KMS XINJIANG INSTITUTE OF ECOLOGY AND GEOGRAPHY，CAS
| 基于高光谱和热红外光谱信息的干旱区土壤水盐反演 | |
| 徐璐 | |
| Subtype | 博士 |
| Thesis Advisor | 王权 ; 李兰海 |
| 2016 | |
| Degree Grantor | 中国科学院大学 |
| Place of Conferral | 北京 |
| Degree Discipline | 地图学与地理信息系统 |
| Keyword | 土壤含水量 土壤含盐量 可见光-近红外光谱 热红外光谱 土壤颜色 偏最小二乘法 光谱指数法 模型评价标准 |
| Abstract | 盐碱土对生态环境和农业生产是一个全球性的威胁。本文利用干旱区盐化荒漠土设计了不同土壤含水量的实验，并利用古尔班通古特沙漠边缘的风沙土设计了不同盐土类型（NaCl、Na2SO4、Na2CO3）和不同土壤含盐量的实验。实验中同步获取了土壤水盐信息和光谱信息（包括可见-近红外光谱和热红外光谱），分析了土壤水盐对光谱的影响，并对土壤水盐进行光谱建模反演。本文采用了偏最小二乘回归（PLSR）和光谱指数两种方法进行建模，并用校正的赤池信息量准则（AICc）和建模标定值标准差与拟合值标准差的比值（相对分析误差，RPD）进行模型评价，基于上述研究得到以下主要结论： （1）由于西北干旱区的土壤盐分以Na2SO4为主，在设计土壤盐分反演实验时应当考虑土柱高度，并且高度在大于7cm时能够较好的模拟野外实际情况。不同盐分类型和盐分含量实验发现，对光谱影响的主要因素是盐分类型而不是盐分含量。 （2）利用高光谱对土壤含水量进行PLSR建模时发现，1370nm和1955nm是对土壤水分含量最为敏感的波段，最终确定的土壤水分反演模型含有6个波段，模型精度为RPD=1.51，R2=0.69。对不同盐分类型的盐分含量反演建模时发现，只有Na2CO3盐分模型中含有可见光波段，且模型精度为RPD=3.49，R2=0.92，其他盐分类型的模型中都仅含近红外波段，均有较好的精度。光谱指数模型对土壤水分反演效果不能达到预测目的，对土壤含盐量可以达到预测目的，但是都不如PLSR模型精度高。 （3）利用热红外光谱对土壤含水量进行PLSR建模时发现，热红外光谱可以较好的预测土壤含水量，且8.596μm和8.769μm对土壤水分信息更为敏感，最优模型包含7个波段，模型精度为RPD=1.58，R2=0.714。对不同盐分类型的盐分含量反演建模时发现，只有Na2SO4选择原始光谱进行土壤盐分模型，敏感波段遍布9-14μm，而其他盐分类型模型都选择导数光谱，且敏感波段分布在8-9μm，都能较好的建立土壤盐分预测模型。光谱指数模型对土壤水盐含量都不能达到预测目的，结果都不如PLSR模型精度高，这可能是由于PLSR模型可以选择多个对土壤属性信息敏感的波段，而光谱指数法只限于一个或两个波段。 （4）野外采样数据对模型验证的结果精度较差，进一步利用野外数据进行PLSR建模发现，可见光波段对模型有很大的贡献，而近红外波段的贡献很小，模型精度为RPD=2.06，R2=0.81。光谱指数法得到的最优模型也达到了预测精度，但是不如PLSR模型精度高。野外数据建模的主要贡献波段是可见光波段，说明土壤颜色是野外土壤光谱的主要影响因素。 （5）利用能够直观反映土壤颜色的照片反演土壤含盐量，通过提取照片中红绿蓝和灰度4种颜色成分，计算每个颜色成分中的各亮度值像元数占总像元数的比例。将亮度值分区可以更充分地利用各亮度值占比，最后确定合适的亮度值分区数为9，并基于此建立土壤含盐量模型，模型精度可达R2=0.9。 |
| Other Abstract | Saline-alkali soil is a world wide threat for ecological environment and agriculture. This thesis designed one experiment of various soil water content with saline desert soil, and another experiment of various soil salt types (NaCl, Na2SO4, and Na2CO3) and contents with aeolian soil. The information of soil water and salt conent and soil spectra including visible-near infrared and thermal infrared spectra were obtained synchronously. Analyzing spectral response to the changes of soil water and salt content, models were built for soil water and salt estimation. The methods of partial least squares regression (PLSR) and spectral index were used for modeling, and the correction of Akaike Information Criterion (AICc) and Relative Percent Deviation (RPD) were used for model evalution. Based on the above studies, the results were concluded below: 1. Duing to salt type of Na2SO4 occupied a main place in the saline soil Northwest China, indoor simulation experiment need considering the height of soil column, and we proved that soil column higher than 7cm was good enough for field situation simulation. The second experiment showed that the main influence on spectra was not salt content but salt type. 2. Modeling soil water content with the method of PLSR based on hyperspectra, we found that the most sensitive bands were 1370nm and 1955nm, and the optimal model contained 6 bands for soil water estimation with RPD of 1.51 and R2 of 0.69. When modeling soil salt content, only the model for salt type of Na2CO3 contained the visible band, and with a RPD of 3.49 and R2 of 0.92. Bands involved in models for other salt types all located at near infrared range, and all models were high accurate. The method of spectral index was not good enough for soil water estimation, but good enough for soil salt estimation, nonetheless the accuracy was inferior to PLSR. 3. Modeling soil water content with the method of PLSR based on thermal infrared spectra, we found that the most sensitive bands were 8.596μm and 8.769μm, and the optimal model contained 7 bands for soil water estimation with a RPD of 1.58 and R2 of 0.71. When modeling soil salt content, only the model for salt type of Na2SO4 chose raw spectra to estimate soil water content, and the sensitive bands spread over 9-14μm. Models for other salt types chose derivative spectra and the sensitive bands centralized in 8-9μm, and all models could predict soil salt content well. The method of spectral index did not work for soil water and salt content estimation, indicating that PLSR worked better than spectral index. This might result from the method of PLSR had the ability to choose more bands than spectral index. 4. The result of model validation with field data was not good enough. To tell the difference between models built with experiment data and field data, we conducted modeling with field data and found that visible spectra contributed much more than infrared spectra, and the model achieved the accuracy with a RPD of 2.06 and R2 of 0.81. The model identified with method of spectral index could also have the ability to perdict soil salt content, but the accuracy was not as good as PLSR. The dominant bands focused on the visible range, indicating that soil color was the main factor influencing soil spectra. 5. Soil photoes reflecting soil color were used to estimate soil salt content. Four color components (RGB and Gray) were drawn from the photoes, and the percentage of each brightness in each color component was calculated. Brightness value partition could take full advantage of the percentages, and the final partition number was identified as nine. Based on these processes, soil salt content was modeled with a R2 of 0.9. |
| Subject Area | 地图学与地理信息系统 |
| Language | 中文 |
| Document Type | 学位论文 |
| Identifier | http://ir.xjlas.org/handle/365004/14710 |
| Collection | 研究系统_荒漠环境研究室 |
| Affiliation | 中科院新疆生态与地理研究所 |
| Recommended Citation GB/T 7714 | 徐璐. 基于高光谱和热红外光谱信息的干旱区土壤水盐反演[D]. 北京. 中国科学院大学,2016. |
| Files in This Item: | There are no files associated with this item. | |||||
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.
Edit Comment