基于NIR光谱和梯度提升决策树的农作物含水率快速检测平台研究

Platform for Rapid Detection of Crop Moisture Content Based on NIR Spectroscopy and GBDT

摘要:

农作物含水率与其质量息息相关，含水率过高易引发霉变，过低则会影响形态完整，然而现有便携检测方法存在精度低、检测速度慢等问题，烘干法处理单样本需40 min，定容称量法检测误差为1.2%，为满足农作物含水率快速检测需求，本文设计了一种基于近红外(Near-infrared, NIR)光谱和梯度提升决策树(Gradient boosting decision tree, GBDT)的农作物含水率快速检测平台。基于Zemax优化设计了离轴切尼-Turner(Czerny-Turner, CT)光学结构，采用ZYNQ芯片内嵌线阵CCD解码算法并与手机连接，在手机上建立萨维茨基-戈雷(Savitzky-Golay, SG)去噪和分段直接标准化(Piecewise direct standardization, PDS)数据迁移算法，通过光谱采样平台采样后再加载手机的GBDT模型预测含水率。Zemax仿真证明离轴设计光学结构在体积和像差控制上均优于传统球面设计，在像面边缘950 nm处仍可以保持2 nm分辨率。SG去噪可使信噪比(SNR)从18.11 dB提升到35.10 dB，基于GBDT建立的含水率预测模型在验证集回归预测中决定系数R2和预测均方根误差(RMSEP)分别为0.877 7和0.135 4%，在测试集中R2和RMSEP分别为0.861 7和0.121 4%，含水率绝对误差小于等于0.4%。检测平台具有体积小、速度快、方便携带等优势，单样本检测时间相较于烘干法（40 min）小，仅需0.5 min，满足农作物含水率快速高效检测需求。

Abstract:

Water content detection is essential for preventing mold growth and ensuring the proper storage and transportation of crops. However, existing portable detection methods often suffer from low accuracy and efficiency. To meet the demand for rapid crop water content measurement, a fast detection platform was proposed based on near-infrared (NIR) spectroscopy and a gradient boosting decision tree (GBDT) model. An off-axis Czerny-Turner (CT) optical structure was optimized by using Zemax, and a ZYNQ chip was employed to implement embedded decoding algorithms for a linear array CCD, with the system connected to a mobile phone. Savitzky-Golay (SG) smoothing and a piecewise direct standardization (PDS) data transfer algorithm were implemented on the mobile device. After spectral acquisition, the GBDT model running on the mobile phone was used to predict crop water content. Zemax simulations demonstrated that the off-axis optical design outperformed traditional spherical designs in terms of volume reduction and aberration control, maintaining a spectral resolution of 2 nm at the 950 nm edge of the image plane. SG smoothing improved the signal-to-noise ratio (SNR) from 18.11 dB to 35.10 dB. The GBDT-based prediction model achieved a coefficient of determination (R2) of 0.877 7 and a root mean square error of prediction (RMSEP) of 0.135 4% on the validation set. On the test set, the model attained an R2 of 0.861 7 and an RMSEP of 0.121 4%, with prediction absolute errors within 0.4%, meeting the requirements for rapid and efficient crop water content detection.