基于高光谱成像技术的亚麻籽粒品种快速分类

Rapid Classification of Flax Seed Grain Varieties Based on Hyperspectral Imaging Technique

摘要: 亚麻是世界上一种重要的经济作物和油料作物，快速筛选不同品种的亚麻籽对于亚麻育种以及农业种植具有重要意义。市场上亚麻籽品种繁多，如何快速、准确、高效地鉴别亚麻籽的品种，从而实现亚麻品种优育以及高产高收亟待解决。本文提出基于高光谱图像技术的K近邻(KNN)分类模型，以期实现不同品种亚麻籽的快速分类。本研究以甘肃省农业科学院提供的10个不同品种的亚麻籽作为本次试验的样本，每个品种随机选取50粒亚麻籽，利用高光谱成像系统采集870.07~1709.45 nm范围内的亚麻籽光谱图像。将采集到的图像黑白校正之后，以50粒亚麻籽为感兴趣区域，求取该区域内平均光谱作为原始光谱数据。由于原始光谱数据两端噪声较大，有效信息会受到噪声干扰，为了增强信噪比，本研究截取950~1680 nm范围内的亚麻籽的光谱波段作为有效波段进行分析。为避免数据在采集过程中受无关信息的干扰较强从而影响建模效果，因此对去噪后的光谱波段信息进行SG (Savitzky Golay)平滑预处理，并在SG平滑预处理的基础上分别进行最大归一化(MN)和二阶求导预处理(2^ndDer)。数据预处理后分别采用竞争性自适应重加权算法(CARS)、连续投影算法(SPA)单一提取和CARS + SPA、CARS-SPA组合筛选方法提取特征波长，利用K近邻算法建立CARS-KNN、SPA-KNN、CARS + SPA-KNN、CARS-SPA-KNN这4种亚麻籽品种鉴别模型。实验结果表明：基于(SG-2^ndDer)-CARS-KNN、(SG-2^ndDer)-SPA-KNN、(SG-2^ndDer)-CARS + SPA-KNN、(SG-2^ndDer)-CARS-SPA-KNN这4种分类模型对亚麻籽的分类准确率最高可以达到100%。故利用近红外高光谱成像技术结合KNN算法对亚麻籽品种进行快速无损鉴别是优异和可靠的方案。

Abstract: Flax is an important economic and oil crop globally; rapid screening of different flax seed varieties is essential for flax breeding and agricultural cultivation. There are many flax seed varieties on the market; how to quickly, accurately, and efficiently identify the flax seed varieties to realize the superiority of flax varieties and high-yield yield needs to be resolved. This paper proposes the K-Nearest Neighbor (KNN) classification model based on the hyperspectral image technique to classify flax seed varieties quickly. In this study, ten different varieties of flax seeds provided by the Gansu Provincial Academy of Agricultural Sciences were used as the samples for this experiment, and 50 flax seeds of each variety were randomly selected. The spectral images of flax seeds in the range of 870.07~1709.45 nm were collected by using a hyperspectral imaging system. After the black-and-white correction of the acquired images, the 50 flax seeds were the region of interest, and the average spectrum in the region was the original spectral data. Due to the significant noise at both ends of the raw spectral data, the practical information will be interfered with by noise; in order to enhance the signal-to-noise ratio, this study intercepted the spectral bands of linseed in the range of 950~1680 nm as the influential bands for analysis. In order to avoid the data in the acquisition process by irrelevant information interference is more muscular, thus affecting the modeling effect, the de-noised spectral band information is pre-processed by SG (Savitzky Golay) smoothing, and based on SG smoothing, respectively, the Maximum Normalization (MN) and second-order derivation pre-processing (2^ndDer). After data preprocessing, the feature wavelengths were extracted by Competitive Adaptive Re-weighting Algorithm (CARS), Successive Projection Algorithm (SPA) single extraction and CARS + SPA, CARS-SPA combined screening methods, respectively, and the K-Nearest Neighbor algorithm was utilized to establish the identification models for the four flax seed varieties, namely, CARS-KNN, SPA-KNN, CARS + SPA-KNN, and CARS-SPA-KNN. The experimental results show that the classification accuracy of flax seed based on the four classification models (SG-2^ndDer)-CARS-KNN, (SG-2^ndDer)-SPA-KNN, (SG-2^ndDer)-CARS + SPA-KNN, and (SG-2^ndDer)-CARS- SPA-KNN) can reach up to 100%. Therefore, using near-infrared hyperspectral imaging technology combined with the KNN algorithm for rapid and non-destructive identification of linseed varieties is an excellent and reliable program.