МАШИННОЕ ОБУЧЕНИЕ ДЛЯ КЛАССИФИКАЦИИ ЗАГРЯЗНЕНИЯ РАЙОНОВ ПАДЕНИЯ РАКЕТ-НОСИТЕЛЕЙ

Abstract

Environmental monitoring of rocket booster impact areas is an important task for ensuring environmental safety. In Kazakhstan, 53 impact areas covering 4.94 million hectares are contaminated with highly toxic rocket fuel. The aim of this study is to evaluate the effectiveness of machine learning algorithms for the classification and clustering of pollution levels. A dataset of 256 soil samples was used. Four algorithms were applied: logistic regression, k-nearest neighbors, random forest, and gradient boosting, along with k-means cluster analysis. All algorithms achieved accuracy of 92.7–98.2%; logistic regression showed the best result (98.2%, F1-score 97.3%). The key pollution predictors were NDMA concentration (importance 0.274) and soil pH (0.224). Cluster analysis revealed an optimal division into two clusters (silhouette 0.889). The results confirm the feasibility of integrating machine learning into geographic information systems for environmental monitoring.