Deep learning for classification of time series spectral images using combined multi-temporal and spectral features

Time series spectral imaging facilitates a comprehensive understanding of the underlying dynamics multi-component systems and processes. Most existing classification strategies focus exclusively on features they tend to fail when spectra between classes closely resemble each other. This work proposes hybrid approach principal component analysis (PCA) deep learning (i.e., long short-term memory (LSTM) model) for incorporating utilizing combined multi-temporal information from time datasets. An example data, consisting times images casein-based biopolymers, was used illustrate evaluate proposed approach. Compared using partial least squares discriminant (PLSDA), PCA-LSTM method applying same pretreatment achieved substantial improvement in pixel-wise accuracy increased 59.97% PLSDA 85.73% PCA-LSTM). When projecting model object-based classification, produced an 100%, correctly classifying whole 21 film samples independent test set, while only led 80.95%. The is powerful versatile distinctive characteristics dependencies multivariate dataset, which could be adapted suit non-congruent over sequences as well spectroscopic data.