Application of machine learning and remote sensing in soil science: current state and practical opportunities

The article examines current approaches to the assessment and mapping of soil properties based on the integration of remote sensing data, ground observations, laboratory measurements, and machine learning methods. The aim of the study is to systematize data sources, preprocessing stages, and analytical algorithms used to improve the accuracy of soil property prediction under conditions of limited field measurements. The methodological basis of the study includes a review and comparative analysis of recent scientific publications, a synthesis of information on multispectral, hyperspectral, radar, and unmanned aerial vehicle data, and an assessment of the practical use of random forest, support vector machines, gradient boosting, neural networks, and probabilistic models. It is shown that the highest stability of soil models is achieved through the combined use of heterogeneous data sources, correct identification of bare soil areas, calculation of spectral, terrain, and climatic predictors, and spatially justified model validation. Ensemble and nonlinear algorithms are effective for digital soil mapping, deep learning methods are promising for processing spatial and temporal data structures, and probabilistic approaches additionally allow prediction uncertainty to be taken into account. It is substantiated that the application of machine learning and remote sensing in soil science enables a transition from point-based measurements to spatially continuous assessments of soil properties and provides an information basis for precision agriculture, agrochemical surveys, and sustainable land resource management.