Hyperspectral imaging technology for determining soluble solids content in Dangshansuli pears with CPO-PLSR modeling optimization

【Objective】Dangshansuli pear (Pyrus bretschneideri Rehd.), as one of the widely popular pear varieties, is renowned for its crisp and sweet flesh, juicy texture, and rich nutrition. Soluble solids content (SSC) serves as a crucial indicator of honey pear quality. Traditional methods for SSC determination rely on sample destruction and the use of chemical reagents, which are time-consuming and labor- intensive for large- scale sample analysis, making them unsatisfactory for market application demands. Therefore, the exploration of a novel technology that enables rapid and non-destructive detection has become a research hotspot. This study proposed an advanced non-destructive quantitative meth-od for SSC determination in Dangshansuli pears, which integrates hyperspectral imaging (HSI) and Crested Porcupine Optimizer-Partial Least Squares Regression (CPO-PLSR) techniques. By combining innovative wavelength selection algorithms with metaheuristic parameter optimization, the study tackled the inherent issues of spectral redundancy and model instability in HSI-based SSC prediction, thereby contributing to the development of precision agriculture and intelligent fruit quality inspection equipment.【Methods】After standing for 24 hours under controlled environmental conditions (25±1 ℃, 60% relative humidity), hyperspectral images of 187 honey pear samples were acquired using the Headwall Micro-Hyperspec VNIR A system (400-1000 nm). A portable digital refractometer, the PAL-1 from Atago Co., Japan, was used to perform three independent measurements of soluble solids content (SSC) in Dangshansuli pear samples, and the average value was taken as the reference. The dataset (187 samples) was partitioned into a training set (140 samples) and a test set (47 samples) at a 3∶1 ratio using the sample set portioning based on the joint x-y distance (SPXY) algorithm. The SPXY algorithm comprehensively considered both spectral features and sample physicochemical properties, enabling a more holistic assessment and partitioning of the dataset. The SSC of the Dangshansuli pears ranged from 9.57 to 13.58 °Brix, with standard deviations of 0.727 3 and 0.559 4 for the training and test sets, respectively. The means of the training and test sets were close, and the overall coefficient of variation was low, indicating that the partitioning of the dataset was reliable. The original spectral images were opened using ENVI software, and raw spectral data were extracted from 20 × 20 pixel square regions of interest (ROIs). After interactive selection and recording of regional spectral averages, computational modeling of hyperspectral characteristics was conducted using MATLAB R2022b environment, which supported initial image partitioning and feature space analysis. The reflectance spectra were preprocessed using centered normalization and moving average smoothing. Following this, three feature selection methods were employed for rigorous dimensionality reduction: Sequentially Projected Algorithm (SPA) (retaining 19 wavelengths, accounting for 5.8% ), Competitive Adaptive Reweighted Sampling (CARS) (142 wavelengths, 43.42% ), and Improved Modified Uninformative Variable Elimination (imUVE) (122 wavelengths, 37.65%). Partial Least Squares Regression (PLSR) achieved effective linear fitting of the data by minimizing the sum of squared errors between predicted and actual values, and was widely used in chemometric analysis. PLSR maps input and output variables into a low-dimensional space and performs principal component decomposition, selecting a set of optimal composite variables that would effectively explain system variations, which were then used for regression modeling. In this experiment, the maximum number of latent principal components for PLSR was set to 20. The Crested Porcupine Optimizer (CPO) algorithm, inspired by the defensive strategies of porcupines, dynamically optimized the number of principal components in the PLSR model through an iterative evolutionary process for automated parameter adjustment, thereby reducing the risks of underfitting and overfitting.【Results】The spectral trend and peak coordinates of the MA-Centered spectra showed no significant changes compared with those before transformation, but the amplitude curves were more clustered, with reflectance standardized to the range of -2.5 to 1.5. This significantly reduced noise interference and scattering effects, thereby improving the resolution and reliability of the spectral data. The SPA excelled in feature wavelength extraction (5.8% ), minimizing computational load while preserving chemically significant regions, particularly near 480 nm (absorption by chlorophyll and carotenoids) and near 960 nm (O-H/NH vibrational overtones). In contrast, CARS and imUVE retained higher redundancy (43.42% and 37.65%, respectively), introducing slight prediction biases. The SPA-CPO-PLSR model exhibited the strongest predictive capability, achieving R2 p = 0.692 9, RMSEP = 0.306 7 °Brix, and RPD = 1.824 1,outperforming traditional PLSR ( R2 p = 0.691 5, RMSEP = 0.307 4) and other feature- based models (CARS-CPO-PLSR: R2 p = 0.585 5; imUVE-CPO-PLSR: R2 p = 0.613 7). CPO optimization significantly enhanced the performance of full-spectrum PLSR, increasing R2 p from 0.231 8 to 0.551 2 and RPD from 1.153 9 to 1.509 6, validating CPO's ability to effectively address underfitting and overfitting issues in traditional PLSR models by adjusting the number of latent variables. The combination of HSI with CPO-PLSR enabled rapid analysis (＜3 seconds per sample) and high stability.【Conclusion】This study established an efficient method for non- destructive SSC (Soluble Solids Content) detection in Dangshansuli pears by combining hyperspectral imaging technology with Partial Least Squares Regression (PLSR) optimized by the Crested Porcupine Optimization (CPO) algorithm. This would provide a robust and non-destructive solution for the rapid assessment of SSC in Dangshansuli pears. The experimental results demonstrated the feasibility of using hyperspectral imaging technology to detect SSC in mangoes, thereby proving the potential of this technology in the internal quality inspection of fruits and offering a new efficient method for SSC determination in Dangshansuli pears. This method would significantly improve prediction accuracy and operational efficiency compared with traditional techniques by automating principal component selection and preferentially extracting information- rich wavelengths. Future research should extend this framework to multi-parameter fruit quality analysis (such as acidity, vitamin content, etc.) and validate its applicability across different varieties and agricultural products.