基于随机森林法的农作物高光谱遥感识别

  目的  不同农作物种类光谱差异小，通过探测众多窄波段范围的细微差别，提取区分不同农作物的特征波段，是目前实现农作物高光谱遥感识别的重要途径。如何提取区分不同农作物的特征波段，进而实现农作物的精确识别是一个挑战。近来出现的随机森林方法在多变量目标的分类识别方法展现了优势，为解决这一难题提供了一个新手段。  方法  利用随机森林法与传统方法分析杭州地区8种典型农作物的反射光谱，提取特征波段并进行分类，对比不同方法的识别效果。  结果  不同作物的反射光谱及其一阶微分、二阶微分、倒数的对数、去包络线法所提取的特征波段只能区分部分作物；随机森林法无需对反射光谱预处理，直接对全波段反射光谱数据处理，不仅筛选出了区分不同作物的特征波段，且运用所选择的波段对作物进行随机森林分类的效果也是最优的。  结论  随机森林法选择的波段（550、2 490、370、770、560、380、540、530、570、350 nm）不仅能区分不同作物，还能反映农作物生化属性的不同，使得用于分类的波段及分类方法体现了不同作物间物化性质的不同，在展现高光谱遥感识别农作物优势的同时，也为大面积农作物遥感精细分类提供借鉴。

Crops identification based on hyperspectral data and random forest method

  Objective  The difference in hyperspectral reflectance of different crops is small, the subtle differences of the spectral feature enable identification of different crops. This aim is to find an effective way to select spectral feature bands from hundreds of bands and indentify different crops.  Method  The random forest method and the traditional methods were applied to analyze the reflective hyperspectra of 8 typical crops in Hangzhou, the spectral feature bands were extracted and the crops were distinguished. The traditional methods included first-order derivative, second-order derivative, logarithm of reciprocal, and the de-enveloping line method.  Result  The reflective hyperspectra of the different crops and the results processed by traditional methods were limited in crops identification, the spectral feature extracted by these traditional ways could not distinguish all of 8 crops. The random forest method was the most effective one for extraction of the spectral feature.  Conclusion  The extracted spectral feature bands were 550, 2 490, 370, 770, 560, 380, 540, 530, 570 and 350 nm, which not only were applied effectively to distinguish 8 crops, but also reflected the biochemical properties of the crops, which made sense to explain the classification with hyperspectral remote sensing.The random forest method fully demonstrated the advantages of hyperspectral remote sensing for crop identification and spectral feature extraction. The random forest classification can provide reference for large-scale crop fine classification with hyperspectral image.