Reevaluating Convolutional Neural Networks for Spectral Analysis: A Focus on Raman Spectroscopy

TL;DR

This work reframes CNN-based Raman spectroscopy classification for autonomous, resource-constrained missions by training on raw spectra and carefully controlling translational invariance. It demonstrates four practical advances: (i) baseline-free accuracy, where 1-D CNNs outperform KNN/SVM on raw data without background correction; (ii) pooling-based robustness that tunes invariance to Raman shifts up to $30\, \mathrm{cm^{-1}}$; (iii) label-efficient learning via SGAN and contrastive pretraining that gains up to $11\%$ with only $10\%$ labels; and (iv) constant-time adaptation by freezing backbones and retraining only the softmax layer for new minerals, outperforming Siamese approaches on embedded hardware. The study also provides in-depth analysis of CNN inductive biases for spectra, interpretable Grad-CAM insights, and transfer-learning limitations, supported by reproducible data splits from the RRUFF database. Collectively, these findings offer a scalable, data-efficient pathway for robust Raman classification in autonomous planetary and oceanic exploration contexts. The open-release of datasets and scripts further enables benchmarking and deployment-ready evaluation.

Abstract

Autonomous Raman instruments on Mars rovers, deep-sea landers, and field robots must interpret raw spectra distorted by fluorescence baselines, peak shifts, and limited ground-truth labels. Using curated subsets of the RRUFF database, we evaluate one-dimensional convolutional neural networks (CNNs) and report four advances: (i) Baseline-independent classification: compact CNNs surpass $k$-nearest-neighbors and support-vector machines on handcrafted features, removing background-correction and peak-picking stages while ensuring reproducibility through released data splits and scripts. (ii) Pooling-controlled robustness: tuning a single pooling parameter accommodates Raman shifts up to $30 \,\mathrm{cm}^{-1}$, balancing translational invariance with spectral resolution. (iii) Label-efficient learning: semi-supervised generative adversarial networks and contrastive pretraining raise accuracy by up to $11\%$ with only $10\%$ labels, valuable for autonomous deployments with scarce annotation. (iv) Constant-time adaptation: freezing the CNN backbone and retraining only the softmax layer transfers models to unseen minerals at $\mathcal{O}(1)$ cost, outperforming Siamese networks on resource-limited processors. This workflow, which involves training on raw spectra, tuning pooling, adding semi-supervision when labels are scarce, and fine-tuning lightly for new targets, provides a practical path toward robust, low-footprint Raman classification in autonomous exploration.

Figures (25)

• Figure 4: Feature extraction process for traditional ML methods. The Raman spectrum is segmented into intervals and peaks are counted within each interval.
• Figure 5: Wavelet peak detection on Andalusite. Upper: RRUFF-raw spectrum—several diagnostic peaks are missed due to fluorescence baseline. Lower: RRUFF-clean spectrum—more peaks are detected, illustrating why peak‑based feature extraction works better on the cleaned data.
• Figure 6: Architecture of the 1-D CNNs model used for Raman spectral classification. The input consists of raw Raman spectra, which pass through three convolutional layers (16, 32, and 64 filters, with kernel sizes of 21, 11, and 5, respectively). Each convolutional layer is followed by Batch Normalization (BN), LeakyReLU activation, and max pooling (2×2). The extracted features are concatenated and processed through a dense layer (2048 neurons) with hyperbolic tangent (Tanh) activation, dropout (0.5), and additional pooling. The final classification layer applies a softmax function to assign the spectrum to one of the mineral classes. This architecture is based on the design proposed by Liu et al.C7AN01371J.
• Figure 7: Top-1 accuracy on the test set for CNNs, CNNs+KNN (feature extraction with CNNs followed by KNN classification), MLP, and KNN classifiers. Results are shown for both RRUFF-raw and RRUFF-clean Raman datasets, with and without data augmentation. CNNs achieve the highest performance across all settings, especially with data augmentation.
• Figure 8: Grad-CAM activation maps for two carbonates. Top: Dolomite (CaMg(CO3)2), correctly classified. Bottom: Rhodochrosite (MnCO3), correctly classified.