UAV ДЕРЕКТЕРІ НЕГІЗІНДЕ АУЫЛ ШАРУАШЫЛЫҒЫ ДАҚЫЛДАРЫН ЖІКТЕУГЕ АРНАЛҒАН ГИБРИДТІ СТЕКИНГ МОДЕЛІ

Abstract

This study presents a two-level hybrid stacking framework for automatic crop classification using multispectral UAV orthomosaics. The proposed architecture combines several gradient boosting methods (LightGBM, XGBoost, and CatBoost), classical tree-based ensemble models (Random Forest and Extra Trees), and an Attention-based multilayer perceptron. During training, out-of-fold probability estimates generated by the base learners are used as inputs to a second-level Extra Trees meta-classifier, which produces the final prediction. The feature representation includes original spectral bands together with eight widely used vegetation indices (NDVI, NDRE, GNDVI, SAVI, MSR, EVI, SIPI, and MSAVI). For each index and band, statistical descriptors such as mean and standard deviation are calculated to capture spatial variability within field segments. The experimental evaluation was carried out on real multispectral UAV data collected in Eastern Kazakhstan. The proposed model achieved high classification accuracy (approximately 95 %) and a macro-averaged F1-score close to 0.95. In addition, full-field spatial segmentation results demonstrated stable performance at the level of about 99%, indicating strong consistency across the study area. Overall, the findings suggest that combining heterogeneous machine learning models within a stacking framework improves robustness and generalization in crop classification tasks.