An accurate firmness detection method for Korla pears based on data augmentation

The firmness of Korla fragrant pears is a core quality attribute that determines their commercial value and storage life. However, traditional penetrometer-based methods are inefficient and destructive, hindering comprehensive and accurate quality assessment; these limitations represent a key technical bottleneck for the standardization and intelligent upgrading of the pear industry. To achieve accurate firmness detection, this study utilized Near-Infrared Spectroscopy (NIRS) over 900–1800 nm to measure pear firmness non-destructively. Four preprocessing methods—Savitzky-Golay Smoothing (SGS), Min-Max Normalization (MMN), Vector Normalization (VN) and Multiplicative Scatter Correction (MSC)—and two wavelength selection algorithms—Uninformative Variable Elimination (UVE) and Successive Projections Algorithm (SPA)—were used to develop four prediction models: Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN). Generative Adversarial Network (GAN), Deep Convolutional GAN (DCGAN), Wasserstein GAN (WGAN), and WGAN with Gradient Penalty (WGAN-GP), were employed for data augmentation by generating 100–400 synthetic samples. The SGS-UVE-PLSR model performed best (R² = 0.78, RMSE = 0.22 N) and was enhanced using DCGAN. With 400 synthetic samples, the improved model (DCGAN-SGS-UVE-PLSR) achieved R² = 0.94 and RMSE = 0.17 N, and demonstrated strong performance on a new dataset (R² = 0.92, RMSE = 0.24 N). This study provides a theoretical framework for quality monitoring of Korla fragrant pears and the non-destructive detection of firmness in other fruits. • Firmness detection method for Korla fragrant pears using data augmentation technology • DCGAN can effectively improve the accuracy of firmness prediction • DCGAN-SGS-UVE-PLSR achieved the best prediction result R²=0.94,RMSE= 0.17 N • Optimal model validation: R²= 0.92, RMSE = 0.24 N