AI-based approaches can improve soil structure evaluation in micromorphological analyses

Soil structure plays an important role in many soil processes such as roots penetration, water retention and the development of microbial habitats. For this reason, studying soil structure is a fundamental step to better understand how these processes work.
Recently, AI-based models development has paved the way for their implementation in analysing and classifying soil features. Convolutional neural networks (CNNs) like U-Net and Mask R-CNN have shown great potential and are often used for segmenting soil pores or plant roots but, despite this, the potential of deep learning in segmenting different types of soil structures remains relatively unexplored. Currently, soil structure evaluation methods often rely on subjective interpretations and therefore subject to human error.
This study explores the potential of an AI-based method as a reliable decision support tool for soil microstructure assessment. The goal of the training was the correct segmentation of different types of soil structures (e.g. Crumb, Granular, Angular, Sub-angular, Massive).
Since some structures were underrepresented in the original dataset, data augmentation was applied to balance the dataset. Subsequently, the dataset was split into training (70%), validation (20%), and test (10%) sets. The training and validation sets were used for model training and validation. The test set, which was excluded from the training phase, was used to evaluate model performance through four accuracy metrics: Precision, Recall, Dice coefficient, and Intersection over Union (IoU).
According to the test set predictions, with a mean Dice coefficient of 0.81 and mean IoU of 0.73 across all soil structure classes, the model demonstrated strong performance in segmenting soil structures. As expected, the model achieved the best results for those structures that were better represented in the training dataset.
Our findings suggest that, although the quality and heterogeneity of the training dataset play a crucial role, AI has the potential to transform soil structure evaluation, providing more objective analyses while reducing the incidence of human bias.