Decoding soil properties from surface cracks using Minkowski functionals, junction crack angle distributions, and AI-based image analysis.

Desiccation cracks depend on the type of soil, with each type exhibiting a distinct pattern. In this study, we examined the evolution of the crack patterns in different soil types as well as with changes within the same soil type. Physico-chemical studies of the sub-classes of soils, one taken from a flooding left bank and the other from a non-flooding right bank, were nearly identical. However, desiccation crack experiments, analysed using morphological descriptors including Minkowski functionals and junction crack angle distribution, exhibited distinct patterns and descriptors, indicating these to be good fingerprints of soil types and sub-types. To refine this analysis, the image dataset from the experiments was used to train convolutional neural network algorithms. 60% data were used for training, and 100% prediction accuracy was achieved in both major and sub-major classification. The results of this study shows the versatility of the study of desiccation crack pattern studies and how coupling it with deep learning leads to accurate identification of the soil types, with applications in agricultural soil assessment, planetary terrain studies, geotechnical engineering, floodplain and river-basin monitoring, and image-based soil classification.