Application of machine learning algorithms to classify soil components with different hydrophilicity

The capabilities of the pilot model classifier trained to recognize microbial activity traces on solid surfaces for studying soils and soil-like bodies have been preliminarily assessed. A database of 500 samples described by the authors and in open sources from 1988 to the present was collected for machine learning; among them, 59 samples represented soil horizons, 146 parent rocks and soil-like bodies, as well as rock-forming minerals, accompanying components of soil formation, xenobiotics common in technogenically transformed landscapes of the world. The samples were envolved in the database as options of dispersion, coverage with biofilms and films of other nature, chemical and physical treatment. The array of sample features significant for machine learning included quantiles of the wetting contact angle distribution and generalizing categorical indicators of surface geometry, mineral composition, and state of organic matter. The classification target function was the presence of microbial activity stable traces on a solid surface. Missing data were reconstructed using Monte Carlo procedure and bootstrapping. As a result of numerical experiments on optimizing the machine learning a balanced training dataset containing 1233 pseudo-sample elements was obtained. Six classifier models with parameter variations were trained and evaluated. The most productive classifier, a five-layer neural network with randomly dropout neurons, demonstrated a prediction accuracy of 0.74 and an ROC AUC of 0.80 on the test sample, which is higher than that of simpler and faster classifiers (accuracy and ROC AUC of 0.70). Based on the disagreement between the classifications of a human expert and trained algorithm common feature of samples that are difficult for machine classification were established: with traces of life activity, carbonate, dispersed, which allows one to determine the direction of collecting information to improve the performance of the classifier. The development of an algorithm for recognizing traces of microbial activity is useful for clarifying the mechanisms of biogeochemical and biogeotechnological processes in soils of various origins, including soil formation and terraforming.