The authors describe an approach aimed at improving the efficiency of machine learning models in solving the task of classifying metageosystems, which enables overcoming the limitations imposed on the use of convolutional neural network ones. The presented solution is based on the use of superficial fully connected patterns trained on a set of information territorial descriptors to be integrated based on data from different sources. A technique for constructing a complex territorial descriptor, calculated on the basis of satellite imagery data, a digital elevation model and an electronic landscape map, was developed, which makes it possible to achieve classification accuracy comparable to that of convolutional models within a specific task. An advantage of the approach proposed in the article for enhancing the efficiency of machine learning models in solving the task of classifying metageosystems is the stability of the developed solution in the face of labeled data shortage as well as the possibility of reuse in the study of new territorial systems, subject to additional training and fine tuning. The proposed method can be used to solve the task of automated monitoring the changes in the land use structure and geophysical envelope as well as automated validation of digital maps of a significant territorial coverage.