|
|
| |
|
|
| This article has been：browse 30times Download 3times |
Scan the code！ |
|
|
| 基于可见光-近红外光谱与机器学习的不同菌草物种叶片光谱特性比较分析 |
|
李声繁1, 吕 师2, 郑人浩1, 林 辉2, 杜泓伟1, 严慧慧2,
周 然1, 王子怡1, 林冬梅2, 林占熺2, 刘凤山2*
|
|
1. 福建农林大学 林学院(碳中和学院), 福州 350002;2. 福建农林大学
国际菌草学院/国家菌草工程技术研究中心, 福州 350002
|
|
| 摘要: |
| 针对不同菌草物种形态相似、传统识别效率低的问题,该研究通过光谱特征分析与机器学习建模,实现对8种菌草的高效识别与分类。采集巨菌草、象草、莱牧1号象草、紫象草、狼尾草、杂交狼尾草、美洲狼尾草、巴新野生蔗8种菌草叶片在400～900 nm波段的光谱反射率数据,结合原始光谱、倒数取对数变换、一阶导数和二阶导数处理,分析光谱特征,并提取多类型植被指数。基于原始光谱、倒数取对数变换、一阶导数、二阶导数、三边参数、植被指数的6类特征集,构建20类梯度化特征组合,采用支持向量机(SVM)与随机森林(RF)算法构建分类模型并评价精度。结果表明:(1)8种菌草在可见光波段呈现典型植被光谱特征,红边区域(700～750 nm)反射率急剧上升。(2)不同光谱处理方法在特定波段(如570～650 nm的倒数取对数变换、730 nm附近的一阶导数、670～760 nm的二阶导数)显著放大了物种间的光谱差异; 红边振幅、红边面积及简单比率(SR)植被指数对物种区分能力最优。(3)模型精度显示,SVM算法整体优于RF,“原始光谱+三边参数”组合精度最高,达70.56%,较RF同组合提升14.03%; RF最优为“原始光谱+倒数取对数变换”组合,但高维特征适配性不足。综上认为,可见光与近红外波段光谱特征结合光谱变换处理及植被指数的构建,辅以支持向量机算法,可为菌草物种快速识别和精准分类提供有效的理论依据和技术支持。 |
| 关键词: 物种识别, 光谱反射率, 特征提取, 支持向量机, 随机森林 |
| DOI：10.11931/guihaia.gxzw202510038 |
| 分类号:Q948 |
| 文章编号:1000-3142(2026)06-1027-19 |
| Fund project:新疆维吾尔自治区重大科技专项项目(2024A030095); 那曲市级科技项目(NQKJ-2024-20); 福建农林大学科技创新专项基金项目(KFB23189A)。 |
|
| Comparative analysis of leaf spectral characteristics of different Juncao species based on visible-near infrared spectroscopy and machine learning |
|
LI Shengfan1, LÜ Shi2, ZHENG Renhao1, LIN Hui2, DU Hongwei1, YAN Huihui2,
ZHOU Ran1, WANG Ziyi1, LIN Dongmei2, LIN Zhanxi2, LIU Fengshan2*
|
|
1. Forestry College( Carbon Neutrality College ), Fujian Agriculture and Forestry University, Fuzhou 350002, China;2. International
College of Juncao Science/National Engineering Research Center of Juncao Technology, Fujian
Agriculture and Forestry University, Fuzhou 350002, China
|
| Abstract: |
| Aiming to address the challenges of morphological similarity among different species of Juncao and the low efficiency of traditional identification methods, this study employs spectral feature analysis and machine learning modeling to achieve efficient identification and classification of eight Juncao species. Spectral reflectance data of leaf samples from eight species, including Cenchrus fungigraminus, Penmisetum purpureum, P. purpureum cv. Laimu-1, P. purpureum cv. Red, P. alopecuroides, P. glaucum × purpureum, P. americanum, and Saccharum officinarum cv. PNG, were collected within the 400-900 nm wavelength range. By processing the original spectra along with their reciprocal taking logarithm transformations, the first derivatives, and the second derivatives, spectral characteristics were analyzed, and multiple vegetation indices were extracted. Based on six types of feature sets — original spectra, reciprocal taking logarithm transformation, the first derivative, the second derivative, three-edge parameters, and vegetation indices — twenty graded feature combinations were constructed. Classification models were developed using support vector machine(SVM)and random forest(RF)algorithms, with model accuracy evaluated accordingly. The results were as follows:(1)The eight Juncao species exhibited typical vegetation spectral characteristics in the visible light region, with reflectance rising sharply in the red-edge region(700-750 nm).(2)Different spectral processing methods significantly amplified inter-species spectral differences in specific bands, such as reciprocal taking logarithm transformation within 570-650 nm, the first derivatives around 730 nm, and the second derivatives in the 670-760 nm range. The red-edge amplitude, red-edge area, and simple ratio(SR)vegetation index demonstrated the strongest discriminative power among species.(3)Model performance indicated that the SVM algorithm generally outperformed RF. The combination of “original spectra + three-edge parameters” achieved the highest accuracy of 70.56% under SVM, which was 14.03% higher than the same combination under RF. The optimal performance for RF was observed with the “original spectra + reciprocal taking logarithm transformation” combination, though it showed limited adaptability to high-dimensional features. Integrating spectral features from visible and near-infrared bands with spectral transformation techniques and vegetation indices, coupled with the SVM algorithm, provides an effective theoretical foundation and technical support for the rapid and accurate identification and classification of Juncao species. |
| Key words: species identification, spectral reflectance, feature extraction, support vector machine, random forest |
|
|
|
|
|
