Attention-Augmented Deep Learning Models for Robust Mango Dry Matter Content Prediction Using Near-Infrared Spectroscopy

Mango is a high-value tropical fruit whose commercial value and consumer acceptance are strongly influenced by internal quality attributes such as dry matter content (DMC), total soluble solids, firmness, and ripeness. Traditional quality assessment methods are often destructive, time-consuming, and unsuitable for large-scale commercial grading. Consequently, non-destructive sensing technologies, particularly near-infrared spectroscopy (NIRS) and hyperspectral imaging (HSI), have emerged as promising tools for rapid fruit quality evaluation. This thesis investigates the application of machine learning (ML) and deep learning (DL) techniques for non-destructive mango quality assessment, with a primary focus on DMC prediction using NIRS data. A systematic review of 85 studies published over the last decade was first conducted to examine advances in data acquisition, preprocessing, dimensionality reduction, ML, DL, and transformer-inspired approaches for mango quality assessment, identifying key research gaps related to model generalisation, reproducibility, and deployment under real-world conditions. Building upon these findings, a large-scale public NIRS dataset collected by CQUiniversity comprising up to 85,401 spectra collected across multiple seasons and instruments was used to investigate the effects of seasonal and instrumental variability on predictive performance. Comparative evaluations involving partial least squares (PLS), artificial neural networks (ANN), one-dimensional convolutional neural networks (1D-CNNs), and attention-augmented CNN architectures incorporating squeeze-and-excitation (SE), efficient channel attention (ECA), convolutional block attention module (CBAM), and multi-head self-attention (MHSA) were performed. Experiments were conducted on both a single-instrument dataset containing 11,691 spectra from 4,650 fruit samples and a large-scale multi-instrument dataset containing 85,401 spectra from 10,640 fruit samples collected across 31 instruments and seven seasons. The influence of spectral preprocessing, wavelength selection, outlier detection, and software environments was also systematically evaluated. Results demonstrated that DL models consistently outperformed the conventional PLS baseline, with ANN and CNN architectures providing the most robust performance across multiple experimental scenarios. While attention mechanisms improved spectral feature representation, they provided only marginal gains over the baseline CNN. The best predictive performance on the single-instrument dataset was achieved by the baseline CNN (R²p = 0.92, RMSEP = 0.75%), whereas CBAM achieved the best result on the multi-instrument dataset (R²p = 0.78, RMSEP = 1.26%). The findings further revealed that seasonal and instrumental variability substantially affect model generalisation and that spectral preprocessing, wavelength selection, and software environments exert a greater influence on predictive performance than attention mechanism selection.