Simple methods for uncertainty estimation in neural networks applied to spectral data processing: A case study on mango dry matter prediction

Abstract

The growing complexity of real-world chemometric applications, particularly in spectroscopy, has exposed the limitations of traditional linear models in capturing non-linear patterns in spectral data. Deep learning models offer a powerful alternative but remain underutilised in chemometrics due to concerns about interpretability and trust, particularly in high-risk applications where uncertainty estimation is critical. This study investigates and compares three uncertainty estimation techniques suitable for neural networks: Monte Carlo Dropout (MC dropout), model averaging, and Stochastic Weight Averaging-Gaussian (SWAG). These methods are evaluated using a spectral deep learning architecture. The analysis focuses on identifying key hyper-parameters affecting both predictive performance and uncertainty calibration. Results show that while MC Dropout offers a good balance between accuracy and uncertainty estimation at low computational cost, model averaging provides robust performance but at the expense of greater training time and storage. SWAG emerges as a middle-ground method requiring careful tuning. Importantly, a trade-off between predictive accuracy and uncertainty calibration is observed, underscoring the need to consider uncertainty as an integral part of model evaluation. These findings highlight the relevance of deep learning uncertainty estimation in chemometrics and open new directions for optimising data acquisition, model calibration, and model selection based on both prediction confidence and performance.