Classification of In Situ Collected Mid-Infrared Spectroscopic Markers for Determining Strawberry Maturity

The strawberry sector aims to adopt rapid, cost-effective, eco-friendly, and non-destructive methods for monitoring physical quality changes during fruit maturation. Portable Mid-Infrared (MIR) spectrometers using micro-electro-mechanical systems have shown promise in assessing strawberry maturity. Not all spectral data contribute meaningfully to biomarker development. Therefore, dimensionality reduction techniques have been explored to improve data analysis and computational efficiency. In this work we explore a widely applied linear method, Principal Component Analysis (PCA). Since PCA may not always capture complex, non-linear patterns in spectral data, we also use deep learning-based methods such as Autoencoders (AEs) and Variational Autoencoders (VAEs). AEs focus on reconstructing input data with minimal loss, while VAEs incorporate probabilistic modeling to generate new and meaningful representations. This paper proposes applying both linear and non-linear methods to achieve optimal transformation of wavenumbers. Furthermore, these techniques are combined with supervised classification methods to enhance the detection of strawberry maturity states using MIR spectral data. We show that 53 principal components (out of 900 features) were sufficient to retain approximately 95% of the total variance, striking a balance between dimensionality reduction and information preservation. Additionally, the 53 components are used in a Support Vector Machine (SVM) model to achieve 97% accuracy in classifying 14 strawberry maturity levels. Our AE on the other hand, uses a 32-dimensional informative latent space to compresses spectral data, and achieve a 94% accuracy in classification using SVM. Finally, VAE also using a 32-dimensional latent space achieves a 97% accuracy with all three classifiers, SMV, Naive Bayes and Random Forest.