Automated spectroscopic modelling with optimised convolutional neural networks.

Area of Science:

• Spectroscopy
• Machine Learning
• Soil Science

Background:

• Convolutional Neural Networks (CNNs) are used for spectroscopic modeling but are often manually tuned.
• Manual tuning can lead to suboptimal model performance and lacks analysis of hyperparameter effects.

Purpose of the Study:

• To develop and demonstrate an automated approach for tuning one-dimensional CNNs (1D-CNNs) for spectroscopic modeling.
• To optimize hyperparameters to maximize model performance in estimating soil organic carbon (SOC).

Main Methods:

• A parametric representation of 1D-CNNs was developed for automated hyperparameter optimization.
• The approach was tested on a large European soil spectroscopic database for SOC content estimation.
• Optimization performance was compared against random search, and hyperparameter importance was analyzed using functional Analysis of Variance.

Main Results:

• The automated optimization approach demonstrated faster convergence and superior results compared to random search.
• The optimized model achieved highly accurate SOC content estimations ([Formula: see text] and [Formula: see text]).
• Hyperparameters related to model training and architecture significantly impacted performance, while spectral preprocessing hyperparameters had minimal effect.

Conclusions:

• Automated hyperparameter optimization simplifies the development of 1D-CNNs for spectroscopic applications.
• The approach enhances model reliability by automatically selecting optimal hyperparameters and preprocessing methods.
• Hyperparameter importance analysis provides insights into the tuning process and model behavior.