Self-supervised and Multi-fidelity Learning for Extended Predictive Soil Spectroscopy

TL;DR

This paper addresses the challenge of scalable, accurate soil spectroscopy by developing a self-supervised learning framework that derives a compact $32$-dimensional latent space from MIR spectra. It then creates a bridge to lower-cost NIR measurements by training an NIR encoder that maps into the MIR latent space while keeping the MIR decoder fixed. Predictive models map latent representations (and converted spectra) to nine soil properties, showing that MIR-derived embeddings yield higher accuracy and consistency, while NIR-to-MIR conversion via the latent space often matches or surpasses NIR-only baselines. The approach offers data efficiency, interpretability through latent-feature correlations, and a practical path for deploying portable NIR devices to leverage large MIR datasets in soil health monitoring.

Abstract

We propose a self-supervised machine learning (SSML) framework for multi-fidelity learning and extended predictive soil spectroscopy based on latent space embeddings. A self-supervised representation was pretrained with the large MIR spectral library and the Variational Autoencoder algorithm to obtain a compressed latent space for generating spectral embeddings. At this stage, only unlabeled spectral data were used, allowing us to leverage the full spectral database and the availability of scan repeats for augmented training. We also leveraged and froze the trained MIR decoder for a spectrum conversion task by plugging it into a NIR encoder to learn the mapping between NIR and MIR spectra in an attempt to leverage the predictive capabilities contained in the large MIR library with a low cost portable NIR scanner. This was achieved by using a smaller subset of the KSSL library with paired NIR and MIR spectra. Downstream machine learning models were then trained to map between original spectra, predicted spectra, and latent space embeddings for nine soil properties. The performance of was evaluated independently of the KSSL training data using a gold-standard test set, along with regression goodness-of-fit metrics. Compared to baseline models, the proposed SSML and its embeddings yielded similar or better accuracy in all soil properties prediction tasks. Predictions derived from the spectrum conversion (NIR to MIR) task did not match the performance of the original MIR spectra but were similar or superior to predictive performance of NIR-only models, suggesting the unified spectral latent space can effectively leverage the larger and more diverse MIR dataset for prediction of soil properties not well represented in current NIR libraries.

Figures (9)

• Figure 1: Illustrative example of functional organic groups and their respective spectral bands at the fundamental mid-infrared (MIR) range and their overtones in the near-infrared (NIR) range for a random soil sample (top): fundamental band for Polysaccharides C-O group is positioned around 1170 cm$^{-1}$ (8547 nm) with its respective fourth overtone represented around 2137 nm; fundamental band for Methyls C-H group is positioned between 1445-1350 cm$^{-1}$ (6920-7407 nm) with its respective third overtone represented between 2307-2469 nm; fundamental band for Carboxylic acids C=O group is positioned around 1725 cm$^{-1}$ (5797 nm) with its respective third overtone represented around 1930 nm; fundamental band for Hydroxyls H-O group is positioned around 3600 cm$^{-1}$ (2778 nm) with its respective second overtone represented around 1400 nm Weyer1985. This physical relationship of fundamental vibrations and overtones across the infrared range forms the basis of a unified latent space hypothesis (bottom).
• Figure 2: Learning strategies employed in this study. Firstly, the self-supervised learning is leveraged to pre-train and learn a latent space representation from mid-infrared (MIR) spectra with variational autoencoders (VAE) and scan repeats. Then, the learned MIR decoder is frozen and aligned with a near-infrared (NIR) encoder to enable the learning and mapping between the two fidelities (NIRtoMIR). Lastly, nine relevant soil properties were mapped from original spectral data, learned latent representations, and multi-fidelity outputs for predictive purposes using either partial least squares regression (PLSR) or multilayer perceptron (MLP). Please note: stacked neural networks are illustrative examples.
• Figure 3: Probability distribution functions (PDF) of the latent space representations obtained from the self-supervised learning (left). The heatmap depicts the non-parametric Xi correlation between latent space features and the original MIR spectra, binned and averaged in increments of 25 cm-1 for the correlation analysis.
• Figure 4: Comparison of original and predicted mid-infrared (MIR) spectra of the independent test samples (top panel, n=206). Mean and standard deviation difference (predicted-observed) is highlighted in the bottom panel.
• Figure 5: Lin's concordance correlation coefficient (CCC) of different strategies for soil properties prediction of an independent test set. Reference models are provided for both near-infrared (NIR) and mid-infrared (MIR) spectra using two fitting algorithms: partial least squares regression (PLSR) and multilayer perceptron (MLP). In addition, MLP is combined with latent space embeddings derived from the self-supervised learning task (SSL) of MIR spectra, as well as with both the predicted MIR spectra and their latent space embeddings derived from multi-fidelity learning (NIRtoMIR). A horizontal line is fixed at CCC = 0.7. Blue color represents prediction models using MIR spectrum as input. Cyan color represents prediction models using latent space representations as input. And Yellow colors represents prediction models using NIR spectrum as inputs.