基于高光谱图像特征融合的榛子水分含量测定

采用高光谱图像技术对榛子水分含量进行快速无损检测。采集200个榛子在400~1 000 nm波段的高光谱图像，提取榛子图像区域的平均光谱信息。利用K-S算法划分样品验证集和预测集，使用四种预处理方法对光谱进行预处理。通过竞争自适应加权算法（Competitive Adaptive Reweighted Sampling，CARS）和逐次投影法（Successive Projection Algorithm，SPA）进行光谱特征的提取；灰度共生矩阵法（Gray Level Co-occurrence Matrix，GLCM）提取图像的纹理特征；分别建立基于光谱特征，图像纹理特征以及两者串联融合的偏最小二乘回归（Partial Least Squares Regression，PLSR）和支持向量回归（Support Vector Regression，SVR）模型对榛子水分进行预测。结果表明，CARS和SPA算法能够有效选择特征波长并提升预测性能；图像特征能够对榛子水分进行预测，基于主成分图像提取的图像特征信息建立的模型预测效果更好。光谱图像特征融合能明显提高对榛子水分含量预测的准确率，CARS提取的特征波段结合主成分图像的纹理特征显示出了更好的效果，SVR模型的RMSECV为0.03，RC 为0.97，RMSEP为0.04，RP为0.96。利用高光谱图像和纹理特征能够对榛子水分进行有效预测，为榛子水分含量检测提供了新的方法。

Determination of Moisture Content of Hazelnuts Based on Hyperspectral Image Feature Fusion

For the rapid, non-destructive detection of moisture content in hazelnuts, hyperspectral image technology was utilized. A dataset comprising hyperspectral images of 200 hazelnuts covering wavelengths of 400~1 000 nm was collected, and the average spectral information of the hazelnuts image regions was extracted. The dataset was divided into sample validation and prediction sets using the K-S algorithm. Additionally, four preprocessing methods were applied to enhance spectra quality. Spectral features were extracted using a competitive adaptive weighting algorithm (CARS) and a successive projection method (SPA). Image texture features were obtained using the gray-scale co-occurrence matrix method (GLCM). Partial least squares regression (PLSR) and support vector regression (SVR) models were developed based on spectral features, image texture features, and the fusion of both to predict hazelnut moisture. The CARS and SPA algorithms effectively selected feature wavelengths and enhanced prediction performance. Furthermore, image features showed potential in predicting hazelnut moisture content, particularly when extracted from principal component images. The fusion of spectral and image features significantly enhances the accuracy of hazelnut moisture content prediction, especially when combining CARS-selected feature wavelengths with texture features from principal component images. The SVR model achieved impressive results, with an RMSECV of 0.03, RC of 0.97, RMSEP of 0.04, and RP of 0.96. This study highlights the effectiveness of hyperspectral image and texture features in predicting hazelnut moisture content, providing a novel approach for moisture detection in hazelnuts.