高光谱透射成像联合特征区间筛选的鸽蛋受精无损检测方法

Non-destructive Fertility Detection of Pigeon Eggs Using Hyperspectral Transmission Imaging Combined with Feature Interval Selection

摘要:

为提升鸽蛋育种与孵化管理的无损检测效率与准确性,构建了一套面向孵化期种鸽蛋受精状态快速判别的高光谱透射成像识别流程。 在高光谱透射式装置下采集原始高光谱数据,进行黑白校正后,将整个蛋体作为感兴趣区域获取其平均光谱信息。 采用了标准正态变量校正 (Standard normal variate, SNV)、 小波变换 (Wavelet transform, WT)、多元散射校正(Multiplicative scatter correction, MSC)3 种预处理算法;通过 KS 样本选取策略按照比例 7∶ 3划分训练集与测试集;结合 BiPLS、iPLS 与 SiPLS 对光谱区间进行特征波段区间筛选;在此基础上分别训练线性模型(LDA、LR)与非线性分类模型(XGBoost),并分析最优预处理—特征选择—建模组合与关键判别波段。 采取一阶导数处理方法,对受精 / 非受精的光谱差异进行详细分析。 结果表明,在孵化前分类效果最好的是 BiPLS-LDA-WT 算法,对受精种蛋的召回率达到 0. 98,准确率为 0. 87,F1 分数为 0. 89;在孵化第 1 天、第 2 天和第 3 天, 分类准确率分别为 0. 92、0. 95 和 1. 00,召回率分别为 0. 98、1. 00、1. 00。 本研究所提出的集成式高光谱特征区间筛选与多模型对比框架具备良好的准确性与可解释性,可为种鸽蛋孵化早期无损受精检测及孵化管理提供技术支撑。

Abstract:

Aiming to improve the efficiency and accuracy of non-destructive testing in pigeon egg breeding and incubation management, a hyperspectral transmission imaging recognition process was established for the rapid determination of the fertilization status of breeding pigeon eggs in the early stages of incubation. Original hyperspectral data were collected by using a hyperspectral transmission device, and after black-and-white correction, the entire egg was taken as the region of interest to obtain its average spectral information. Three preprocessing algorithms were applied: standard normal variate (SNV), wavelet transform (WT), and multiplicative scatter correction (MSC). Using a KS sample selection strategy, the data were split into training and test sets at a ratio of 7 ∶ 3. Feature wavelength intervals were selected by BiPLS, iPLS, and SiPLS for spectral interval screening. Based on this, linear models (LDA, LR )and nonlinear classification models (XGBoost )were trained, and the optimal preprocessing-feature selection-modeling combination and key discriminative wavelengths were analyzed. First derivative processing was used for a detailed analysis of the spectral differences between fertilized and unfertilized eggs. The results showed that, before incubation, the BiPLS-LDA-WT algorithm achieved the best classification performance, with a recall rate of 0. 98 for fertilized eggs, an accuracy of 0. 87, and an F1-score of 0. 89. On the 1st, 2nd, and 3rd days of incubation, the classification accuracies were 0. 92, 0. 95, and 1. 00, respectively, with recall rates of 0. 98, 1. 00, and 1. 00. The proposed integrated hyperspectral feature interval selection and multi-model comparison framework demonstrated good accuracy and interpretability, providing technical support for early non-destructive fertilization detection and incubation management of breeding pigeon eggs.