Convolutional neural networks for mapping of lake sediment core particle size using hyperspectral imaging

The particle size of lake sediments integrates important environmental information, and the detection of changes in this variable over time provides important information for understanding ecosystem and sedimentary pro-cesses. Although standard machine learning regression algorithms especially random forest (RF) have shown great potential for particle size mapping in sediment cores using hyperspectral imaging, no research has yet applied deep learning approaches. One-dimensional convolutional neural networks (CNN) have recently been developed and applied in several applications in the field of spectroscopy. This study addresses this issue by developing and applying a new methodology based on a 1D convolutional autoencoder as the feature extractor and a 1D CNN architecture for regression. The proposed architecture was applied to hyperspectral images of nine lake sediment cores across Canada and evaluated against results of the RF algorithm. However, in order for the results to be comparable, the RF algorithm was performed on features that also resulted from the convolutional autoencoder. According to the leave-one-out evaluation method, the proposed CNN method represented an improvement of 14%, 4.58, 5.45, and 0.83 for R2, MAE, RMSE, and RPD, respectively, relative to the best RF algorithm. Our findings show that the proposed method can be reliably used to reconstruct particle size in sediment cores from lakes of varying climatic and environmental characteristics.

Automated summary

The particle size of lake sediments contains important environmental information and detecting its changes over time is crucial for understanding ecosystems and sedimentary processes. This study proposes a new method that utilizes a one-dimensional convolutional autoencoder and a one-dimensional convolutional neural network (CNN) for regression analysis, and successfully reconstructs particle size from hyperspectral images of lake sediment cores. The proposed CNN method outperforms the traditional random forest algorithm in terms of predictive accuracy.