Arid
基于可见-近红外光谱的我国干旱区土壤碳酸钙含量预测
其他题名Predicting Soil Calcium Carbonate Content based on Visible-near Infrared Reflectance Spectroscopy in the Arid Region of China
林卡
出版年2018
学位类型硕士
导师李德成
学位授予单位中国科学院大学
中文摘要土壤信息是精准农业、土壤污染治理和生态环境模拟等研究和应用的基础。土壤信息获取的传统途径为室内化学分析测试,但该方法周期长、成本高,并且有可能污染环境,难以满足现代社会对快速获取土壤属性信息及保护生态环境的需求。光谱分析技术具有速度快、成本低和实时性的优势,目前已经成为土壤信息获取的重要手段之一。然而,当前的土壤光谱建模与预测研究多集中在局部小区域范围或特定土壤类型的土壤样本。由于土壤景观类型的多样性,土壤光谱模型的推广性或普适性往往比较差,严重限制了光谱分析技术的应用潜力在土壤学上的发挥,制约了准确土壤信息的快速廉价获取。如何基于土壤光谱准确预测任意局部地区样本的土壤性质,是一个具有挑战性的问题。鉴于此,本文从影响土壤光谱预测的主要因素(样本覆盖度、样本信噪比和建模样本等)出发,提出了一种建立区域土壤光谱库与构建建模样本集相结合的思路与方法,利用可见-近红外光谱(350~2500 nm)对任意局部地区样本的土壤性质进行预测研究。以土壤碳酸钙含量为目标土壤属性,以包括新疆、甘肃、青海和内蒙古的西北干旱区为研究区,开展了以下三个方面的研究工作:(1)建立西北干旱区土壤碳酸钙光谱库;(2) 探究碳酸钙含量与光谱反演效果关系:利用黑河流域土壤样本数据库,根据不同指标将黑河流域干旱土数据集划分为不同碳酸钙含量子集,采用PLSR算法对土壤碳酸钙含量进行反演,对比不同子集的建模结果,以此探究光谱反演土壤碳酸钙含量的效果与碳酸钙含量的定量关系;(3)研究建模样本集的选取方法:对同地数据集与异地数据集建模预测结果进行比较,然后根据光谱分类和光谱匹配的方法在异地数据集中挑选建模样本集,建模预测并进行精度对比,以获得基于碳酸钙-可见近红外光谱数据库的最佳建模样本集构建方法和土壤碳酸钙含量的最佳预测结果。结果发现: (1)随着碳酸钙含量升高,干旱土的可见-近红外波段整体的光谱反射率也逐渐升高,该趋势在近红外波段更加明显。逐波段对土壤光谱反射率和碳酸钙含量进行相关性分析,光谱反射率与土壤碳酸钙含量成正相关。 (2)利用可见近红外光谱反演土壤碳酸钙,其反演效果与碳酸钙含量关系不大。因此在利用光谱预测土壤碳酸钙含量时不需对不同含量的碳酸钙样本进行划分。(3)同地数据集预测土壤碳酸钙含量的精度远胜于异地数据集。同地数据集的预测R2=0.86,RPD=1.96,异地数据集的预测R2=0.72,RPD=1.41。(4)基于光谱库光谱反射特性的光谱分类策略,极易恶化碳酸钙含量的变异和分布,导致局部模型精度的降低(R2=0.67,RPD=1.27),其预测精度不如全局模型。(5)基于待测样本光谱特性的光谱匹配可以较好地预测土壤碳酸钙含量,其中以基于吸收谷特征参数的欧氏距离匹配建模预测效果最好(R2=0.74,RPD=1.46)。(6)在异地数据集中随机添加同地数据可以有效提高光谱预测土壤碳酸钙的精度,吸收谷特征参数匹配结合同地数据校正,R2>0.74, RPD≥1.6。综上,本文提出的基于待测样本光谱特性的光谱匹配技术与同地数据相结合的思路方法,能够有效地利用区域土壤光谱数据库,对土壤碳酸钙含量的预测取得了较高的预测精度,实现了利用土壤光谱准确预测任意局部地区样本的碳酸钙含量。这有助于充分发挥光谱分析技术在土壤学中的应用潜力,同时可以推动准确土壤信息的快速廉价获取,提高对土壤、农业和生态等研究和应用的服务能力。
英文摘要Soil information is the basis for research and application of precision agriculture, soil pollution control and ecological environment simulation. Current methods of soil information acquisition are mainly based on traditionally indoor physical and chemical analysis. But these methods are costly, laborous, time-consuming, and easy to cause environment contamination. Instead, spectral analysis has the advantages of high analytical speed, low cost and real-time. Thus it is a potential and promising method for soil information acquisition. However, most soil spectral analysis and prediction studies are either limited to small-scale areas or based on a specific soil type. And comsequently the soil spectral models are highly dependent on the model-building area. Due to the diversity of soil-landscape types, the extrapolation of these models are often rather poor. This, to a big extent, restricts the applications of spectral analysis in soil science. Therefore, it is a challenging issue to accurately predict the soil properties of samples from any given local area using soil spectrum.As far as soil samples are considered, their coverage, signal-to-noise ratio and calibrarion set selection are main influencing factors for soil spectral prediction. Taking care of these factors would improve the extrapolation performance of soil spectral models. Based on this idea, we propose in this paper a new approach to predicting soil properties of the samples from any given local area using soil Vis-NIR spectroscopy (350~2500 nm) through combining regional soil spectral library with model-building sample set. Taking soil calcium carbonate content as a target property, we choose northwestern arid region as our study areas, including Xinjiang, Gansu, Qinghai, and Inner Mongolia, and the following three aspects of research work have been carried out: (1)Establishment of a Vis-NIR spectral library in the arid area of northwestern China.(2)Relationship between calcium carbonate content and spectral inversion effect: Using the soil sample data collected in the Heihe River Basin, according to different indicators, the aridisol data set was divided into different calcium carbonate content subsets, and the soil calcium carbonate content is predicted using the PLSR model. The performances of different subsets were compared with each other to explore the quantitative relationship between the effect of spectral inversion and the content of calcium carbonate.(3)Selection methods of model-building sample: Comparing the prediction results of the local data sets and non-local data sets, and then selecting model samples based on non-local data sets according to spectral classification and matching. Comparing the prediction accuracy to identify the best modeling sample selection method and the best prediction result based on Vis-NIR spectral library. This thesis gets the following conclusions: (1)With the increase of calcium carbonate content, the spectral reflectance of visible near-infrared bands gradually increases, and this trend is more obvious in the near-infrared bands. Correlation of soil spectral reflectance and calcium carbonate content were analyzed on a band-by-band basis. Spectral reflectance was positively correlated with soil calcium carbonate content.(2)Calcium carbonate can enhance spectral reflectance of visible near-infrared bands, but the effect is not so significantly reflected in the prediction of soil calcium carbonate content using the Vis-NIR spectral reflectance. Therefore, it seems unnecessary to divide soil samples by content of soil calcium carbonate when using spectra to predict calcium carbonate contents. (3)The prediction accuracy of soil calcium carbonate content of local datasets is much better than non-local datasets. The R2 and RPD of the local data set were 0.86 and 1.96, while those of non-local data set were 0.72 and 1.41.(4)Spectral classification methods based on the spectral reflectance characteristics of the spectral library can easily deteriorate the variation and distribution of calcium carbonate content, resulting in a decrease in the accuracy of the local model (R2 = 0.67, RPD = 1.27), and its prediction accuracy is not as good as the global model.(5) Spectral matching based on the spectral characteristics of the test samples can predict the soil calcium carbonate content more better, and the Euclidean distance matching based on the absorption valley characteristic parameters predicts the best (R2=0.74, RPD=1.46).(6) Randomly adding the local correction data to non-local matching sets can effectively improve the prediction accuracy. Absorption valley characteristic parameter matching combined with local correction data can achieve the best prediction effect, R2>0.74, RPD≥1.6.In summary, this paper propose a new approach to accurately predict the soil calcium carbonate content of the samples from any given local area using soil Vis-NIR spectroscopy. This approach combines spectral matching technology (based on the spectral characteristics of the test sample) with local data, and thus we can effectively utilize the regional soil spectral library to obtain a high prediction accuracy of soil calcium carbonate content. This work would be helpful for the application of spectral analysis technology in soil science. At the same time, it can promote the rapid and inexpensive acquisition of accurate soil information and improve the service capabilities of soil information, and provide support for the research and application of soil science, agricultural science and ecology.
中文关键词可见-近红外光谱 ; 区域光谱数据库 ; 土壤碳酸钙 ; 土壤光谱建模与预测 ; 西北干旱区
英文关键词Visible near-infrared reflection spectra Regional spectral library Soil calcium carbonate Soil spectral modeling and predicting Northwest arid region
语种中文
国家中国
来源学科分类土壤学
来源机构中国科学院南京土壤研究所
资源类型学位论文
条目标识符http://119.78.100.177/qdio/handle/2XILL650/288103
推荐引用方式
GB/T 7714
林卡. 基于可见-近红外光谱的我国干旱区土壤碳酸钙含量预测[D]. 中国科学院大学,2018.
条目包含的文件
条目无相关文件。
个性服务
推荐该条目
保存到收藏夹
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[林卡]的文章
百度学术
百度学术中相似的文章
[林卡]的文章
必应学术
必应学术中相似的文章
[林卡]的文章
相关权益政策
暂无数据
收藏/分享

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。