）三原色方法将土壤LIBS光谱应用于土壤空间变异分析和土壤制图。结果表明，RGB三原色法能保留土壤LIBS光谱中更多的信息，从而提高土壤空间变异的表达能力；构建了土壤LIBS光谱与土壤综合肥力之间的回归模型和分类模型，评估并比较回归模型和分类模型对土壤综合肥力的预测能力。